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

逐字稿文字預覽

OpenAI released GPT 6.1 SOL on September 29.
Some say this model slashed costs by 50% for their $200 tier.
Others say it unceremoniously cut their subscription usage by 50% while calling it an efficiency boost.
Why are there two completely opposite reactions to the same announcement?
Because cheaper, and, more expensive, happened on the same day.
Depending on whether you pay per token or per month.
This isn't about whether the new model is good.
It's about OpenAI saying two contradictory things to different users.
Today, we won't discuss if 6.1 is worth using.
We'll clarify three things.
First, why can the bills for models with the same list price differ by 10 times?
Second, which official data matches independent evaluations, and which do not?
Third, how should you determine if it is right for you?
Okay, let's start with the first one.
GPT-6.1 is not the flagship model.
It is an upgraded version of the GPT-6 SOL released on September 22nd.
OpenAI's current product line looks like this.
At the top, GPT-6 Ultra is the most powerful and expensive.
In the middle, GPT-6 SOL is slightly weaker and cheaper.
At the bottom, GPT-6 Lite is the weakest and cheapest.
The official positioning for 6.1 SOL is this.
Providing Near Ultra Intelligence at 1-5th the price of Ultra.
The pricing is as follows.
Ultra is $10 input, $50 output
6.1 SOL is $2 input, $10 output
Light is $0.10 input, $0.50 output
If you calculate Ultra divided by SOL, it is exactly 5 times
This is what the official statement means by 1 fifth the price
But that is just the unit price
The real cost is a completely different story
Why?
On September 29th, someone on X posted a chart with a headline meaning that for the same $2
$10.00 price tag, the bill can differ by 10 times. What were they comparing?
GPT 6.1 SOL at $2 input, $10 output. Claude Sonnet 5.5 at $2 input, $10 output.
The prices are identical, but the per-task cost calculated by artificial analysis shows 5.5
is $7.60, while 6.1 SOL is $0.72. A difference of 10.6 times. Why? Because Sonnet 5.5 is
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
6.1 SOL
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
There is a chart on OpenAI's release page for SWE Bench.
It measures AI's ability to complete complex software engineering tasks in real codebases.
O1 has five reasoning levels, low, medium, high, ultra, and max.
The numbers are, low is 70.0% at $0.17.
Medium is 73.2% at $0.42.
High is 75.2% at $0.65.
Ultra is 74.9% at $0.79
Max is 71.9% at $1.57
You see, the high level is the best
The max level actually regressed
And it costs 2.4 times more than the high level
Compared to Claude 3.5 Sonnet's best at 74.1% for $4.43
O1's high level scores higher and is 6.8 times cheaper
The official claim says it's, approaching, Sonnet
but the numbers show it surpasses Sonnet.
The official statement downplayed the result.
The second number, 71.9% equals 71.9%.
This is O1's ultra and max levels.
The scores are identical, both 71.9%, but max costs double.
Official data tells you, don't set it to max.
This is also true on the Artificial Analysis Leaderboard.
their coding agent index shows the ultra level is three points higher than max
higher reasoning intensity isn't always better
the third number 52 versus 58
from artificial analysis's intelligence index
this is their comprehensive score from actual evaluations
the scores are
claude 3.5 opus at 58
sonnet 3.5 at 56
extra at 53
flash at 53
and 01 at 52.01 is the lowest.
But the cost is also the lowest.
Per task, it's 72 cents.
Sonnet is $3.26, and Opus 3.5 is $5.98.
Claude 3.5 Sonnet is $7.60.
So the question isn't which is best,
but whether your task needs a 58 or if a 52 is enough.
The fourth number, 646 vs 643 vs 728.
This is the big code bench leaderboard score
Not an official list, but an independent one
The results are like this
01-646 points
Previous version, 643 points
GPT40, 728 points
01 is only 3 points higher than the previous version
And it lags behind 40 by 82 points
They outright say it moved a lot on their own benchmarks
But barely moved on independent ones
The term used is
Benchmarking to the test on their own list
Cognition's Frontier Code 1.1 Extended is the same.
01 is 60.4% and the previous version was 60.7%, almost flat.
The official and independent leaderboards tell two different stories.
OK, pause.
We now know four numbers.
First, it beats Sonnet on some tasks.
Second, Max is more expensive and worse.
Third, the comprehensive score is the lowest of the bunch, but the cost is also the lowest.
四、インデピンチマーク インプーブメント
What do these four numbers tell you when put together?
6.1 Sonnet is not a cheap version of Opus.
It is an efficiency model for specific tasks.
It is not an all-rounder.
It is specialized.
So the question arises.
How do you decide if it's right for you?
We divide scenarios into three categories.
Code fixing, batch tasks, and repetitive work.
If you have to fix bugs daily or run massive repetitive knowledge tasks.
6.1 sonnet is perfect for this
a developer conducted an experiment
they put 105 bugs into 2 code repositories
opus scored 45 and cost 33 dollars
6.1 sonnet scored 44 and cost 6 dollars and 56 cents
opus 3.5 scored 41.7 and cost 58 dollars
6.1 sonnet nearly matched opus at one fifth the cost
however they noted n equals 1
meaning they only did this experiment once
On the Click, Benchmark for White Collar Knowledge Work
6.1 Sonnet scored 80.0, costing 18 cents
Claud 3.5 scored 81.4, costing 89
For these tasks, using 6.1 sonnet is more than enough.
Don't use Macs.
The second category, Creative Visuals, Games, and Complex Reasoning.
如果您做3D scenes, visual effects, or game design, tasks needing imagination,
6.1 Sonnet may not be enough.
BridgeBench asked it to render a sunset seascape.
The conclusion was it looked like a sunset in a different pric

來源：transcript.txt

本次執行流程與問題處理

下達指令
指令：videonote https://youtu.be/yi-m_iGnGx0?si=FgymJ0Bi1qF8Mhvv --language zh

來源：https://youtu.be/yi-m_iGnGx0?si=FgymJ0Bi1qF8Mhvv

辨識語言：zh ｜參數：--language zh
預期流程
依指令產生的處理順序：
1. 下載與語音辨識
2. 來源選擇與品質檢查
3. 文字校正與句子單元
4. 筆記與發布
5. 公開網頁發布
1. 下載與語音辨識
工具：videonote-transcribe-url（主辨識管線：下載、Whisper 辨識與選稿）
處理：下載音訊、以 Whisper 辨識並選出最佳逐字稿
產出：yt_dlp_meta.json（影片元資料）、audio_16k_mono.wav（16kHz 單聲道音訊）、transcript.whisper.json（Whisper 辨識結果）、transcript.json（正式逐字稿）、transcript_selection_report.json（來源選擇報告）
耗時：132.8 秒
關係：依序執行
結果：完成
2. 全文快速校讀
工具：videonote_rapid_transcript_review.py
處理：先速讀完整逐字稿，再修正高可信突兀詞句並列出待確認項目
產出：rapid_transcript_review_report.json（產出檔案）
耗時：41.8 秒
關係：依序執行
結果：完成
3. 建立句子單元
工具：videonote-build-sentence-units
處理：將逐字稿切成帶時間軸的句子單元
產出：sentence_units.json（句子單元）
耗時：0.8 秒
關係：依序執行
結果：完成
4. 句子密度平衡
工具：videonote_balance_sentence_units.py
處理：平衡句子長度與密度
產出：sentence_units.json（句子單元）
耗時：0.8 秒
關係：依序執行
結果：完成
5. 時間軸稽核
工具：videonote_timeline_audit.py
處理：檢查時間軸與音訊對齊
產出：timeline_audit_report.json（時間軸稽核）
耗時：0.9 秒
關係：依序執行
結果：完成
6. 品質檢查
工具：videonote_controlled_quality.py（檢查時間軸與品質）
處理：檢查語言、時間軸與可讀性品質
產出：controlled_quality_report.json（品質報告）、language_sanity_report.json（語言檢查）
耗時：0.9 秒
關係：依序執行
結果：完成
7. 播放時間軸回填
工具：videonote_rebase_display_timeline.py
處理：將網頁播放時間軸回填為校正字幕時間
產出：display_timeline_rebase_report.json（時間軸回填報告）
耗時：1.0 秒
關係：依序執行
結果：完成
8. 時間軸整合
工具：videonote_timeline_reconcile.py
處理：整合字幕、Whisper 與 VAD 時間證據
產出：timeline_evidence.json（時間軸證據）
耗時：1.0 秒
關係：依序執行
結果：完成
9. 逐字稿交叉檢查
工具：videonote_cross_check_transcript.py
處理：交叉檢查逐字稿與來源字幕
產出：transcript_crosscheck_report.json（交叉檢查）
耗時：1.0 秒
關係：依序執行
結果：完成
10. 來源句整理
工具：videonote_prepare_source_entries.py
處理：整理來源句子供時間軸與頁面使用
產出：sentence_units.json（句子單元）
耗時：0.0 秒
關係：依序執行
結果：完成
11. 生成筆記摘要
工具：videonote-generate-note（生成筆記摘要）
處理：生成筆記摘要與重點
產出：note.md（筆記 Markdown）、note.html（筆記 HTML）
耗時：154.9 秒
關係：依序執行
結果：完成
12. 建立筆記頁面
工具：videonote-render-html（產生筆記 HTML）
處理：產生可閱讀的筆記 HTML 頁面
產出：note.html（筆記 HTML）
耗時：1.0 秒
關係：依序執行
結果：完成
13. 發布公開頁面
工具：videonote-publish（發布公開頁面）
處理：發布到公開網站與每日列表
產出：公開筆記網址、result.json（管線結果）
耗時：1.1 秒
關係：依序執行
結果：完成

手動來源字幕

reference_subtitle.txt 下載9,494 bytes
reference_subtitle.vtt 下載86,809 bytes

自動來源字幕

source_subtitles.clean.json 下載54,130 bytes
subtitle_acquisition.json 下載1,312 bytes

Whisper 字幕

尚未產生

Qwen 字幕

尚未產生

Parakeet 字幕

尚未產生

MOSS 字幕

尚未產生

聲學時間證據（PyAnnote / WhisperX）

尚未產生

校正前後逐字稿

校正前 transcript.json 純文字 JSON 8,730 / 67,443 bytes
校正後 transcript.json 純文字 JSON 8,730 / 67,443 bytes

目前待解決問題與初步改善方向

診斷筆記：目前沒有可可靠理解的逐字稿待修正
asr_language_failure；wrong_language_after_segmented_asr；宣告語言與辨識結果不一致；原始證據已保留。；筆記仍已建立並可進入精進模式；未捏造影片摘要或翻譯。
逐字稿品質檢核未完全通過待修正
requested_language_repair_inconclusive；保留 197 段 Whisper 逐字稿，已先行發布。
逐字稿語言與宣告語言不一致待修正
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。
字幕時間軸待檢待修正
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。
逐字稿全文快速校讀未完成待修正
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。；RAPID_REVIEW_STATUS=REVIEW_REQUIRED_INVALID_OUTPUT
逐字稿品質檢核未完全通過待修正
{'stage': 'canonical_transcript_quality', 'action': 'published_with_warning', 'signals': ['transcript_selection_not_resolved', 'asr_language_mismatch:zh']}
Qwen3-ASR 未完成，需確認 qwen_asr_lab 環境與音訊後重試精進模式。

進一步精進字幕與筆記
使用 Workbench 上傳校正字幕、中文參考稿或回饋 SRT，重新產生雙語與學習單元。
開啟 Workbench

逐字稿文字預覽

本次執行流程與問題處理

下達指令

預期流程

1. 下載與語音辨識

2. 全文快速校讀

3. 建立句子單元

4. 句子密度平衡

5. 時間軸稽核

6. 品質檢查

7. 播放時間軸回填

8. 時間軸整合

9. 逐字稿交叉檢查

10. 來源句整理

11. 生成筆記摘要

12. 建立筆記頁面

13. 發布公開頁面