The short version
● Astra $8 per million input, $40 output. A 20% cut from GPT-6 Astra, and the first OpenAI flagship to launch cheaper than the model it replaces.
● Sol $2 in, $10 out. Unchanged from GPT-6.1 Sol, and OpenAI says many of its own staff run Sol as their default.
● Where ChatGPT Work and
Codex for Plus, Pro, Enterprise and Edu. Both in the API. Neither is in regular Chat.
● The catch Artificial Analysis puts Astra second behind Claude Fable 5.2, and several well-known testers who usually praise OpenAI releases did not this time.
OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on Tuesday 17 November. The model IDs are gpt-6.2-astra and gpt-6.2-sol. The interesting part is not the benchmark table, it is that Astra costs 20% less than the model it replaces while scoring roughly the same on independent indexes, and that the people who test these models hardest came away unimpressed.
What it costs
| Model | Input / M | Output / M | Change |
|---|
| GPT-6.2 Astra | $8.00 | $40.00 | 20% cheaper than GPT-6 Astra |
| GPT-6.2 Sol | $2.00 | $10.00 | No change from GPT-6.1 Sol |
Input pricing is for uncached tokens. For context on where that sits against the rest of the market, our AI API pricing tracker carries the per-million rates for every current frontier model.
Who can actually use it today
Access is narrower than a normal OpenAI launch, and the rollout is staged:
- ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu. Not regular Chat.
- Pro 500 subscribers get Ultrafast on both models.
- Pro 200 subscribers get Ultrafast on Sol first.
- API access to both, Ultrafast included.
Availability also depends on workspace permissions and how far the rollout has reached your account, so an Enterprise seat is not a guarantee the models are visible yet. If you are comparing plans before paying, our ChatGPT listing tracks what each tier includes.
The benchmarks, from both sides
OpenAI's own numbers are genuine improvements, and mostly about cost rather than raw capability:
- Agents' Last Exam — Astra at xhigh scores 5.2 points above GPT-6 Astra at max, at less than half the cost per task.
- ExploitBench Internal Port — Astra gains 22 points.
- FrontierCode 2.0 — Sol comes close to Astra on merge-ready changes to real codebases at roughly one-fifth the cost.
- OpenAI composite — 170.1, against 168.4 for GPT-6 Astra.
Independent measurement is flatter. Artificial Analysis scored Astra 62 on its Intelligence Index, its best result on that index, which places it second overall behind Claude Fable 5.2. Epoch AI had Astra and Fable tied at 174 against 169 for GPT-6 Astra, with uncertainty ranges that overlap. On the Coding Agent Index, Astra was flat and Sol dropped three points, most of that traced to a 14-point fall on SWE-Atlas-QnA-AA where it more often left out key facts about a codebase.
The one standout is Harvey LAB-AA, a legal-work evaluation, where OpenAI's chart puts Astra at 15.0% and says that more than doubles Fable's pass rate.
Why experienced testers pushed back
This is the part worth reading before you migrate anything. Several people whose model reviews are widely followed reported the same class of problem, which is control rather than capability.
Matt Shumer, an investor and HyperWrite's former CEO, said the models “were essentially unusable” and skipped his usual written review. He found Astra underperformed as a sub-agent, with Claude supervisors redoing its work.
Theo Browne called Astra the worst model OpenAI has released. On a compiler rewrite it spent seven days without an improvement and then produced a version 56% slower; a second attempt aimed at speed stalled for four days. He said long projects were hard to keep moving and hard to audit, and that he still reaches for Claude Opus 5.5 for development work.
Ethan Mollick of Wharton described Astra as very uneven and harder to steer than rival models, and said it was the first OpenAI frontier release since GPT-4.5 he did not read as a clear step forward.
Dan Shipper of Every reported Astra stopping early on end-to-end builds, pushing brevity so hard that explanations went missing, and producing design output with a house style that was difficult to vary.
It was not unanimous. Ben Hylak and Ashe Magalhaes both rated Astra highly for motion graphics, VFX, animated visuals and agentic video editing — which is a fair signal about where its gains landed.
Two things in OpenAI's own system card line up with the complaints: the model can fail to disclose obstacles it hits, and tighter cybersecurity controls can block legitimate vulnerability research. Browne noted OpenAI fixed some of his issues before launch.
What to do about it
- Already paying for GPT-6 Astra — switch the model ID. Same work, 20% less, and the independent scores are level or better.
- Running agents with a supervisor model — test Astra as the sub-agent before you commit. That is the exact shape of workload the critics described failing.
- Long autonomous coding runs — keep your current model until you have run your own multi-day test. Two separate testers reported stalls measured in days.
- Cost-sensitive routine work — Sol at $2/$10 is the value pick, and OpenAI says its own staff default to it. Check the SWE-Atlas regression if you query it about codebases.
- Video and motion work — this is where the positive reports cluster. Worth a trial.
Where these land against everything else shipping this month is in our November model ranking, and the coding-specific picture is in best AI coding models.
FAQ
How much does GPT-6.2 Astra cost?
$8 per million uncached input tokens and $40 per million output tokens. That is 20% below GPT-6 Astra, making it the first OpenAI flagship to arrive cheaper than its predecessor rather than more expensive.
Is GPT-6.2 Sol cheaper than GPT-6.1 Sol?
No. Sol stays at $2 per million input and $10 per million output, unchanged. The price movement in this release was all on Astra.
Can I use GPT-6.2 in normal ChatGPT?
Not at launch. The rollout covers ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu accounts, plus the API. OpenAI did not add either model to regular Chat.
Is GPT-6.2 Astra better than Claude Fable 5.2?
Not on independent measurement. Artificial Analysis scored Astra 62 and placed it second behind Fable 5.2, and Epoch AI had the two tied at 174 with overlapping uncertainty. Astra does lead on Harvey LAB-AA, a legal evaluation, at 15.0%.
Why are developers calling GPT-6.2 unusable?
The complaints are about steering rather than scores: stopping early on long tasks, being hard to direct, not reporting obstacles, and underperforming as a sub-agent under another model's supervision. Matt Shumer, Theo Browne, Ethan Mollick and Dan Shipper all reported versions of this. Reviewers working on motion graphics and video rated it well.
Should I migrate from GPT-6 Astra to GPT-6.2 Astra?
For ordinary request-response work, yes — it is the same capability for 20% less. For multi-day autonomous runs or agent hierarchies, run your own test first, because that is where the reported failures concentrate.