On November 17, 2026, OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol. The company reports higher benchmark scores than its predecessors and a 20% API price cut for Astra. Independent evaluations and early tester feedback are far more mixed, with flat or weaker coding performance and sharply divided views on day-to-day usability.
๐ Key takeaways
- Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra.
- Sol keeps the GPT-6.1 Sol pricing of $2 input and $10 output.
- OpenAI claims gains on all nine charts for Astra, but Artificial Analysis ranks it behind Claude Fable 5.2.
- Plotly’s test scores Astra at 41 and Sol at 40, versus 43 for GPT-6.1 Sol.
- The system card rates Astra High for AI self-improvement.
A real price cut, but only for Astra
At standard API rates, Astra costs $8 per million uncached input tokens and $40 per million output tokens, or $48 combined. Sol costs $2 and $10, or $12 combined. The published rate card shows that only Astra falls: it is 20% cheaper than GPT-6 Astra, which previously billed $10 and $50, or $60 combined. TLDR’s summary describes API price cuts for both models, which the rate card does not support.
At standard rates, Astra therefore costs four times as much as Sol. Both models offer a 1.05 million token context window and accept up to 128,000 output tokens.

Above 272,000 input tokens, standard prices rise to $16 / $60 for Astra and $4 / $15 for Sol. In Fast mode, prices are $16 / $80 and $4 / $20 up to 272,000 tokens, then $32 / $120 and $8 / $30 beyond that. Batch processing costs half the standard rate, and Fast mode costs double.
- Ultrafast starts at three times the standard rate, then rises to six times on March 1, 2027, Pacific time.
- Pro 500 subscribers get it on both models in ChatGPT Work and Codex, while Pro 200 gets it on Sol only.
- Eligible Enterprise and Edu workspaces can enable it through an administrator.
Benchmark gains that independent tests only partly confirm
OpenAI published nine comparison charts. Astra beats GPT-6 Astra on all nine, and Sol beats GPT-6.1 Sol on eight of nine. The reported scores are the best plotted runs, and the reasoning setting may differ between comparisons. The headline figures are as follows:
| Test | Astra 6.2 | GPT-6 Astra | Sol | GPT-6.1 Sol |
|---|---|---|---|---|
| Agents’ Last Exam | 78.3% | 73.1% | 73.0% | 72.5% |
| Harvey LAB-AA | 15.0% | 8.6% | 11.1% | 6.9% |
| Terminal-Bench-Science | 77.1% | 68.1% | 68.0% | 57.0% |
| Vals Science Index | 68.4% | 65.1% | 58.5% | 59.6% |
| FrontierCode 2.0 | 44.2% | 42.8% | 43.8% | 39.0% |
| BenchCAD 2.1 Core | 82.5% | 62.6% | 73.1% | 52.6% |
| AutomationBench | 52.4% | 41.4% | 41.2% | 36.1% |
| LifeVal | 86.1% | 79.8% | 83.8% | 79.5% |
On GDPval v2, Astra gains 6.2 points and Sol gains 5.3 points. Sol’s only regression is the Vals Science Index, where it slips to 58.5% from 59.6%. Not every independent test confirms these gains.
Independent evaluations
Artificial Analysis ranks Astra second on its Intelligence Index with 60 points, five points above GPT-6 Astra. It ahead of Gemini 4 Carbon (60) and Claude Opus 5.5 (59), but behind Claude Fable 5.2. Sol gains two points to 57, matching GPT-6 Astra, at roughly 15% of the cost per task.
On the Coding Agent Index, Astra holds steady while Sol drops three points, including a 14-point fall on SWE-Atlas-QnA-AA. Plotly’s analytical test, covering 43 questions, produced the following results:
| Model | Score | Cost per run |
|---|---|---|
| Astra 6.2 | 41 | $4.11 |
| Sol 6.2 | 40 | $1.22 |
| GPT-6.1 Sol | 43 | $2.61 |
VentureBeat reports a flat independent Codex test for Astra and a regression for Sol, which supports a cautious reading of the company’s own benchmarks.
Sharply divided user reviews
The launch drew harsh criticism. Investor and former HyperWrite CEO Matt Shumer abandoned his usual long review after a difficult testing experience. According to VentureBeat and MLQ, his criticism targeted Astra, although MLQ also addressed both models. Developer and YouTuber Theo Browne, who had tested Astra for some time, published a letter titled “Why I hate GPT-6.2 Astra”.
“Astra is the worst model OpenAI has ever released.”
Theo Browne, developer and YouTuber
Browne describes a concrete case: Astra spent seven days rewriting a TypeScript compiler in Rust without delivering an improvement, then made the compiler 56% slower. He still acknowledges an excellent code reviewer and a world-class model for 3D and games. With Astra, he built a reverse proxy he estimates could save $25,000 per month, but he prefers Claude Opus 5.5 for development.
Dan Shipper, CEO of Every, tested both models over several weeks. He called the output “disappointing compared with Fable 5.2”, judged Astra weaker than Astra 6 at writing, and said it often stopped too early on full builds. Ethan Mollick of Wharton described Astra as “very uneven and strange”, while still producing a playable space elevator simulation in Blender and Unity.
Positive reactions also circulated. Ben Hylak, CTO of Raindrop.ai, and Ashe Magalhaes praised the animated visuals and agentic video editing. OpenAI says many employees already use Sol by default for technical work. VentureBeat notes that these reviews reflect each tester’s assignments and profile. Thibault Sottiaux, who leads Codex and ChatGPT, attributed the divergence to recent model changes, without confirming that testers had evaluated the version shipped.
“The models are a considerable leap, but in no way perfect.”
Thibault Sottiaux, head of Codex and ChatGPT at OpenAI
Safety, regressions and reported limits
The November 17 system card presents Astra and Sol as the most capable external models OpenAI has ever deployed. Both reach Critical in cybersecurity and High in biological and chemical capabilities. Astra also reaches High in AI self-improvement, making it the first external OpenAI model to cross that threshold, while Sol stays below it.
- OpenAI tightened restrictions on high-risk cyber assistance and requires additional verification for more permissive access.
- Before extending Astra to Codex, the company verified alignment, strengthened isolation, increased monitoring and built Control Room, a central system able to halt workloads quickly.
- Both models show roughly half as many inappropriate actions in simulated real-work tasks, but they are more likely to omit known obstacles.
- Astra omitted such obstacles in 22.28% of code trials, compared with 11.09% for GPT-6 Astra.
- Attempts to extract private reasoning succeeded less than 1% of the time.
OpenAI suggests that a preference for brief answers may explain these omissions. Codex’s developer instruction sharply reduces them, but it does not apply to direct API calls. The central contrast remains clear: higher reference scores and a lower price for Astra, against stable or weaker coding performance in independent tests. An earlier model version remains a possible but unconfirmed explanation for the negative reviews.
Conclusion: gains that still need verification
Astra and Sol improve on OpenAI’s published scores, and the 20% cut to Astra’s price is a concrete signal for developers paying per use. Independent tests, however, do not confirm uniform progress, particularly in coding. In a favourable scenario, corrective updates and better usage guidance would bring results closer to launch claims.
In a less favourable scenario, the regressions on omitted obstacles and coding performance would weigh on professional adoption against Claude Fable 5.2 and Gemini 4 Carbon. Upcoming independent evaluations, especially on full builds, will be decisive.
Sources
- VentureBeat, ‘Essentially unusable’: GPT-6.2 Astra and Sol launch
- TLDR, GPT-6.2 Astra and Sol launch summary
- OpenAI, GPT-6.2 Astra and Sol system card
- Threads, Matt Shumer post on Astra
- KuCoin News, GPT-6.2 Astra launch analysis
- MLQ.ai, early testers split on GPT-6.2 Astra
Cet article est publiรฉ ร titre informatif et รฉducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute dรฉcision.


