gpt-6.1-sol 发布，五分之一的价格，真能接近 Astra？深入解读！

三少科技 首頁 每日列表 原始影片
實際影片長度：10:41.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:04.720
enOpenAI released GPT 6.1 SOL on September 29.
0:04.720–0:06.237
enSome say this model
0:06.237–0:08.133
enslashed costs by 50% for
0:08.133–0:09.460
entheir $200 tier.
0:09.460–0:11.393
enOthers say it unceremoniously cut
0:11.393–0:12.860
entheir subscription usage
0:12.860–0:14.127
enby 50% while calling it
0:14.127–0:15.327
enan efficiency boost.
0:15.327–0:16.260
enWhy are there two
0:16.260–0:17.460
encompletely opposite
0:17.460–0:18.660
enreactions to the same
0:18.660–0:19.527
enannouncement?
0:19.527–0:20.527
enBecause cheaper,
0:20.527–0:21.027
enand,
0:21.027–0:21.960
enmore expensive,
0:21.960–0:23.360
enhappened on the same day.
0:23.360–0:24.760
enDepending on whether you
0:24.760–0:26.227
enpay per token or per month.
0:26.227–0:27.627
enThis isn't about whether
0:27.627–0:28.827
enthe new model is good.
0:28.827–0:30.227
enIt's about OpenAI saying
0:30.227–0:31.693
entwo contradictory things
0:31.693–0:32.827
ento different users.
0:32.827–0:33.327
enToday,
0:33.327–0:35.460
enwe won't discuss if 6.1 is worth using.
0:35.460–0:37.060
enWe'll clarify three things.
0:37.060–0:37.560
enFirst,
0:37.560–0:38.693
enwhy can the bills for
0:38.693–0:39.827
enmodels with the same
0:39.827–0:41.493
enlist price differ by 10 times?
0:41.493–0:41.993
enSecond,
0:41.993–0:43.127
enwhich official data
0:43.127–0:45.127
enmatches independent evaluations,
0:45.127–0:46.060
enand which do not?
0:46.060–0:46.560
enThird,
0:46.560–0:47.960
enhow should you determine
0:47.960–0:49.160
enif it is right for you?
0:49.160–0:49.660
enOkay,
0:49.660–0:51.393
enlet's start with the first one.
0:51.393–0:53.327
enGPT-6.1 is not the flagship model.
0:53.327–0:54.356
enIt is an upgraded
0:54.356–0:55.921
enversion of the GPT-6 SOL
0:55.921–0:57.800
enreleased on September 22nd.
0:57.920–0:59.387
enOpenAI's current product
0:59.387–1:00.587
enline looks like this.
1:00.587–1:01.187
enAt the top,
1:01.187–1:02.453
enGPT-6 Ultra is the most
1:02.453–1:03.980
enpowerful and expensive.
1:04.300–1:05.189
enIn the middle,
1:05.189–1:07.780
enGPT-6 SOL is slightly weaker and cheaper.
1:08.140–1:09.057
enAt the bottom,
1:09.057–1:11.580
enGPT-6 Lite is the weakest and cheapest.
1:12.220–1:15.240
enThe official positioning for 6.1 SOL is this.
1:15.660–1:16.860
enProviding Near Ultra
1:16.860–1:18.127
enIntelligence at 1-5th
1:18.127–1:19.193
enthe price of Ultra.
1:19.500–1:20.967
enThe pricing is as follows.
1:20.967–1:24.200
enUltra is $10 input, $50 output
1:24.200–1:28.040
en6.1 SOL is $2 input, $10 output
1:28.040–1:31.860
enLight is $0.10 input, $0.50 output
1:31.860–1:34.368
enIf you calculate Ultra divided by SOL,
1:34.368–1:35.700
enit is exactly 5 times
1:35.700–1:36.567
enThis is what the
1:36.567–1:38.033
enofficial statement means
1:38.033–1:39.100
enby 1 fifth the price
1:39.100–1:40.767
enBut that is just the unit price
1:40.767–1:43.920
enThe real cost is a completely different story
1:43.920–1:45.100
enWhy?
1:45.100–1:46.242
enOn September 29th,
1:46.242–1:47.455
ensomeone on X posted a
1:47.455–1:48.739
enchart with a headline
1:48.739–1:50.380
enmeaning that for the same $2
1:50.380–1:51.547
en$10.00 price tag,
1:51.547–1:53.569
enthe bill can differ by 10 times.
1:53.569–1:55.280
enWhat were they comparing?
1:56.080–1:58.574
enGPT 6.1 SOL at $2 input,
1:58.574–1:59.886
en$10 output.
1:59.886–2:03.168
enClaude Sonnet 5.5 at $2 input,
2:03.168–2:04.480
en$10 output.
2:04.660–2:06.341
enThe prices are identical,
2:06.341–2:07.716
enbut the per-task cost
2:07.716–2:09.397
encalculated by artificial
2:09.397–2:10.620
enanalysis shows 5.5
2:10.620–2:11.702
enis $7.60,
2:11.702–2:14.273
enwhile 6.1 SOL is $0.72.
2:14.273–2:17.384
enA difference of 10.6 times.
2:17.384–2:17.925
enWhy?
2:17.925–2:20.360
enBecause Sonnet 5.5 is
2:20.360–2:21.120
zh很 wordy
2:21.120–2:25.460
zh它映出了一共193,000典的登場在本項上
2:25.460–2:28.820
zh那是最高的一種的模式測試驗
2:28.820–2:31.160
zh它是60%更多的 Opus 5.5
2:31.160–2:33.660
zh和7倍的GPT6Ultra
2:33.660–2:35.280
en6.1 SOL
2:35.280–2:41.220
zhA發現它的登場是10%-30%的高度
2:41.220–2:45.320
zh但它是10%的一共1%的上一共1%
2:45.320–2:47.320
zh它是比方的一页价值,
2:47.320–2:48.720
zh但是因为它说少
2:48.720–2:50.400
zh你费同一项供应价
2:50.400–2:51.680
zh但一项供应价格更
2:51.680–2:53.040
zh而另一项供应价格更多
2:53.040–2:55.040
zh当然,这些人会比较不同
2:55.040–2:57.200
zh这就是第一个我们必须理解
2:57.200–2:59.560
zh在这些人的价格上不符定在这些价格上
3:00.440–3:02.160
zh与其中的价格出现
3:02.160–3:03.360
zh我们看看这些 data
3:04.000–3:05.520
zh2从公司的公司
3:05.520–3:07.560
zh2从公司的公司
3:07.560–3:09.040
zh4个人数在 total
3:09.760–3:10.480
zh第一
3:10.480–3:13.680
en75.2 vs 74.1
3:13.680–3:14.847
enThere is a chart on
3:14.847–3:16.326
enOpenAI's release page
3:16.326–3:17.260
enfor SWE Bench.
3:17.980–3:19.380
enIt measures AI's ability
3:19.380–3:20.513
ento complete complex
3:20.513–3:21.780
ensoftware engineering
3:21.780–3:23.180
entasks in real codebases.
3:23.340–3:25.933
enO1 has five reasoning levels,
3:25.933–3:26.433
enlow,
3:26.433–3:27.073
enmedium,
3:27.073–3:27.592
enhigh,
3:27.592–3:28.214
enultra,
3:28.214–3:28.940
enand max.
3:29.720–3:31.365
enThe numbers are,
3:31.365–3:33.480
enlow is 70.0% at $0.17.
3:35.100–3:38.380
enMedium is 73.2% at $0.42.
3:39.280–3:42.460
enHigh is 75.2% at $0.65.
3:42.460–3:46.520
enUltra is 74.9% at $0.79
3:46.520–3:50.980
enMax is 71.9% at $1.57
3:50.980–3:53.740
enYou see, the high level is the best
3:53.740–3:55.680
enThe max level actually regressed
3:55.680–3:56.978
enAnd it costs 2.4 times
3:56.978–3:58.420
enmore than the high level
3:58.420–4:00.717
enCompared to Claude 3.5
4:00.717–4:03.980
enSonnet's best at 74.1% for $4.43
4:03.980–4:05.939
enO1's high level scores
4:05.939–4:08.620
enhigher and is 6.8 times cheaper
4:08.620–4:10.608
enThe official claim says it's,
4:10.608–4:11.563
enapproaching,
4:11.563–4:12.063
enSonnet
4:12.460–4:14.793
enbut the numbers show it surpasses Sonnet.
4:15.180–4:17.847
enThe official statement downplayed the result.
4:17.847–4:19.544
enThe second number,
4:19.544–4:21.800
en71.9% equals 71.9%.
4:21.800–4:24.120
enThis is O1's ultra and max levels.
4:24.840–4:27.225
enThe scores are identical,
4:27.225–4:28.309
enboth 71.9%,
4:28.309–4:30.260
enbut max costs double.
4:31.680–4:33.246
enOfficial data tells you,
4:33.246–4:34.440
endon't set it to max.
4:35.220–4:36.487
enThis is also true on the
4:36.487–4:38.487
enArtificial Analysis Leaderboard.
4:38.487–4:39.887
entheir coding agent index
4:39.887–4:41.258
enshows the ultra level is
4:41.258–4:43.060
enthree points higher than max
4:43.060–4:44.245
enhigher reasoning
4:44.245–4:46.300
enintensity isn't always better
4:46.300–4:49.480
enthe third number 52 versus 58
4:49.480–4:52.720
enfrom artificial analysis's intelligence index
4:52.720–4:53.453
enthis is their
4:53.453–4:54.920
encomprehensive score from
4:54.920–4:56.053
enactual evaluations
4:56.053–4:58.600
enthe scores are
4:58.600–5:01.080
enclaude 3.5 opus at 58
5:01.080–5:03.640
ensonnet 3.5 at 56
5:03.640–5:05.120
enextra at 53
5:05.120–5:06.700
enflash at 53
5:06.700–5:09.800
enand 01 at 52.01 is the lowest.
5:11.300–5:13.440
enBut the cost is also the lowest.
5:13.660–5:15.540
enPer task, it's 72 cents.
5:15.800–5:17.673
enSonnet is $3.26,
5:17.673–5:20.080
enand Opus 3.5 is $5.98.
5:20.500–5:23.080
enClaude 3.5 Sonnet is $7.60.
5:23.620–5:25.620
enSo the question isn't which is best,
5:25.780–5:27.220
enbut whether your task
5:27.220–5:29.140
enneeds a 58 or if a 52 is enough.
5:29.500–5:32.305
enThe fourth number,
5:32.305–5:34.760
en646 vs 643 vs 728.
5:34.760–5:37.227
enThis is the big code bench leaderboard score
5:37.227–5:38.427
enNot an official list,
5:38.427–5:39.740
enbut an independent one
5:39.740–5:41.140
enThe results are like this
5:41.140–5:43.660
en01-646 points
5:43.660–5:46.580
enPrevious version, 643 points
5:46.580–5:49.500
enGPT40, 728 points
5:49.500–5:50.512
en01 is only 3 points
5:50.512–5:52.400
enhigher than the previous version
5:52.400–5:55.060
enAnd it lags behind 40 by 82 points
5:55.060–5:56.243
enThey outright say it
5:56.243–5:58.260
enmoved a lot on their own benchmarks
5:58.260–6:00.520
enBut barely moved on independent ones
6:00.520–6:02.000
enThe term used is
6:02.000–6:04.540
enBenchmarking to the test on their own list
6:04.760–6:06.194
enCognition's Frontier
6:06.194–6:08.080
enCode 1.1 Extended is the same.
6:08.680–6:10.324
en01 is 60.4% and the
6:10.324–6:12.954
enprevious version was 60.7%,
6:12.954–6:14.160
enalmost flat.
6:14.660–6:15.593
enThe official and
6:15.593–6:17.127
enindependent leaderboards
6:17.127–6:18.727
entell two different stories.
6:18.727–6:19.420
enOK, pause.
6:20.380–6:21.780
enWe now know four numbers.
6:22.220–6:22.720
enFirst,
6:22.720–6:24.440
enit beats Sonnet on some tasks.
6:25.080–6:25.580
enSecond,
6:25.580–6:27.500
enMax is more expensive and worse.
6:27.700–6:28.200
enThird,
6:28.200–6:29.600
enthe comprehensive score
6:29.600–6:31.067
enis the lowest of the bunch,
6:31.067–6:32.800
enbut the cost is also the lowest.
6:32.800–6:35.260
zh四、インデピンチマーク インプーブメント
6:35.260–6:36.260
enWhat do these four
6:36.260–6:38.260
ennumbers tell you when put together?
6:38.260–6:40.540
en6.1 Sonnet is not a cheap version of Opus.
6:40.540–6:43.100
enIt is an efficiency model for specific tasks.
6:43.100–6:44.860
enIt is not an all-rounder.
6:44.860–6:46.540
enIt is specialized.
6:46.540–6:48.700
enSo the question arises.
6:48.700–6:51.420
enHow do you decide if it's right for you?
6:51.420–6:54.300
enWe divide scenarios into three categories.
6:54.300–6:55.202
enCode fixing,
6:55.202–6:56.104
enbatch tasks,
6:56.104–6:57.580
enand repetitive work.
6:57.580–6:58.780
enIf you have to fix bugs
6:58.780–6:59.913
endaily or run massive
6:59.913–7:01.580
enrepetitive knowledge tasks.
7:01.580–7:03.900
en6.1 sonnet is perfect for this
7:03.900–7:06.080
ena developer conducted an experiment
7:06.080–7:09.860
enthey put 105 bugs into 2 code repositories
7:09.860–7:12.920
enopus scored 45 and cost 33 dollars
7:12.920–7:15.025
en6.1 sonnet scored 44 and
7:15.025–7:17.340
encost 6 dollars and 56 cents
7:17.340–7:22.080
enopus 3.5 scored 41.7 and cost 58 dollars
7:22.080–7:23.461
en6.1 sonnet nearly
7:23.461–7:26.040
enmatched opus at one fifth the cost
7:26.040–7:28.100
enhowever they noted n equals 1
7:28.100–7:30.500
enmeaning they only did this experiment once
7:30.500–7:31.233
enOn the Click,
7:31.233–7:33.633
enBenchmark for White Collar Knowledge Work
7:33.633–7:35.455
en6.1 Sonnet scored 80.0,
7:35.455–7:37.020
encosting 18 cents
7:37.020–7:40.740
enClaud 3.5 scored 81.4, costing 89
7:41.195–7:42.365
enFor these tasks,
7:42.365–7:44.955
enusing 6.1 sonnet is more than enough.
7:44.955–7:46.155
enDon't use Macs.
7:46.155–7:47.643
enThe second category,
7:47.643–7:48.966
enCreative Visuals,
7:48.966–7:49.466
enGames,
7:49.466–7:51.115
enand Complex Reasoning.
7:51.115–7:52.285
zh如果您做3D scenes,
7:52.285–7:53.545
envisual effects,
7:53.545–7:54.715
enor game design,
7:54.715–7:56.875
entasks needing imagination,
7:57.115–7:59.195
en6.1 Sonnet may not be enough.
7:59.515–8:00.848
enBridgeBench asked it to
8:00.848–8:02.315
enrender a sunset seascape.
8:02.315–8:03.515
enThe conclusion was it
8:03.515–8:04.782
enlooked like a sunset in
8:04.782–8:06.315
ena different price bracket,
8:06.315–8:08.115
enand the sea was barely rendered.
8:08.115–8:10.248
enSomeone posted a tweet that got over 1,
8:10.248–8:10.848
en200 likes.
8:10.848–8:12.848
enOPUS 3.5 is much better than GPT 6.1.
8:12.915–8:15.048
enSomeone else made a comparison video.
8:15.048–8:15.848
enWith the same,
8:15.848–8:17.048
enhigh effort setting,
8:17.048–8:18.382
enthe conclusion was that
8:18.382–8:20.315
enOPUS 3 won and 50 Jae-on in quality,
8:20.315–8:20.815
enbugs,
8:20.815–8:21.882
enand color schemes.
8:21.882–8:24.282
enSONNET is only better in token efficiency.
8:24.282–8:25.482
enOthers have also used
8:25.482–8:27.015
enlogic questions to test it.
8:27.015–8:28.615
en6.1 SONNET set to max failed.
8:28.615–8:31.675
enOPUS 3.5 set to medium passed easily.
8:31.675–8:32.987
enFor these tasks,
8:32.987–8:34.675
enstick to CLOD or OPUS.
8:34.675–8:36.186
enThe third category,
8:36.186–8:37.875
enscientific research.
8:37.875–8:40.075
enThe company admitted this themselves.
8:40.075–8:42.951
enOn the Terminal Bench Science benchmark,
8:42.951–8:44.485
en6.1 SONNET is 57.0,
8:44.485–8:46.115
enwhile OPUS is 68.1%.
8:46.115–8:48.273
enthe official word is that the most
8:48.273–8:49.891
enchallenging scientific
8:49.891–8:51.895
enresearch tasks should use opus
8:51.895–8:53.599
enfor these tasks use opus
8:53.599–8:55.815
enand don't try to save a few bucks
8:55.815–8:58.527
enalright now you know 6.1
8:58.527–9:01.375
ensonnet isn't a cheap opus
9:01.375–9:04.095
enit's an efficiency tool for specific tasks
9:04.095–9:06.946
enif your task aligns it's
9:06.946–9:09.655
engreat if not it's a trap
9:09.655–9:12.115
enand there's one more thing you need to know
9:12.115–9:13.366
enOn September 29,
9:13.366–9:14.350
enthe same day,
9:14.350–9:16.495
enOpenAI's TBO posted a tweet.
9:16.615–9:17.832
enThe quota for the new
9:17.832–9:19.264
enPro 200 plan was reduced
9:19.264–9:20.195
enfrom 20x to 10x.
9:20.355–9:22.955
enOld users temporarily retained the 20x limit.
9:22.955–9:23.822
enOn the same day,
9:23.822–9:25.095
enthey launched Pro 500.
9:25.355–9:28.095
en$500 per month for a 25x quota.
9:28.655–9:30.675
enAPI users saw it as getting cheaper.
9:31.155–9:32.488
enSubscription users saw
9:32.488–9:34.155
enit as getting more expensive.
9:34.155–9:36.288
enSo this isn't just a new model release.
9:37.095–9:38.495
enThis is OpenAI repricing
9:38.495–9:40.162
enfor different user segments.
9:40.162–9:41.362
endo you understand now
9:41.362–9:42.488
enwhy some people are
9:42.488–9:44.615
encheering while others are cursing
9:44.615–9:46.015
enbecause they are looking
9:46.015–9:47.348
enat two different things
9:47.348–9:49.766
enone last thing if you
9:49.766–9:53.055
enask me if 6.15 is worth using
9:53.055–9:54.352
enmy answer is don't ask
9:54.352–9:56.515
enif it's worth it ask if it's suitable
9:56.515–9:58.715
enit is suitable for code
9:58.715–10:00.915
enrepair and batch tasks
10:00.915–10:03.215
enuse the higher tier and turn on max
10:03.215–10:05.335
enfor creative and complex reasoning
10:05.335–10:07.655
encurrent tests show it can't beat claude
10:07.655–10:09.695
enContinue using Claude or Opus
10:09.695–10:11.295
enAs for scientific research,
10:11.295–10:13.295
enthe official guide already told you
10:13.295–10:15.028
enOne more thing about using Opus
10:15.028–10:16.393
enIf you are a subscriber,
10:16.393–10:17.855
enyour quota has been cut
10:17.855–10:19.188
enYou need to recalculate
10:19.188–10:20.388
enif it's enough for you
10:20.388–10:21.639
enIf it's not enough,
10:21.639–10:23.455
eneither upgrade to Pro500
10:23.455–10:26.155
enOr switch to API and pay by token
10:26.155–10:27.505
enThis isn't a question of
10:27.505–10:28.855
enwhich model is stronger
10:28.855–10:31.655
enThis is a question of OpenAI's repricing
10:31.655–10:33.455
enThat's all for today's content
10:33.455–10:34.422
enIf you liked it,
10:34.422–10:35.539
enremember to like,
10:35.539–10:36.060
enfollow,
10:36.060–10:36.655
enand share
10:36.655–10:38.910
enThis is my motivation to keep creating.
10:38.910–10:39.935
enSee you next time!
0:00.000–0:04.720
OpenAI released GPT 6.1 SOL on September 29.
0:04.720–0:06.237
Some say this model
0:06.237–0:08.133
slashed costs by 50% for
0:08.133–0:09.460
their $200 tier.
0:09.460–0:11.393
Others say it unceremoniously cut
0:11.393–0:12.860
their subscription usage
0:12.860–0:14.127
by 50% while calling it
0:14.127–0:15.327
an efficiency boost.
0:15.327–0:16.260
Why are there two
0:16.260–0:17.460
completely opposite
0:17.460–0:18.660
reactions to the same
0:18.660–0:19.527
announcement?
0:19.527–0:20.527
Because cheaper,
0:20.527–0:21.027
and,
0:21.027–0:21.960
more expensive,
0:21.960–0:23.360
happened on the same day.
0:23.360–0:24.760
Depending on whether you
0:24.760–0:26.227
pay per token or per month.
0:26.227–0:27.627
This isn't about whether
0:27.627–0:28.827
the new model is good.
0:28.827–0:30.227
It's about OpenAI saying
0:30.227–0:31.693
two contradictory things
0:31.693–0:32.827
to different users.
0:32.827–0:33.327
Today,
0:33.327–0:35.460
we won't discuss if 6.1 is worth using.
0:35.460–0:37.060
We'll clarify three things.
0:37.060–0:37.560
First,
0:37.560–0:38.693
why can the bills for
0:38.693–0:39.827
models with the same
0:39.827–0:41.493
list price differ by 10 times?
0:41.493–0:41.993
Second,
0:41.993–0:43.127
which official data
0:43.127–0:45.127
matches independent evaluations,
0:45.127–0:46.060
and which do not?
0:46.060–0:46.560
Third,
0:46.560–0:47.960
how should you determine
0:47.960–0:49.160
if it is right for you?
0:49.160–0:49.660
Okay,
0:49.660–0:51.393
let's start with the first one.
0:51.393–0:53.327
GPT-6.1 is not the flagship model.
0:53.327–0:54.356
It is an upgraded
0:54.356–0:55.921
version of the GPT-6 SOL
0:55.921–0:57.800
released on September 22nd.
0:57.920–0:59.387
OpenAI's current product
0:59.387–1:00.587
line looks like this.
1:00.587–1:01.187
At the top,
1:01.187–1:02.453
GPT-6 Ultra is the most
1:02.453–1:03.980
powerful and expensive.
1:04.300–1:05.189
In the middle,
1:05.189–1:07.780
GPT-6 SOL is slightly weaker and cheaper.
1:08.140–1:09.057
At the bottom,
1:09.057–1:11.580
GPT-6 Lite is the weakest and cheapest.
1:12.220–1:15.240
The official positioning for 6.1 SOL is this.
1:15.660–1:16.860
Providing Near Ultra
1:16.860–1:18.127
Intelligence at 1-5th
1:18.127–1:19.193
the price of Ultra.
1:19.500–1:20.967
The pricing is as follows.
1:20.967–1:24.200
Ultra is $10 input, $50 output
1:24.200–1:28.040
6.1 SOL is $2 input, $10 output
1:28.040–1:31.860
Light is $0.10 input, $0.50 output
1:31.860–1:34.368
If you calculate Ultra divided by SOL,
1:34.368–1:35.700
it is exactly 5 times
1:35.700–1:36.567
This is what the
1:36.567–1:38.033
official statement means
1:38.033–1:39.100
by 1 fifth the price
1:39.100–1:40.767
But that is just the unit price
1:40.767–1:43.920
The real cost is a completely different story
1:43.920–1:45.100
Why?
1:45.100–1:46.242
On September 29th,
1:46.242–1:47.455
someone on X posted a
1:47.455–1:48.739
chart with a headline
1:48.739–1:50.380
meaning that for the same $2
1:50.380–1:51.547
$10.00 price tag,
1:51.547–1:53.569
the bill can differ by 10 times.
1:53.569–1:55.280
What were they comparing?
1:56.080–1:58.574
GPT 6.1 SOL at $2 input,
1:58.574–1:59.886
$10 output.
1:59.886–2:03.168
Claude Sonnet 5.5 at $2 input,
2:03.168–2:04.480
$10 output.
2:04.660–2:06.341
The prices are identical,
2:06.341–2:07.716
but the per-task cost
2:07.716–2:09.397
calculated by artificial
2:09.397–2:10.620
analysis shows 5.5
2:10.620–2:11.702
is $7.60,
2:11.702–2:14.273
while 6.1 SOL is $0.72.
2:14.273–2:17.384
A difference of 10.6 times.
2:17.384–2:17.925
Why?
2:17.925–2:20.360
Because Sonnet 5.5 is
2:20.360–2:21.120
很 wordy
2:21.120–2:25.460
它映出了一共193,000典的登場在本項上
2:25.460–2:28.820
那是最高的一種的模式測試驗
2:28.820–2:31.160
它是60%更多的 Opus 5.5
2:31.160–2:33.660
和7倍的GPT6Ultra
2:33.660–2:35.280
6.1 SOL
2:35.280–2:41.220
A發現它的登場是10%-30%的高度
2:41.220–2:45.320
但它是10%的一共1%的上一共1%
2:45.320–2:47.320
它是比方的一页价值,
2:47.320–2:48.720
但是因为它说少
2:48.720–2:50.400
你费同一项供应价
2:50.400–2:51.680
但一项供应价格更
2:51.680–2:53.040
而另一项供应价格更多
2:53.040–2:55.040
当然,这些人会比较不同
2:55.040–2:57.200
这就是第一个我们必须理解
2:57.200–2:59.560
在这些人的价格上不符定在这些价格上
3:00.440–3:02.160
与其中的价格出现
3:02.160–3:03.360
我们看看这些 data
3:04.000–3:05.520
2从公司的公司
3:05.520–3:07.560
2从公司的公司
3:07.560–3:09.040
4个人数在 total
3:09.760–3:10.480
第一
3:10.480–3:13.680
75.2 vs 74.1
3:13.680–3:14.847
There is a chart on
3:14.847–3:16.326
OpenAI's release page
3:16.326–3:17.260
for SWE Bench.
3:17.980–3:19.380
It measures AI's ability
3:19.380–3:20.513
to complete complex
3:20.513–3:21.780
software engineering
3:21.780–3:23.180
tasks in real codebases.
3:23.340–3:25.933
O1 has five reasoning levels,
3:25.933–3:26.433
low,
3:26.433–3:27.073
medium,
3:27.073–3:27.592
high,
3:27.592–3:28.214
ultra,
3:28.214–3:28.940
and max.
3:29.720–3:31.365
The numbers are,
3:31.365–3:33.480
low is 70.0% at $0.17.
3:35.100–3:38.380
Medium is 73.2% at $0.42.
3:39.280–3:42.460
High is 75.2% at $0.65.
3:42.460–3:46.520
Ultra is 74.9% at $0.79
3:46.520–3:50.980
Max is 71.9% at $1.57
3:50.980–3:53.740
You see, the high level is the best
3:53.740–3:55.680
The max level actually regressed
3:55.680–3:56.978
And it costs 2.4 times
3:56.978–3:58.420
more than the high level
3:58.420–4:00.717
Compared to Claude 3.5
4:00.717–4:03.980
Sonnet's best at 74.1% for $4.43
4:03.980–4:05.939
O1's high level scores
4:05.939–4:08.620
higher and is 6.8 times cheaper
4:08.620–4:10.608
The official claim says it's,
4:10.608–4:11.563
approaching,
4:11.563–4:12.063
Sonnet
4:12.460–4:14.793
but the numbers show it surpasses Sonnet.
4:15.180–4:17.847
The official statement downplayed the result.
4:17.847–4:19.544
The second number,
4:19.544–4:21.800
71.9% equals 71.9%.
4:21.800–4:24.120
This is O1's ultra and max levels.
4:24.840–4:27.225
The scores are identical,
4:27.225–4:28.309
both 71.9%,
4:28.309–4:30.260
but max costs double.
4:31.680–4:33.246
Official data tells you,
4:33.246–4:34.440
don't set it to max.
4:35.220–4:36.487
This is also true on the
4:36.487–4:38.487
Artificial Analysis Leaderboard.
4:38.487–4:39.887
their coding agent index
4:39.887–4:41.258
shows the ultra level is
4:41.258–4:43.060
three points higher than max
4:43.060–4:44.245
higher reasoning
4:44.245–4:46.300
intensity isn't always better
4:46.300–4:49.480
the third number 52 versus 58
4:49.480–4:52.720
from artificial analysis's intelligence index
4:52.720–4:53.453
this is their
4:53.453–4:54.920
comprehensive score from
4:54.920–4:56.053
actual evaluations
4:56.053–4:58.600
the scores are
4:58.600–5:01.080
claude 3.5 opus at 58
5:01.080–5:03.640
sonnet 3.5 at 56
5:03.640–5:05.120
extra at 53
5:05.120–5:06.700
flash at 53
5:06.700–5:09.800
and 01 at 52.01 is the lowest.
5:11.300–5:13.440
But the cost is also the lowest.
5:13.660–5:15.540
Per task, it's 72 cents.
5:15.800–5:17.673
Sonnet is $3.26,
5:17.673–5:20.080
and Opus 3.5 is $5.98.
5:20.500–5:23.080
Claude 3.5 Sonnet is $7.60.
5:23.620–5:25.620
So the question isn't which is best,
5:25.780–5:27.220
but whether your task
5:27.220–5:29.140
needs a 58 or if a 52 is enough.
5:29.500–5:32.305
The fourth number,
5:32.305–5:34.760
646 vs 643 vs 728.
5:34.760–5:37.227
This is the big code bench leaderboard score
5:37.227–5:38.427
Not an official list,
5:38.427–5:39.740
but an independent one
5:39.740–5:41.140
The results are like this
5:41.140–5:43.660
01-646 points
5:43.660–5:46.580
Previous version, 643 points
5:46.580–5:49.500
GPT40, 728 points
5:49.500–5:50.512
01 is only 3 points
5:50.512–5:52.400
higher than the previous version
5:52.400–5:55.060
And it lags behind 40 by 82 points
5:55.060–5:56.243
They outright say it
5:56.243–5:58.260
moved a lot on their own benchmarks
5:58.260–6:00.520
But barely moved on independent ones
6:00.520–6:02.000
The term used is
6:02.000–6:04.540
Benchmarking to the test on their own list
6:04.760–6:06.194
Cognition's Frontier
6:06.194–6:08.080
Code 1.1 Extended is the same.
6:08.680–6:10.324
01 is 60.4% and the
6:10.324–6:12.954
previous version was 60.7%,
6:12.954–6:14.160
almost flat.
6:14.660–6:15.593
The official and
6:15.593–6:17.127
independent leaderboards
6:17.127–6:18.727
tell two different stories.
6:18.727–6:19.420
OK, pause.
6:20.380–6:21.780
We now know four numbers.
6:22.220–6:22.720
First,
6:22.720–6:24.440
it beats Sonnet on some tasks.
6:25.080–6:25.580
Second,
6:25.580–6:27.500
Max is more expensive and worse.
6:27.700–6:28.200
Third,
6:28.200–6:29.600
the comprehensive score
6:29.600–6:31.067
is the lowest of the bunch,
6:31.067–6:32.800
but the cost is also the lowest.
6:32.800–6:35.260
四、インデピンチマーク インプーブメント
6:35.260–6:36.260
What do these four
6:36.260–6:38.260
numbers tell you when put together?
6:38.260–6:40.540
6.1 Sonnet is not a cheap version of Opus.
6:40.540–6:43.100
It is an efficiency model for specific tasks.
6:43.100–6:44.860
It is not an all-rounder.
6:44.860–6:46.540
It is specialized.
6:46.540–6:48.700
So the question arises.
6:48.700–6:51.420
How do you decide if it's right for you?
6:51.420–6:54.300
We divide scenarios into three categories.
6:54.300–6:55.202
Code fixing,
6:55.202–6:56.104
batch tasks,
6:56.104–6:57.580
and repetitive work.
6:57.580–6:58.780
If you have to fix bugs
6:58.780–6:59.913
daily or run massive
6:59.913–7:01.580
repetitive knowledge tasks.
7:01.580–7:03.900
6.1 sonnet is perfect for this
7:03.900–7:06.080
a developer conducted an experiment
7:06.080–7:09.860
they put 105 bugs into 2 code repositories
7:09.860–7:12.920
opus scored 45 and cost 33 dollars
7:12.920–7:15.025
6.1 sonnet scored 44 and
7:15.025–7:17.340
cost 6 dollars and 56 cents
7:17.340–7:22.080
opus 3.5 scored 41.7 and cost 58 dollars
7:22.080–7:23.461
6.1 sonnet nearly
7:23.461–7:26.040
matched opus at one fifth the cost
7:26.040–7:28.100
however they noted n equals 1
7:28.100–7:30.500
meaning they only did this experiment once
7:30.500–7:31.233
On the Click,
7:31.233–7:33.633
Benchmark for White Collar Knowledge Work
7:33.633–7:35.455
6.1 Sonnet scored 80.0,
7:35.455–7:37.020
costing 18 cents
7:37.020–7:40.740
Claud 3.5 scored 81.4, costing 89
7:41.195–7:42.365
For these tasks,
7:42.365–7:44.955
using 6.1 sonnet is more than enough.
7:44.955–7:46.155
Don't use Macs.
7:46.155–7:47.643
The second category,
7:47.643–7:48.966
Creative Visuals,
7:48.966–7:49.466
Games,
7:49.466–7:51.115
and Complex Reasoning.
7:51.115–7:52.285
如果您做3D scenes,
7:52.285–7:53.545
visual effects,
7:53.545–7:54.715
or game design,
7:54.715–7:56.875
tasks needing imagination,
7:57.115–7:59.195
6.1 Sonnet may not be enough.
7:59.515–8:00.848
BridgeBench asked it to
8:00.848–8:02.315
render a sunset seascape.
8:02.315–8:03.515
The conclusion was it
8:03.515–8:04.782
looked like a sunset in
8:04.782–8:06.315
a different price bracket,
8:06.315–8:08.115
and the sea was barely rendered.
8:08.115–8:10.248
Someone posted a tweet that got over 1,
8:10.248–8:10.848
200 likes.
8:10.848–8:12.848
OPUS 3.5 is much better than GPT 6.1.
8:12.915–8:15.048
Someone else made a comparison video.
8:15.048–8:15.848
With the same,
8:15.848–8:17.048
high effort setting,
8:17.048–8:18.382
the conclusion was that
8:18.382–8:20.315
OPUS 3 won and 50 Jae-on in quality,
8:20.315–8:20.815
bugs,
8:20.815–8:21.882
and color schemes.
8:21.882–8:24.282
SONNET is only better in token efficiency.
8:24.282–8:25.482
Others have also used
8:25.482–8:27.015
logic questions to test it.
8:27.015–8:28.615
6.1 SONNET set to max failed.
8:28.615–8:31.675
OPUS 3.5 set to medium passed easily.
8:31.675–8:32.987
For these tasks,
8:32.987–8:34.675
stick to CLOD or OPUS.
8:34.675–8:36.186
The third category,
8:36.186–8:37.875
scientific research.
8:37.875–8:40.075
The company admitted this themselves.
8:40.075–8:42.951
On the Terminal Bench Science benchmark,
8:42.951–8:44.485
6.1 SONNET is 57.0,
8:44.485–8:46.115
while OPUS is 68.1%.
8:46.115–8:48.273
the official word is that the most
8:48.273–8:49.891
challenging scientific
8:49.891–8:51.895
research tasks should use opus
8:51.895–8:53.599
for these tasks use opus
8:53.599–8:55.815
and don't try to save a few bucks
8:55.815–8:58.527
alright now you know 6.1
8:58.527–9:01.375
sonnet isn't a cheap opus
9:01.375–9:04.095
it's an efficiency tool for specific tasks
9:04.095–9:06.946
if your task aligns it's
9:06.946–9:09.655
great if not it's a trap
9:09.655–9:12.115
and there's one more thing you need to know
9:12.115–9:13.366
On September 29,
9:13.366–9:14.350
the same day,
9:14.350–9:16.495
OpenAI's TBO posted a tweet.
9:16.615–9:17.832
The quota for the new
9:17.832–9:19.264
Pro 200 plan was reduced
9:19.264–9:20.195
from 20x to 10x.
9:20.355–9:22.955
Old users temporarily retained the 20x limit.
9:22.955–9:23.822
On the same day,
9:23.822–9:25.095
they launched Pro 500.
9:25.355–9:28.095
$500 per month for a 25x quota.
9:28.655–9:30.675
API users saw it as getting cheaper.
9:31.155–9:32.488
Subscription users saw
9:32.488–9:34.155
it as getting more expensive.
9:34.155–9:36.288
So this isn't just a new model release.
9:37.095–9:38.495
This is OpenAI repricing
9:38.495–9:40.162
for different user segments.
9:40.162–9:41.362
do you understand now
9:41.362–9:42.488
why some people are
9:42.488–9:44.615
cheering while others are cursing
9:44.615–9:46.015
because they are looking
9:46.015–9:47.348
at two different things
9:47.348–9:49.766
one last thing if you
9:49.766–9:53.055
ask me if 6.15 is worth using
9:53.055–9:54.352
my answer is don't ask
9:54.352–9:56.515
if it's worth it ask if it's suitable
9:56.515–9:58.715
it is suitable for code
9:58.715–10:00.915
repair and batch tasks
10:00.915–10:03.215
use the higher tier and turn on max
10:03.215–10:05.335
for creative and complex reasoning
10:05.335–10:07.655
current tests show it can't beat claude
10:07.655–10:09.695
Continue using Claude or Opus
10:09.695–10:11.295
As for scientific research,
10:11.295–10:13.295
the official guide already told you
10:13.295–10:15.028
One more thing about using Opus
10:15.028–10:16.393
If you are a subscriber,
10:16.393–10:17.855
your quota has been cut
10:17.855–10:19.188
You need to recalculate
10:19.188–10:20.388
if it's enough for you
10:20.388–10:21.639
If it's not enough,
10:21.639–10:23.455
either upgrade to Pro500
10:23.455–10:26.155
Or switch to API and pay by token
10:26.155–10:27.505
This isn't a question of
10:27.505–10:28.855
which model is stronger
10:28.855–10:31.655
This is a question of OpenAI's repricing
10:31.655–10:33.455
That's all for today's content
10:33.455–10:34.422
If you liked it,
10:34.422–10:35.539
remember to like,
10:35.539–10:36.060
follow,
10:36.060–10:36.655
and share
10:36.655–10:38.910
This is my motivation to keep creating.
10:38.910–10:39.935
See you next time!
0:00.000–0:04.720
enOpenAI released GPT 6.1 SOL on September 29.
OpenAI released GPT 6.1 SOL on September 29.
0:04.720–0:06.237
enSome say this model
Some say this model
0:06.237–0:08.133
enslashed costs by 50% for
slashed costs by 50% for
0:08.133–0:09.460
entheir $200 tier.
their $200 tier.
0:09.460–0:11.393
enOthers say it unceremoniously cut
Others say it unceremoniously cut
0:11.393–0:12.860
entheir subscription usage
their subscription usage
0:12.860–0:14.127
enby 50% while calling it
by 50% while calling it
0:14.127–0:15.327
enan efficiency boost.
an efficiency boost.
0:15.327–0:16.260
enWhy are there two
Why are there two
0:16.260–0:17.460
encompletely opposite
completely opposite
0:17.460–0:18.660
enreactions to the same
reactions to the same
0:18.660–0:19.527
enannouncement?
announcement?
0:19.527–0:20.527
enBecause cheaper,
Because cheaper,
0:20.527–0:21.027
enand,
and,
0:21.027–0:21.960
enmore expensive,
more expensive,
0:21.960–0:23.360
enhappened on the same day.
happened on the same day.
0:23.360–0:24.760
enDepending on whether you
Depending on whether you
0:24.760–0:26.227
enpay per token or per month.
pay per token or per month.
0:26.227–0:27.627
enThis isn't about whether
This isn't about whether
0:27.627–0:28.827
enthe new model is good.
the new model is good.
0:28.827–0:30.227
enIt's about OpenAI saying
It's about OpenAI saying
0:30.227–0:31.693
entwo contradictory things
two contradictory things
0:31.693–0:32.827
ento different users.
to different users.
0:32.827–0:33.327
enToday,
Today,
0:33.327–0:35.460
enwe won't discuss if 6.1 is worth using.
we won't discuss if 6.1 is worth using.
0:35.460–0:37.060
enWe'll clarify three things.
We'll clarify three things.
0:37.060–0:37.560
enFirst,
First,
0:37.560–0:38.693
enwhy can the bills for
why can the bills for
0:38.693–0:39.827
enmodels with the same
models with the same
0:39.827–0:41.493
enlist price differ by 10 times?
list price differ by 10 times?
0:41.493–0:41.993
enSecond,
Second,
0:41.993–0:43.127
enwhich official data
which official data
0:43.127–0:45.127
enmatches independent evaluations,
matches independent evaluations,
0:45.127–0:46.060
enand which do not?
and which do not?
0:46.060–0:46.560
enThird,
Third,
0:46.560–0:47.960
enhow should you determine
how should you determine
0:47.960–0:49.160
enif it is right for you?
if it is right for you?
0:49.160–0:49.660
enOkay,
Okay,
0:49.660–0:51.393
enlet's start with the first one.
let's start with the first one.
0:51.393–0:53.327
enGPT-6.1 is not the flagship model.
GPT-6.1 is not the flagship model.
0:53.327–0:54.356
enIt is an upgraded
It is an upgraded
0:54.356–0:55.921
enversion of the GPT-6 SOL
version of the GPT-6 SOL
0:55.921–0:57.800
enreleased on September 22nd.
released on September 22nd.
0:57.920–0:59.387
enOpenAI's current product
OpenAI's current product
0:59.387–1:00.587
enline looks like this.
line looks like this.
1:00.587–1:01.187
enAt the top,
At the top,
1:01.187–1:02.453
enGPT-6 Ultra is the most
GPT-6 Ultra is the most
1:02.453–1:03.980
enpowerful and expensive.
powerful and expensive.
1:04.300–1:05.189
enIn the middle,
In the middle,
1:05.189–1:07.780
enGPT-6 SOL is slightly weaker and cheaper.
GPT-6 SOL is slightly weaker and cheaper.
1:08.140–1:09.057
enAt the bottom,
At the bottom,
1:09.057–1:11.580
enGPT-6 Lite is the weakest and cheapest.
GPT-6 Lite is the weakest and cheapest.
1:12.220–1:15.240
enThe official positioning for 6.1 SOL is this.
The official positioning for 6.1 SOL is this.
1:15.660–1:16.860
enProviding Near Ultra
Providing Near Ultra
1:16.860–1:18.127
enIntelligence at 1-5th
Intelligence at 1-5th
1:18.127–1:19.193
enthe price of Ultra.
the price of Ultra.
1:19.500–1:20.967
enThe pricing is as follows.
The pricing is as follows.
1:20.967–1:24.200
enUltra is $10 input, $50 output
Ultra is $10 input, $50 output
1:24.200–1:28.040
en6.1 SOL is $2 input, $10 output
6.1 SOL is $2 input, $10 output
1:28.040–1:31.860
enLight is $0.10 input, $0.50 output
Light is $0.10 input, $0.50 output
1:31.860–1:34.368
enIf you calculate Ultra divided by SOL,
If you calculate Ultra divided by SOL,
1:34.368–1:35.700
enit is exactly 5 times
it is exactly 5 times
1:35.700–1:36.567
enThis is what the
This is what the
1:36.567–1:38.033
enofficial statement means
official statement means
1:38.033–1:39.100
enby 1 fifth the price
by 1 fifth the price
1:39.100–1:40.767
enBut that is just the unit price
But that is just the unit price
1:40.767–1:43.920
enThe real cost is a completely different story
The real cost is a completely different story
1:43.920–1:45.100
enWhy?
Why?
1:45.100–1:46.242
enOn September 29th,
On September 29th,
1:46.242–1:47.455
ensomeone on X posted a
someone on X posted a
1:47.455–1:48.739
enchart with a headline
chart with a headline
1:48.739–1:50.380
enmeaning that for the same $2
meaning that for the same $2
1:50.380–1:51.547
en$10.00 price tag,
$10.00 price tag,
1:51.547–1:53.569
enthe bill can differ by 10 times.
the bill can differ by 10 times.
1:53.569–1:55.280
enWhat were they comparing?
What were they comparing?
1:56.080–1:58.574
enGPT 6.1 SOL at $2 input,
GPT 6.1 SOL at $2 input,
1:58.574–1:59.886
en$10 output.
$10 output.
1:59.886–2:03.168
enClaude Sonnet 5.5 at $2 input,
Claude Sonnet 5.5 at $2 input,
2:03.168–2:04.480
en$10 output.
$10 output.
2:04.660–2:06.341
enThe prices are identical,
The prices are identical,
2:06.341–2:07.716
enbut the per-task cost
but the per-task cost
2:07.716–2:09.397
encalculated by artificial
calculated by artificial
2:09.397–2:10.620
enanalysis shows 5.5
analysis shows 5.5
2:10.620–2:11.702
enis $7.60,
is $7.60,
2:11.702–2:14.273
enwhile 6.1 SOL is $0.72.
while 6.1 SOL is $0.72.
2:14.273–2:17.384
enA difference of 10.6 times.
A difference of 10.6 times.
2:17.384–2:17.925
enWhy?
Why?
2:17.925–2:20.360
enBecause Sonnet 5.5 is
Because Sonnet 5.5 is
2:20.360–2:21.120
zh很 wordy
很 wordy
2:21.120–2:25.460
zh它映出了一共193,000典的登場在本項上
它映出了一共193,000典的登場在本項上
2:25.460–2:28.820
zh那是最高的一種的模式測試驗
那是最高的一種的模式測試驗
2:28.820–2:31.160
zh它是60%更多的 Opus 5.5
它是60%更多的 Opus 5.5
2:31.160–2:33.660
zh和7倍的GPT6Ultra
和7倍的GPT6Ultra
2:33.660–2:35.280
en6.1 SOL
6.1 SOL
2:35.280–2:41.220
zhA發現它的登場是10%-30%的高度
A發現它的登場是10%-30%的高度
2:41.220–2:45.320
zh但它是10%的一共1%的上一共1%
但它是10%的一共1%的上一共1%
2:45.320–2:47.320
zh它是比方的一页价值,
它是比方的一页价值,
2:47.320–2:48.720
zh但是因为它说少
但是因为它说少
2:48.720–2:50.400
zh你费同一项供应价
你费同一项供应价
2:50.400–2:51.680
zh但一项供应价格更
但一项供应价格更
2:51.680–2:53.040
zh而另一项供应价格更多
而另一项供应价格更多
2:53.040–2:55.040
zh当然,这些人会比较不同
当然,这些人会比较不同
2:55.040–2:57.200
zh这就是第一个我们必须理解
这就是第一个我们必须理解
2:57.200–2:59.560
zh在这些人的价格上不符定在这些价格上
在这些人的价格上不符定在这些价格上
3:00.440–3:02.160
zh与其中的价格出现
与其中的价格出现
3:02.160–3:03.360
zh我们看看这些 data
我们看看这些 data
3:04.000–3:05.520
zh2从公司的公司
2从公司的公司
3:05.520–3:07.560
zh2从公司的公司
2从公司的公司
3:07.560–3:09.040
zh4个人数在 total
4个人数在 total
3:09.760–3:10.480
zh第一
第一
3:10.480–3:13.680
en75.2 vs 74.1
75.2 vs 74.1
3:13.680–3:14.847
enThere is a chart on
There is a chart on
3:14.847–3:16.326
enOpenAI's release page
OpenAI's release page
3:16.326–3:17.260
enfor SWE Bench.
for SWE Bench.
3:17.980–3:19.380
enIt measures AI's ability
It measures AI's ability
3:19.380–3:20.513
ento complete complex
to complete complex
3:20.513–3:21.780
ensoftware engineering
software engineering
3:21.780–3:23.180
entasks in real codebases.
tasks in real codebases.
3:23.340–3:25.933
enO1 has five reasoning levels,
O1 has five reasoning levels,
3:25.933–3:26.433
enlow,
low,
3:26.433–3:27.073
enmedium,
medium,
3:27.073–3:27.592
enhigh,
high,
3:27.592–3:28.214
enultra,
ultra,
3:28.214–3:28.940
enand max.
and max.
3:29.720–3:31.365
enThe numbers are,
The numbers are,
3:31.365–3:33.480
enlow is 70.0% at $0.17.
low is 70.0% at $0.17.
3:35.100–3:38.380
enMedium is 73.2% at $0.42.
Medium is 73.2% at $0.42.
3:39.280–3:42.460
enHigh is 75.2% at $0.65.
High is 75.2% at $0.65.
3:42.460–3:46.520
enUltra is 74.9% at $0.79
Ultra is 74.9% at $0.79
3:46.520–3:50.980
enMax is 71.9% at $1.57
Max is 71.9% at $1.57
3:50.980–3:53.740
enYou see, the high level is the best
You see, the high level is the best
3:53.740–3:55.680
enThe max level actually regressed
The max level actually regressed
3:55.680–3:56.978
enAnd it costs 2.4 times
And it costs 2.4 times
3:56.978–3:58.420
enmore than the high level
more than the high level
3:58.420–4:00.717
enCompared to Claude 3.5
Compared to Claude 3.5
4:00.717–4:03.980
enSonnet's best at 74.1% for $4.43
Sonnet's best at 74.1% for $4.43
4:03.980–4:05.939
enO1's high level scores
O1's high level scores
4:05.939–4:08.620
enhigher and is 6.8 times cheaper
higher and is 6.8 times cheaper
4:08.620–4:10.608
enThe official claim says it's,
The official claim says it's,
4:10.608–4:11.563
enapproaching,
approaching,
4:11.563–4:12.063
enSonnet
Sonnet
4:12.460–4:14.793
enbut the numbers show it surpasses Sonnet.
but the numbers show it surpasses Sonnet.
4:15.180–4:17.847
enThe official statement downplayed the result.
The official statement downplayed the result.
4:17.847–4:19.544
enThe second number,
The second number,
4:19.544–4:21.800
en71.9% equals 71.9%.
71.9% equals 71.9%.
4:21.800–4:24.120
enThis is O1's ultra and max levels.
This is O1's ultra and max levels.
4:24.840–4:27.225
enThe scores are identical,
The scores are identical,
4:27.225–4:28.309
enboth 71.9%,
both 71.9%,
4:28.309–4:30.260
enbut max costs double.
but max costs double.
4:31.680–4:33.246
enOfficial data tells you,
Official data tells you,
4:33.246–4:34.440
endon't set it to max.
don't set it to max.
4:35.220–4:36.487
enThis is also true on the
This is also true on the
4:36.487–4:38.487
enArtificial Analysis Leaderboard.
Artificial Analysis Leaderboard.
4:38.487–4:39.887
entheir coding agent index
their coding agent index
4:39.887–4:41.258
enshows the ultra level is
shows the ultra level is
4:41.258–4:43.060
enthree points higher than max
three points higher than max
4:43.060–4:44.245
enhigher reasoning
higher reasoning
4:44.245–4:46.300
enintensity isn't always better
intensity isn't always better
4:46.300–4:49.480
enthe third number 52 versus 58
the third number 52 versus 58
4:49.480–4:52.720
enfrom artificial analysis's intelligence index
from artificial analysis's intelligence index
4:52.720–4:53.453
enthis is their
this is their
4:53.453–4:54.920
encomprehensive score from
comprehensive score from
4:54.920–4:56.053
enactual evaluations
actual evaluations
4:56.053–4:58.600
enthe scores are
the scores are
4:58.600–5:01.080
enclaude 3.5 opus at 58
claude 3.5 opus at 58
5:01.080–5:03.640
ensonnet 3.5 at 56
sonnet 3.5 at 56
5:03.640–5:05.120
enextra at 53
extra at 53
5:05.120–5:06.700
enflash at 53
flash at 53
5:06.700–5:09.800
enand 01 at 52.01 is the lowest.
and 01 at 52.01 is the lowest.
5:11.300–5:13.440
enBut the cost is also the lowest.
But the cost is also the lowest.
5:13.660–5:15.540
enPer task, it's 72 cents.
Per task, it's 72 cents.
5:15.800–5:17.673
enSonnet is $3.26,
Sonnet is $3.26,
5:17.673–5:20.080
enand Opus 3.5 is $5.98.
and Opus 3.5 is $5.98.
5:20.500–5:23.080
enClaude 3.5 Sonnet is $7.60.
Claude 3.5 Sonnet is $7.60.
5:23.620–5:25.620
enSo the question isn't which is best,
So the question isn't which is best,
5:25.780–5:27.220
enbut whether your task
but whether your task
5:27.220–5:29.140
enneeds a 58 or if a 52 is enough.
needs a 58 or if a 52 is enough.
5:29.500–5:32.305
enThe fourth number,
The fourth number,
5:32.305–5:34.760
en646 vs 643 vs 728.
646 vs 643 vs 728.
5:34.760–5:37.227
enThis is the big code bench leaderboard score
This is the big code bench leaderboard score
5:37.227–5:38.427
enNot an official list,
Not an official list,
5:38.427–5:39.740
enbut an independent one
but an independent one
5:39.740–5:41.140
enThe results are like this
The results are like this
5:41.140–5:43.660
en01-646 points
01-646 points
5:43.660–5:46.580
enPrevious version, 643 points
Previous version, 643 points
5:46.580–5:49.500
enGPT40, 728 points
GPT40, 728 points
5:49.500–5:50.512
en01 is only 3 points
01 is only 3 points
5:50.512–5:52.400
enhigher than the previous version
higher than the previous version
5:52.400–5:55.060
enAnd it lags behind 40 by 82 points
And it lags behind 40 by 82 points
5:55.060–5:56.243
enThey outright say it
They outright say it
5:56.243–5:58.260
enmoved a lot on their own benchmarks
moved a lot on their own benchmarks
5:58.260–6:00.520
enBut barely moved on independent ones
But barely moved on independent ones
6:00.520–6:02.000
enThe term used is
The term used is
6:02.000–6:04.540
enBenchmarking to the test on their own list
Benchmarking to the test on their own list
6:04.760–6:06.194
enCognition's Frontier
Cognition's Frontier
6:06.194–6:08.080
enCode 1.1 Extended is the same.
Code 1.1 Extended is the same.
6:08.680–6:10.324
en01 is 60.4% and the
01 is 60.4% and the
6:10.324–6:12.954
enprevious version was 60.7%,
previous version was 60.7%,
6:12.954–6:14.160
enalmost flat.
almost flat.
6:14.660–6:15.593
enThe official and
The official and
6:15.593–6:17.127
enindependent leaderboards
independent leaderboards
6:17.127–6:18.727
entell two different stories.
tell two different stories.
6:18.727–6:19.420
enOK, pause.
OK, pause.
6:20.380–6:21.780
enWe now know four numbers.
We now know four numbers.
6:22.220–6:22.720
enFirst,
First,
6:22.720–6:24.440
enit beats Sonnet on some tasks.
it beats Sonnet on some tasks.
6:25.080–6:25.580
enSecond,
Second,
6:25.580–6:27.500
enMax is more expensive and worse.
Max is more expensive and worse.
6:27.700–6:28.200
enThird,
Third,
6:28.200–6:29.600
enthe comprehensive score
the comprehensive score
6:29.600–6:31.067
enis the lowest of the bunch,
is the lowest of the bunch,
6:31.067–6:32.800
enbut the cost is also the lowest.
but the cost is also the lowest.
6:32.800–6:35.260
zh四、インデピンチマーク インプーブメント
四、インデピンチマーク インプーブメント
6:35.260–6:36.260
enWhat do these four
What do these four
6:36.260–6:38.260
ennumbers tell you when put together?
numbers tell you when put together?
6:38.260–6:40.540
en6.1 Sonnet is not a cheap version of Opus.
6.1 Sonnet is not a cheap version of Opus.
6:40.540–6:43.100
enIt is an efficiency model for specific tasks.
It is an efficiency model for specific tasks.
6:43.100–6:44.860
enIt is not an all-rounder.
It is not an all-rounder.
6:44.860–6:46.540
enIt is specialized.
It is specialized.
6:46.540–6:48.700
enSo the question arises.
So the question arises.
6:48.700–6:51.420
enHow do you decide if it's right for you?
How do you decide if it's right for you?
6:51.420–6:54.300
enWe divide scenarios into three categories.
We divide scenarios into three categories.
6:54.300–6:55.202
enCode fixing,
Code fixing,
6:55.202–6:56.104
enbatch tasks,
batch tasks,
6:56.104–6:57.580
enand repetitive work.
and repetitive work.
6:57.580–6:58.780
enIf you have to fix bugs
If you have to fix bugs
6:58.780–6:59.913
endaily or run massive
daily or run massive
6:59.913–7:01.580
enrepetitive knowledge tasks.
repetitive knowledge tasks.
7:01.580–7:03.900
en6.1 sonnet is perfect for this
6.1 sonnet is perfect for this
7:03.900–7:06.080
ena developer conducted an experiment
a developer conducted an experiment
7:06.080–7:09.860
enthey put 105 bugs into 2 code repositories
they put 105 bugs into 2 code repositories
7:09.860–7:12.920
enopus scored 45 and cost 33 dollars
opus scored 45 and cost 33 dollars
7:12.920–7:15.025
en6.1 sonnet scored 44 and
6.1 sonnet scored 44 and
7:15.025–7:17.340
encost 6 dollars and 56 cents
cost 6 dollars and 56 cents
7:17.340–7:22.080
enopus 3.5 scored 41.7 and cost 58 dollars
opus 3.5 scored 41.7 and cost 58 dollars
7:22.080–7:23.461
en6.1 sonnet nearly
6.1 sonnet nearly
7:23.461–7:26.040
enmatched opus at one fifth the cost
matched opus at one fifth the cost
7:26.040–7:28.100
enhowever they noted n equals 1
however they noted n equals 1
7:28.100–7:30.500
enmeaning they only did this experiment once
meaning they only did this experiment once
7:30.500–7:31.233
enOn the Click,
On the Click,
7:31.233–7:33.633
enBenchmark for White Collar Knowledge Work
Benchmark for White Collar Knowledge Work
7:33.633–7:35.455
en6.1 Sonnet scored 80.0,
6.1 Sonnet scored 80.0,
7:35.455–7:37.020
encosting 18 cents
costing 18 cents
7:37.020–7:40.740
enClaud 3.5 scored 81.4, costing 89
Claud 3.5 scored 81.4, costing 89
7:41.195–7:42.365
enFor these tasks,
For these tasks,
7:42.365–7:44.955
enusing 6.1 sonnet is more than enough.
using 6.1 sonnet is more than enough.
7:44.955–7:46.155
enDon't use Macs.
Don't use Macs.
7:46.155–7:47.643
enThe second category,
The second category,
7:47.643–7:48.966
enCreative Visuals,
Creative Visuals,
7:48.966–7:49.466
enGames,
Games,
7:49.466–7:51.115
enand Complex Reasoning.
and Complex Reasoning.
7:51.115–7:52.285
zh如果您做3D scenes,
如果您做3D scenes,
7:52.285–7:53.545
envisual effects,
visual effects,
7:53.545–7:54.715
enor game design,
or game design,
7:54.715–7:56.875
entasks needing imagination,
tasks needing imagination,
7:57.115–7:59.195
en6.1 Sonnet may not be enough.
6.1 Sonnet may not be enough.
7:59.515–8:00.848
enBridgeBench asked it to
BridgeBench asked it to
8:00.848–8:02.315
enrender a sunset seascape.
render a sunset seascape.
8:02.315–8:03.515
enThe conclusion was it
The conclusion was it
8:03.515–8:04.782
enlooked like a sunset in
looked like a sunset in
8:04.782–8:06.315
ena different price bracket,
a different price bracket,
8:06.315–8:08.115
enand the sea was barely rendered.
and the sea was barely rendered.
8:08.115–8:10.248
enSomeone posted a tweet that got over 1,
Someone posted a tweet that got over 1,
8:10.248–8:10.848
en200 likes.
200 likes.
8:10.848–8:12.848
enOPUS 3.5 is much better than GPT 6.1.
OPUS 3.5 is much better than GPT 6.1.
8:12.915–8:15.048
enSomeone else made a comparison video.
Someone else made a comparison video.
8:15.048–8:15.848
enWith the same,
With the same,
8:15.848–8:17.048
enhigh effort setting,
high effort setting,
8:17.048–8:18.382
enthe conclusion was that
the conclusion was that
8:18.382–8:20.315
enOPUS 3 won and 50 Jae-on in quality,
OPUS 3 won and 50 Jae-on in quality,
8:20.315–8:20.815
enbugs,
bugs,
8:20.815–8:21.882
enand color schemes.
and color schemes.
8:21.882–8:24.282
enSONNET is only better in token efficiency.
SONNET is only better in token efficiency.
8:24.282–8:25.482
enOthers have also used
Others have also used
8:25.482–8:27.015
enlogic questions to test it.
logic questions to test it.
8:27.015–8:28.615
en6.1 SONNET set to max failed.
6.1 SONNET set to max failed.
8:28.615–8:31.675
enOPUS 3.5 set to medium passed easily.
OPUS 3.5 set to medium passed easily.
8:31.675–8:32.987
enFor these tasks,
For these tasks,
8:32.987–8:34.675
enstick to CLOD or OPUS.
stick to CLOD or OPUS.
8:34.675–8:36.186
enThe third category,
The third category,
8:36.186–8:37.875
enscientific research.
scientific research.
8:37.875–8:40.075
enThe company admitted this themselves.
The company admitted this themselves.
8:40.075–8:42.951
enOn the Terminal Bench Science benchmark,
On the Terminal Bench Science benchmark,
8:42.951–8:44.485
en6.1 SONNET is 57.0,
6.1 SONNET is 57.0,
8:44.485–8:46.115
enwhile OPUS is 68.1%.
while OPUS is 68.1%.
8:46.115–8:48.273
enthe official word is that the most
the official word is that the most
8:48.273–8:49.891
enchallenging scientific
challenging scientific
8:49.891–8:51.895
enresearch tasks should use opus
research tasks should use opus
8:51.895–8:53.599
enfor these tasks use opus
for these tasks use opus
8:53.599–8:55.815
enand don't try to save a few bucks
and don't try to save a few bucks
8:55.815–8:58.527
enalright now you know 6.1
alright now you know 6.1
8:58.527–9:01.375
ensonnet isn't a cheap opus
sonnet isn't a cheap opus
9:01.375–9:04.095
enit's an efficiency tool for specific tasks
it's an efficiency tool for specific tasks
9:04.095–9:06.946
enif your task aligns it's
if your task aligns it's
9:06.946–9:09.655
engreat if not it's a trap
great if not it's a trap
9:09.655–9:12.115
enand there's one more thing you need to know
and there's one more thing you need to know
9:12.115–9:13.366
enOn September 29,
On September 29,
9:13.366–9:14.350
enthe same day,
the same day,
9:14.350–9:16.495
enOpenAI's TBO posted a tweet.
OpenAI's TBO posted a tweet.
9:16.615–9:17.832
enThe quota for the new
The quota for the new
9:17.832–9:19.264
enPro 200 plan was reduced
Pro 200 plan was reduced
9:19.264–9:20.195
enfrom 20x to 10x.
from 20x to 10x.
9:20.355–9:22.955
enOld users temporarily retained the 20x limit.
Old users temporarily retained the 20x limit.
9:22.955–9:23.822
enOn the same day,
On the same day,
9:23.822–9:25.095
enthey launched Pro 500.
they launched Pro 500.
9:25.355–9:28.095
en$500 per month for a 25x quota.
$500 per month for a 25x quota.
9:28.655–9:30.675
enAPI users saw it as getting cheaper.
API users saw it as getting cheaper.
9:31.155–9:32.488
enSubscription users saw
Subscription users saw
9:32.488–9:34.155
enit as getting more expensive.
it as getting more expensive.
9:34.155–9:36.288
enSo this isn't just a new model release.
So this isn't just a new model release.
9:37.095–9:38.495
enThis is OpenAI repricing
This is OpenAI repricing
9:38.495–9:40.162
enfor different user segments.
for different user segments.
9:40.162–9:41.362
endo you understand now
do you understand now
9:41.362–9:42.488
enwhy some people are
why some people are
9:42.488–9:44.615
encheering while others are cursing
cheering while others are cursing
9:44.615–9:46.015
enbecause they are looking
because they are looking
9:46.015–9:47.348
enat two different things
at two different things
9:47.348–9:49.766
enone last thing if you
one last thing if you
9:49.766–9:53.055
enask me if 6.15 is worth using
ask me if 6.15 is worth using
9:53.055–9:54.352
enmy answer is don't ask
my answer is don't ask
9:54.352–9:56.515
enif it's worth it ask if it's suitable
if it's worth it ask if it's suitable
9:56.515–9:58.715
enit is suitable for code
it is suitable for code
9:58.715–10:00.915
enrepair and batch tasks
repair and batch tasks
10:00.915–10:03.215
enuse the higher tier and turn on max
use the higher tier and turn on max
10:03.215–10:05.335
enfor creative and complex reasoning
for creative and complex reasoning
10:05.335–10:07.655
encurrent tests show it can't beat claude
current tests show it can't beat claude
10:07.655–10:09.695
enContinue using Claude or Opus
Continue using Claude or Opus
10:09.695–10:11.295
enAs for scientific research,
As for scientific research,
10:11.295–10:13.295
enthe official guide already told you
the official guide already told you
10:13.295–10:15.028
enOne more thing about using Opus
One more thing about using Opus
10:15.028–10:16.393
enIf you are a subscriber,
If you are a subscriber,
10:16.393–10:17.855
enyour quota has been cut
your quota has been cut
10:17.855–10:19.188
enYou need to recalculate
You need to recalculate
10:19.188–10:20.388
enif it's enough for you
if it's enough for you
10:20.388–10:21.639
enIf it's not enough,
If it's not enough,
10:21.639–10:23.455
eneither upgrade to Pro500
either upgrade to Pro500
10:23.455–10:26.155
enOr switch to API and pay by token
Or switch to API and pay by token
10:26.155–10:27.505
enThis isn't a question of
This isn't a question of
10:27.505–10:28.855
enwhich model is stronger
which model is stronger
10:28.855–10:31.655
enThis is a question of OpenAI's repricing
This is a question of OpenAI's repricing
10:31.655–10:33.455
enThat's all for today's content
That's all for today's content
10:33.455–10:34.422
enIf you liked it,
If you liked it,
10:34.422–10:35.539
enremember to like,
remember to like,
10:35.539–10:36.060
enfollow,
follow,
10:36.060–10:36.655
enand share
and share
10:36.655–10:38.910
enThis is my motivation to keep creating.
This is my motivation to keep creating.
10:38.910–10:39.935
enSee you next time!
See you next time!

影片筆記：gpt-6.1-sol 发布，五分之一的价格，真能接近 Astra？深入解读！

<!-- VIDEONOTE_PUBLICATION_WARNINGS_BEGIN -->

⚠️ 可先使用，仍有待修正項目

本筆記已完成可閱讀的逐字稿與時間跳轉；下列為待處理問題，建議播放原片確認。

  • 待修正：診斷筆記：目前沒有可可靠理解的逐字稿（asr_language_failure；wrong_language_after_segmented_asr）
  • 待修正：逐字稿品質檢核未完全通過（requested_language_repair_inconclusive；保留 197 段 Whisper 逐字稿，已先行發布。）
  • 待修正：逐字稿語言與宣告語言不一致（已保留來源字幕/Whisper 內容並繼續產生筆記；請確認影片語言。；若語音辨識輸出英文，也會被語言檢查標記，不會覆蓋日文底稿。）
  • 待修正：字幕時間軸待檢（已保留來源字幕並照常發布；未取得足夠可靠錨點。；建議在字幕編輯器重點聽辨時間軸可能漂移區段。）
  • 待修正：逐字稿全文快速校讀未完成（已保留自動字幕/原始逐字稿並繼續生成筆記。；未完成的全文速讀項目可在精進模式回聽確認。）
  • 尚有 1 項待修正細節，請查看工作清單。

<!-- VIDEONOTE_PUBLICATION_WARNINGS_END -->

一句話總結

影片深入解析 OpenAI 發布 GPT-6.1 SOL 後的定價爭議與實際效能，指出該模型在程式除錯與批量任務上具備極高性價比（約為高階模型的 1/5 成本），但在創意視覺、複雜推理與科學研究領域表現不佳，建議用戶應根據具體場景選擇模型，而非盲目追求最高等級。

核心重點

  1. 定價策略的雙面性：OpenAI 對 API 按量付費用戶與訂閱制用戶實施了截然不同的策略。API 用戶感受到成本降低（約 1/5），而訂閱用戶則面臨配額削減（如 Pro 200 從 20x 降至 10x），導致雙方對「變便宜」或「變貴」有截然相反的評價。
  2. GPT-6.1 SOL 的定位：官方宣稱其提供接近 Ultra 的智慧，但價格僅為 Ultra 的 1/5。實際數據顯示，它在程式修復、白領知識工作等批量任務中，能以極低成本達到接近 Opus 的表現，但在高階推理與創意任務上，性能落後於 Claude Opus 3.5 或 GPT-4o。
  3. 效能數據的矛盾：
  • SWE Bench：O1 High 等級表現最佳且性價比高，而 Max 等級不僅價格昂貴，性能甚至出現倒退。
  • 基準測試差異：官方基準測試顯示進步顯著，但獨立第三方基準測試（如 Big Code Bench）顯示幾乎無進步，甚至落後於 GPT-4o。
  • 綜合評分：O1 系列在 Artificial Analysis 綜合評分中最低，但單次任務成本也是最低，顯示其適合「精度要求不高但量大」的場景。
  1. 場景適用性建議：
  • 適合：程式除錯、批量任務、白領知識工作（建議使用 SOL 或 Sonnet 等級，避免使用 Max）。
  • 不適合：創意視覺渲染、複雜邏輯推理、科學研究（建議堅持使用 Claude Opus 3.5 或高階層）。

詳細大綱

一、 定價爭議的核心：兩種對立的敘事

  • 背景：9 月 29 日 OpenAI 發布 GPT-6.1 SOL。
  • 矛盾現象：
  • API 用戶視角：認為成本降低 50% 甚至更多。
  • 訂閱用戶視角：認為訂閱用量削減 50%（配額變少）。
  • 原因分析：OpenAI 對不同用戶群體實施差異化策略。取決於用戶是「按 Token 付費」還是「按月訂閱」。
  • 影片目標：釐清定價邏輯與適用場景，而非單純討論模型好壞。

二、 產品線定位與基礎定價

  • 產品階層與價格：
  1. GPT-6 Ultra：最強、最貴（輸入 $10，輸出 $50）。
  2. GPT-6 SOL：稍弱、較便宜（GPT-6.1 SOL 為 9/22 發布之 SOL 的升級版，輸入 $2，輸出 $10）。
  3. GPT-6 Lite：最弱、最便宜（輸入 $0.10，輸出 $0.50）。
  • 官方說法：GPT-6.1 SOL 提供接近 Ultra 的智慧，價格為 Ultra 的 1/5（$10/$2 = 5 倍）。

三、 實際成本差異：按 Token 付費的陷阱

  • 爭議來源：X 平台圖表顯示，相同標價下，帳單可能相差 10 倍。
  • 數據對比（GPT-6.1 SOL vs. Claude Sonnet 5.5）：
  • Sonnet 5.5：單次任務成本 $7.60。原因：模型被描述為「很 wordy」（冗長），輸出 Token 數高達 193,000。
  • 6.1 SOL：單次任務成本 $0.72。原因：輸出較短。
  • 差異：約 10.6 倍。
  • 結論：按 Token 付費時，需考慮模型輸出的冗長程度，單純比較單價可能產生誤導。

四、 四個關鍵數據指標分析

  1. SWE Bench（軟體工程任務）：
  • O1 各等級表現：Low (70.0%), Medium (73.2%), High (75.2%), Ultra (74.9%), Max (71.9%)。
  • 發現：High 等級表現最佳且成本 ($0.65) 低於 Ultra ($0.79) 和 Max ($1.57)。Max 等級性能倒退且價格是 High 的 2.4 倍。
  • 對比 Claude 3.5 Sonnet：Sonnet 最佳分數 74.1% ($4.43)，O1 High 分數更高且便宜 6.8 倍。
  • 建議：官方數據顯示 Max 等級性價比低，不建議設定為 Max。
  1. Artificial Analysis Leaderboard（編碼代理指數）：
  • Ultra 等級分數比 Max 高 3 分。
  • 結論：更高的推理強度不一定帶來更好結果。
  1. Artificial Analysis Intelligence Index（綜合評分）：
  • 評分：Claude 3.5 Opus (58), Sonnet 3.5 (56), Extra (53), Flash (53), O1 (52)。
  • 發現：O1 綜合評分最低，但單次任務成本 ($0.72) 也是最低（Sonnet $3.26, Opus 3.5 $5.98, Claude 3.5 Sonnet $7.60）。
  • 結論：問題不在於誰最好，而在於任務是否需要 58 分的精度，還是 52 分足夠。
  1. Big Code Bench（獨立基準測試）：
  • 分數：O1 (646), 前版 (643), GPT-4o (728)。
  • 發現：O1 僅比前版高 3 分，落後 GPT-4o 82 分。
  • 對比 Cognition's Frontier Code 1.1 Extended：O1 (60.4%) 與前版 (60.7%) 幾乎持平。
  • 結論：官方基準測試進步大，但獨立基準測試幾乎無進步。官方與獨立排行榜講述兩個不同的故事。

五、 場景適用性建議

  • 核心定義：GPT-6.1 SOL 不是廉價版 Opus，而是針對特定任務的效率模型，非全能型。
  • 場景一：程式除錯與批量任務（適合）
  • 案例：修復 105 個 Bug
  • Opus：得分 45，成本 $33。
  • 6.1 Sonnet：得分 44，成本 $6.56。
  • Opus 3.5：得分 41.7，成本 $58。
  • 結論：6.1 Sonnet 以 1/5 成本接近 Opus 表現。但需注意 N=1 的實驗限制。
  • 白領知識工作：6.1 Sonnet 得分 80.0 ($0.18)，Claude 3.5 得分 81.4 ($0.89)。對於此類任務，6.1 Sonnet 足夠，不需使用 Max。
  • 場景二：創意視覺、遊戲與複雜推理（不適合）
  • 案例：渲染日落海景
  • 結論：看起來像不同價格區間的日落，海洋渲染幾乎未呈現。
  • 對比：Opus 3.5 在品質、Bug 和配色上勝出。Sonnet 僅在 Token 效率上較好。
  • 邏輯測試：6.1 Sonnet 設為 Max 失敗，Opus 3.5 設為 Medium 輕鬆通過。
  • 建議：此類任務應堅持使用 Claude 或 Opus。
  • 場景三：科學研究（不適合）
  • Terminal Bench Science 數據：6.1 Sonnet (57.0%), Opus (68.1%)。
  • 官方建議：最具挑戰性的科學研究任務應使用 Opus，不要為了省錢而嘗試其他模型。

六、 訂閱制用戶的配額變動與策略

  • 9 月 29 日變動：
  • Pro 200 計劃：配額從 20x 降至 10x（老用戶暫時保留 20x）。
  • Pro 500 計劃：新推出，$500/月，配額 25x。
  • 用戶視角差異：
  • API 用戶：覺得變便宜了。
  • 訂閱用戶：覺得變貴了（配額減少）。
  • 最終建議：
  • 不要問是否「值得」，要問是否「適合」。
  • 程式修復與批量任務：適合使用 6.1 Sonnet。
  • 創意與複雜推理：使用高階層並開啟 Max（但測試顯示無法擊敗 Claude），或繼續使用 Claude/Opus。
  • 科學研究：遵循官方建議使用 Opus。
  • 訂閱用戶應對：若配額不足，需重新計算。若不足，可升級至 Pro 500 或切換至 API 按 Token 付費。

工具 / 模型 / 名詞整理

  • OpenAI 產品/模型：
  • GPT-6.1 SOL
  • GPT-6 Ultra
  • GPT-6 SOL
  • GPT-6 Lite
  • GPT-6.1 Sonnet (影片口語混用 SOL 與 Sonnet，指代同一產品線)
  • O1 (系列)
  • O1 Low, Medium, High, Ultra, Max (推理等級)
  • GPT-4o
  • Pro 200 (訂閱計劃)
  • Pro 500 (訂閱計劃)
  • TBO (影片提及 OpenAI 的 TBO 發布推文，疑為特定內部代號或口誤，保留原樣)
  • 競爭對手產品/模型：
  • Claude Sonnet 5.5
  • Claude 3.5 Opus
  • Claude 3.5 Sonnet
  • Opus 3.5
  • Extra (模型名稱)
  • Flash (模型名稱)
  • Cognition's Frontier Code 1.1 Extended
  • 基準測試/數據平台：
  • SWE Bench
  • Artificial Analysis Leaderboard (Coding Agent Index)
  • Artificial Analysis Intelligence Index
  • Big Code Bench
  • Benchmark for White Collar Knowledge Work
  • BridgeBench
  • Terminal Bench Science
  • Cognition's Frontier (基準測試名稱)
  • 平台/媒體：
  • X (Twitter)
  • Click (可能指 ClickBench 或特定測試平台，保留原樣)

操作流程整理

  1. 評估任務類型：
  • 若是程式除錯、批量任務或白領知識工作，考慮使用 GPT-6.1 SOL/Sonnet。
  • 若是創意視覺、複雜推理或科學研究，考慮使用 Claude Opus 3.5 或 OpenAI Ultra/Max。
  1. 選擇定價模式：
  • 若為 API 用戶，關注單次任務成本與輸出 Token 數量（避免冗長模型）。
  • 若為訂閱用戶，檢查配額限制（Pro 200 降至 10x，Pro 500 為 25x），計算是否划算。
  1. 設定模型等級：
  • 對於 GPT-6.1 SOL，建議設定為 High 等級以獲得最佳性價比，避免使用 Max（性能倒退且昂貴）。
  • 對於高階任務，若需極致品質，使用 Opus 3.5 或 Ultra。
  1. 監控與調整：
  • 比較官方基準測試與獨立第三方測試結果。
  • 根據實際輸出成本與效能表現，動態調整模型選擇。

值得注意的限制或風險

  1. Max 等級的性能倒退：在 SWE Bench 測試中，O1 Max 等級的性能（71.9%）低於 High 等級（75.2%），且價格昂貴，不建議使用。
  2. 獨立基準測試落後：在 Big Code Bench 等獨立測試中，O1 系列表現落後於 GPT-4o，顯示官方基準測試可能無法完全反映實際競爭優勢。
  3. 訂閱配額削減：Pro 200 計劃配額從 20x 降至 10x，可能導致訂閱用戶實際可用量大幅減少，需重新評估訂閱方案。
  4. 實驗限制：部分效能比較基於 N=1 的實驗，可能存在偶然性，不具統計學意義。
  5. 模型冗長問題：部分模型（如 Sonnet 5.5）輸出 Token 數過高，導致實際成本遠高於標價，需謹慎評估輸出長度。

逐字稿辨識疑點

  • GPT 6.1 SOL：影片主要使用此名稱，但後段混用 GPT-6.1 Sonnet。需查證 OpenAI 官方是否確實命名為 "SOL" 或為聽寫錯誤（通常為 Sonnet）。
  • TBO：影片提及 "OpenAI's TBO posted a tweet"，此縮寫非公開知名產品代號，疑為口誤或內部代號。
  • 很 wordy / 它映出了一共193,000典的登場在本項上 / 那是最高的一種的模式測試驗 / 它是60%更多的 Opus 5.5 / 和7倍的GPT6Ultra / A發現它的登場是10%-30%的高度 / 但它是10%的一共1%的上一共1% / 它是比方的一页价值 / 但是因为它说少 / 你费同一项供应价 / 但一项供应价格更 / 而另一项供应价格更多 / 当然,这些人会比较不同 / 这就是第一个我们必须理解 / 在这些人的价格上不符定在这些价格上 / 与其中的价格出现 / 我们看看这些 data / 2从公司的公司 / 2从公司的公司 / 4个人数在 total / 第一 / 75.2 vs 74.1：此段為大量聽寫錯誤或語意不清的片段，無法精確還原原意，僅能標記為需查證的雜訊。
  • 四、インデピンチマーク インプーブメント：此處出現日文字符，疑為口誤或語音識別錯誤，可能意指 "Fourth, Independent Benchmark Improvement" 或類似含義。
  • Macs：影片建議 "Don't use Macs"，結合上下文應指 "Max" 等級，疑為口誤。
  • 50 Jae-on：影片提及 "OPUS 3 won and 50 Jae-on in quality"，疑為 "50% better on" 或類似含義的聽寫錯誤。
  • CLOD：影片建議 "stick to CLOD or OPUS"，疑為 "Claude" 的聽寫錯誤。
  • 6.15：影片結尾問 "if 6.15 is worth using"，疑為 "6.1 SOL" 或 "6.1 Sonnet" 的聽寫錯誤。

可延伸追問

  1. OpenAI 官方是否已確認 GPT-6.1 SOL 的正式名稱？還是僅為影片作者的暫定稱呼？
  2. SWE Bench 測試中 Max 等級性能倒退的原因是什麼？是否為模型過度推理導致？
  3. 對於訂閱用戶，Pro 500 計劃的 25x 配額是否足以覆蓋高頻使用場景？
  4. 獨立基準測試（如 Big Code Bench）與官方基準測試差異巨大的原因何在？是否反映評估標準的不同？
  5. 在程式除錯任務中，如何準確評估 N=1 實驗的可靠性？是否有更大規模的測試數據支持？

尚未產生學習筆記

請在 Telegram 指令最後加上「學習」，例如：videonote 網址 英文 雙語 學習