Leaderboard · #28 of 36 · OpenAI · Hardest problems

GPT-6.2 Astra

OpenAI's new top model: 62 on Artificial Analysis's index, second to Claude Fable 5.2's 65 at the same $8 and $40 list price, but on Vals AI's finance test it gets 57.3% right against 56.7% for GPT-6.2 Sol at a quarter of the price.

Updated Nov 18, 2026 · kept by WrenJSON

Key facts

Score
58 / 100, rank #28
Tests72Price19Peoplen/aFit100
Price
$8 to read · $40 to write, per 1M tokens (OpenAI API pricing)
Typical mix
$16 per 1M tokens (three parts reading to one part writing)
Reads at once
1M tokens (about 750,000 words)
Where you can get it
OpenAI API, Amazon Bedrock, ChatGPT Plus, Pro, Enterprise and Edu
Host it yourself
No
Released
Nov 17, 2026
Best for
Problems where a wrong answer costs more than the bill, Research-grade analysis, Reading long contracts and filings

GPT-6.2 Astra costs $8 per million tokens it reads (input) and $40 per million tokens it writes (output) on OpenAI's list price (OpenAI API pricing, read Nov 18, 2026). A token is about three-quarters of a word. See what every model costs for the same job.

Independent tests

TestResultWho ran itRead
Artificial Analysis Intelligence Index (max)62Artificial Analysis LLM leaderboardNov 18, 2026
Vals Finance Index57.31% (rank 22)Vals AI: Vals Finance IndexNov 18, 2026
Vals Tax Agent Bench21.84% (rank 55)Vals AI: Tax Agent BenchNov 18, 2026
Vals Legal Research Bench42.79% (rank 31)Vals AI: Legal Research BenchNov 18, 2026

What a job costs

The cost of doing each job 1,000 times on GPT-6.2 Astra, worked out from OpenAI’s list price and a fixed size for each job. A product built on the model charges its own price on top.

1,000 of theseWhat one job meansCost
Support repliesthe customer message, the account notes and two help articles read; a 300-word reply written$40
Documents processeda two-page invoice or form read; the fields pulled out$28
Calls summariseda 40-minute call transcript read; a summary and next steps written$104
Contracts revieweda 40-page contract read against a checklist; a memo of issues written$300
Drafts writtena brief and one example read; a 600-word email or proposal section written$44

What people running it say

The launch thread on Hacker News, a developer forum, drew 1,283 points and 798 comments in a day, and the early replies are mixed: one commenter calls Sol 'a marginal step forward' after a day, another reports Astra refusing a task on hardware they own, others praise the speed against Gemini 4 Carbon. No reports yet from operations teams.

Caveats

  • The price rises on any request that feeds it more than 272,000 tokens of text at once: reading costs double and writing 1.5 times as much for the whole request, OpenAI's model page says.
  • On Vals AI's tax test it gets 21.8% of tasks right, below GPT-6.2 Sol's 24.8%; on legal research 42.8% against Sol's 37.0%.
  • The ChatGPT plans come from 9to5Mac's report of 17 Nov, which does not mention the Business plan; OpenAI's launch page refused automated reading.
  • Microsoft Foundry, Google Cloud and Microsoft 365 Copilot are not listed because no page read on 18 Nov names GPT-6.2 there.
  • Scores within a point of GPT-6 Astra, which the method treats as a tie: no user reports are scored yet, so its place rests on tests, price and where it is sold.

Questions people ask

How much does GPT-6.2 Astra cost?

GPT-6.2 Astra costs $8 per million tokens it reads (input) and $40 per million tokens it writes (output) on OpenAI's list price (OpenAI API pricing, read Nov 18, 2026). A token is about three-quarters of a word.

What does GPT-6.2 Astra cost for 1,000 support replies?

About $40 at OpenAI's list price, counting one reply as the customer message, the account notes and two help articles read; a 300-word reply written. On the same basis: $28 for 1,000 documents processed, $104 for 1,000 calls summarised, $300 for 1,000 contracts reviewed, $44 for 1,000 drafts written. A product built on GPT-6.2 Astra charges its own price on top.

How good is GPT-6.2 Astra?

GPT-6.2 Astra ranks #28 of 36 on the Models at Work leaderboard with a score of 58 out of 100 (November 2026 edition). OpenAI's new top model: 62 on Artificial Analysis's index, second to Claude Fable 5.2's 65 at the same $8 and $40 list price, but on Vals AI's finance test it gets 57.3% right against 56.7% for GPT-6.2 Sol at a quarter of the price.

What is GPT-6.2 Astra best for?

GPT-6.2 Astra is best for problems where a wrong answer costs more than the bill, research-grade analysis, reading long contracts and filings.

How does GPT-6.2 Astra do on independent tests?

Artificial Analysis Intelligence Index (max): 62 (Artificial Analysis LLM leaderboard); Vals Finance Index: 57.31% (rank 22) (Vals AI: Vals Finance Index); Vals Tax Agent Bench: 21.84% (rank 55) (Vals AI: Tax Agent Bench); Vals Legal Research Bench: 42.79% (rank 31) (Vals AI: Legal Research Bench).

How much can GPT-6.2 Astra read at once?

GPT-6.2 Astra can take in 1M tokens (about 750,000 words) in one go, according to OpenAI. Vendors call this the context window.

Where can a business get GPT-6.2 Astra?

GPT-6.2 Astra is available through OpenAI API, Amazon Bedrock, ChatGPT Plus, Pro, Enterprise and Edu.

Can a company host GPT-6.2 Astra itself?

No. GPT-6.2 Astra is not released for self-hosting; it is available through OpenAI and the products and clouds that carry it.

Compare GPT-6.2 Astra

How the score works: Score is 0 to 100 and computed, not typed. 40% is independent testing (Artificial Analysis, Arena, Scale, Epoch, Vals and others; each test is scaled against the best score on this board, and tests of real work, such as finished freelance projects or professional tasks, count double the exams). 30% is price: the vendor's list price for a typical mix of reading and writing, on a fixed scale from $0.05 to $60 per million tokens, so a model that costs a fifth as much gains real ground. 15% is what people report after running it, from Signals and linked threads. 15% is fit: how much text it can take in at once and where a business can get it (the products and clouds that carry it), with a small credit for open weights. A missing component drops out and the rest are reweighted. A model with no independent test yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted. Our coverage: read the article.