OpenAI released two models on November 17, 2026: GPT-6.2 Astra, its new top model for demanding projects, and GPT-6.2 Sol, a cheaper model for everyday work. Both are available in ChatGPT Work and Codex on paid plans from Plus upward and in the API. OpenAI’s help center says the rollout “does not add either model to regular Chat”, and Free and Go plans do not get them.
In the API, Astra costs $8 per million input tokens and $40 per million output, the same rates Anthropic lists for Claude Fable 5.2 (both checked November 18, 2026). OpenAI’s own tests put Astra ahead of Fable 5.2 on several professional benchmarks. The independent Artificial Analysis index gives a more mixed picture. We have not tested the models; the figures below come from OpenAI, Anthropic, and Artificial Analysis.
GPT-6.2 Astra and Sol at a glance
- Released: November 17, 2026
- Model IDs: gpt-6.2-astra and gpt-6.2-sol
- Where: ChatGPT Work on web, mobile, and desktop; Codex app and CLI; OpenAI API
- Inputs: text and images; output is text
- Context: up to 1 million tokens; Codex defaults to 272,000 for Astra and 500,000 for Sol
What is new
1. Two models with different jobs
Astra is meant for the hardest work, such as tracing a bug across several services or building a financial model from source files. Sol is the default for everyday tasks such as editing code or drafting a report, and new chats in ChatGPT Work start on it, according to OpenAI’s model guide. OpenAI says both models finish tasks with fewer output tokens than their predecessors.
2. Documents, slides, and spreadsheets
OpenAI says it trained Astra to follow the conventions of fields like law, finance, and accounting, with more direct writing. In slides, Astra can match a template’s layout, colors, and fonts from the original file or a screenshot. In spreadsheets, OpenAI’s example is a valuation workbook laid out so a reader can follow how the estimate is built.
3. Code, websites, and 3D
OpenAI reports cleaner code and more focused tests from early testers, more precise computer use in CAD and simulation tools, and better visual judgment in web design. On AA-3D, where human judges compare blinded 3D pages built from the same prompt, Astra ranks first with 1,318 Elo, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
Who can use GPT-6.2
| Plan | Access | Notes |
|---|---|---|
| Free and Go | No | Not included in any surface |
| Plus ($20 a month) | Work and Codex | Not in regular Chat |
| Pro ($100, $200, $500) | Work and Codex | Faster Ultrafast mode on some Pro tiers |
| Business | Per OpenAI’s pricing docs | The announcement does not list Business |
| Enterprise and Edu | Once an admin enables them | Off by default |
| OpenAI API | gpt-6.2-astra and gpt-6.2-sol | Available at launch |
Source: OpenAI announcement and ChatGPT Work and Codex pricing, November 18, 2026.
Pricing: token price versus task price
| Per million tokens | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 |
| Claude Fable 5.2 | $8.00 | $0.20 (cache hits) | $40.00 |
Sources: OpenAI API pricing and Anthropic pricing, November 18, 2026.
Equal token rates do not mean equal bills, because models spend different numbers of tokens on the same task. Sol is exactly four times cheaper than Astra per token. The “about five times cheaper” figure often quoted for Sol is something else: OpenAI’s cost per task on one coding test, Cognition’s FrontierCode 2.0, where it says Sol scored within 0.5 points of Astra.
What OpenAI’s benchmarks claim
The announcement states most results as differences between models rather than as scores. The table lists them as OpenAI gives them.
| Benchmark (what it tests) | OpenAI’s claim |
|---|---|
| Agents’ Last Exam (long multistep professional work) | Astra 5.5 points ahead of Claude Fable 5.2 at 5% of the cost |
| Harvey LAB-AA v1.1 (legal deliverables) | Astra more than doubles Fable 5.2’s pass rate at about one sixth of the cost |
| Terminal-Bench-Science 0.1 (science work in a terminal) | Astra 9 points above GPT-6 Astra at under a third of the cost |
| GDPval v2, internal (long knowledge-work projects) | Astra up 6.2 points from GPT-6 Astra with 22% fewer output tokens |
| FrontierCode 2.0 (merge-ready code changes) | Sol within 0.5 points of Astra at about one fifth of the cost |
OpenAI’s footnotes add limits. The Agents’ Last Exam result for Fable 5.2 leaves out nine tasks. On Terminal-Bench-Science, OpenAI used a third-party run of Fable 5.2 in a different setup, while Anthropic reported 74.6% in Claude Code. GDPval v2 is internal, so outsiders cannot rerun it.
What independent testing shows
Artificial Analysis combines nine evaluations across knowledge work, coding, and science into its Intelligence Index v5.0.1 (Astra, Sol, checked November 18, 2026).
| Model and effort | Index score | Cost per index task | Minutes per task |
|---|---|---|---|
| GPT-6.2 Astra, Medium | 57 | $1.62 | 3.8 |
| GPT-6.2 Astra, Xhigh | 62 | $2.52 | 6.5 |
| GPT-6.2 Astra, Max | 62 | $4.00 | 11.5 |
| GPT-6.2 Sol, Medium | 53 | $0.41 | 3.1 |
| GPT-6.2 Sol, Max | 57 | $1.01 | 7.9 |
| Claude Fable 5.2, High | 62 | $9.08 | 14.9 |
| Claude Fable 5.2, Max | 65 | $26.54 | 41.1 |
Fable 5.2 at Max has the highest score in the index. At a score of 62, Astra at Xhigh cost about 28% of Fable 5.2 at High per task, and Astra at Max about 44%. The savings depend on the setting.
An equal index score is also not proof of equal quality. The index is a rounded average of nine tests, so two models with the same number can do well on different things, and the cost per task is a weighted average that will not match your workload. These numbers say which settings are worth testing first, not which model wins on your tasks.
A finding to check before trusting agent output
The most practical warning sits in the system card. In two of OpenAI’s tests, both models were more likely than GPT-6 Astra to leave out obstacles that blocked their work, such as a missing file or an unavailable search tool. A report can therefore look complete when part of the job was not done. OpenAI says new instructions in Codex substantially reduce this and that it is still investigating. In ChatGPT Work or your own agent, ask the model to list anything it could not access or finish, and check the output against your sources.
The system card also rates both models at OpenAI’s Critical level for cybersecurity capability, and Astra is the first released model at its High level for AI self-improvement. The added safeguards also block some legitimate security research, which OpenAI directs to its Daybreak program.
What to try first
For people on Plus or a higher plan, the change happens in Work and Codex. A reasonable test is to give Astra one demanding task you already know well, such as a deck in a company template, and compare the result with what you get now; for drafts and routine edits, Sol uses less of your allowance. Our explainers on GPT-6 Astra and GPT-6 Sol and Luna cover the earlier models.
For API teams, the published data supports running your own tasks on Astra at Xhigh and on Sol before moving a budget. It does not yet show that either model beats Fable 5.2 on your work.
What is still unclear
| Question | Status on November 18, 2026 |
|---|---|
| Astra versus Fable 5.2 on real tasks | Only vendor tests and an aggregate index so far |
| GPT-6.2 in regular Chat | Not added by this rollout; no date given |
| Business plan access | Listed in pricing docs, missing from the announcement |
| Exact benchmark scores | Mostly given as differences between models |
| Fix for omitted obstacles outside Codex | Still under investigation |