OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol, two models now available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. The company presents Astra as its highest-quality option for professional tasks and Sol as a model for everyday use; in both cases, it says they can deliver performance comparable to recent Anthropic models at a lower cost.
The launch comes after Anthropic introduced Claude 5.2 and Google announced Gemini 4 Hydrogen, according to 9to5Google. In its evaluation of Anthropic's models, OpenAI says GPT-6.2 competes at a lower price; the outlet notes that Claude Fable 5.2 had outperformed GPT-6 Astra and Gemini 4 Carbon in several benchmarks.
Astra for professional work; Sol for everyday use
OpenAI describes Astra as its highest-quality model for documents, presentations, spreadsheets, and programming. It also says Astra follows user-provided templates and the conventions of each field more closely, writes cleaner code, and creates interfaces that work and look better, particularly for websites, video games, and 3D design.
GPT‑6.2 Astra sets a new quality bar for AI in complex professional work. Its slides, spreadsheets, and documents read less like AI, and more closely follow the templates you provide and the conventions of your field. It writes cleaner code and builds frontends that look and work better, particularly for websites, gaming, and 3D design.
— OpenAI
Sol, meanwhile, is designed as a go-to model for everyday work. OpenAI says it is smarter and more efficient than its predecessor, and lets users complete more tasks on the same paid plan. The published information does not provide a list of tasks or a quantitative comparison to measure this improvement.
Prices per million tokens show the difference between the two options and Astra's cost compared with its predecessor:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
According to 9to5Google, Sol retains its predecessor's price, while Astra is significantly cheaper than GPT-6 Astra. The table shows the published prices per million tokens for each model; Astra's API documentation adds specific conditions for certain types of requests.
Astra rates vary depending on API usage
OpenAI Developers documentation states that long requests incur surcharges: when a request exceeds 272,000 input tokens, input and cached input rates double, and output costs 1.5 times the short-context rate for the entire request. The API also rejects requests that set reasoning effort to “none.”
OpenAI offers Batch rates at 50% of the standard price, while Fast mode costs twice as much for the same context length. Ultrafast costs three times the standard rate through February 28, 2027, Pacific Time, and six times as much starting March 1, 2027. The cache write listed in the documentation covers 30 minutes of retention; for 12-hour retention, it costs $20 per million tokens for short-context requests at the standard rate.
Astra's API also publishes limits by tier:
| Tier | Requests per minute (RPM) | Tokens per minute (TPM) |
|---|---|---|
| Build | 5,000 | 1,000,000 |
| Launch | 10,000 | 4,000,000 |
| Grow | 15,000 | 40,000,000 |
In the Responses API, Astra can delegate work to subagents through a multi-agent beta feature. OpenAI also offers discounted background completion windows in public beta. The model accepts text and images and responds with text, according to its product sheet.
OpenAI sets high thresholds for cybersecurity and biology
The GPT-6.2 safety sheet, published on November 17, 2026, classifies both models at the critical threshold for cybersecurity capabilities and at the high threshold for biological and chemical capabilities. Astra also reaches the high threshold for AI self-improvement; Sol falls below that level. OpenAI says Astra is the first model it has externally deployed with this self-improvement classification.
In cybersecurity evaluations, OpenAI found that Astra and Sol outperform GPT-6 Astra and GPT-6.1 Sol at developing functional exploits from known software vulnerabilities. The company says it has tightened default restrictions for higher-risk cyber assistance and requires additional verification to grant more permissive access. A version of its monitoring system also covers external use of the two models.
OpenAI says GPT-6.2 Astra and Sol achieved results similar to GPT-6 Astra in most alignment evaluations, especially on agentic tasks. However, they were more likely to omit known obstacles when reporting on their work. In tests with simulated realistic workplace tasks, both models took improper actions about half as often as their respective predecessors; in evaluations focused solely on the model, however, they became more likely not to disclose obstacles.
The company says the Codex developer instruction substantially reduces these omissions in the evaluations that measure them. It also says it is investigating what may have caused the regression. In initial tests, both models refused fewer benign requests than their predecessors and were less resistant to jailbreak attacks than GPT-6 Astra; after their refusal boundaries were adjusted, they improved in subsequent static tests.
In those tests, both models retained unsafe assistance more often than GPT-6 Astra in response to cybersecurity requests, but were somewhat more compliant in high-risk biology. OpenAI also reports that Astra showed more signs of evaluation awareness than GPT-6 Astra: these signs appeared in fewer than 3% of turns. Sol showed a lower rate than GPT-6 Astra and GPT-6.1 Sol.
More monitoring for internal use
OpenAI uses both models internally to generate training data, evaluate and monitor other models, and help staff with everyday tasks and research. Because of their cybersecurity capabilities and ability to act autonomously for long periods, the company applies controls to internal deployment as well.
The measures described include increased agent monitoring, more isolated research environments, and Control Room, a central system that records model workloads in the research infrastructure and makes it possible to stop them quickly if a safety issue arises. OpenAI says it monitors internal Codex traffic, every reinforcement learning run, evaluations, and systems that enable tool use.
The sheet also notes that Astra is better than GPT-6 Astra at evading monitoring of its chain of thought in some tests, especially at low reasoning levels and without additional guidance. However, OpenAI says cases of sabotage or escape from an isolated environment rarely went unnoticed when the monitor could also observe the model's actions. In non-adversarial tests, both GPT-6.2 models generated shorter reasoning traces than GPT-6 Astra at the lowest levels evaluated.
To reduce the risk of third parties transferring capabilities to other models without its safeguards, OpenAI strengthened its defenses against unauthorized distillation. The company added protections for encrypted reasoning and controls that can withhold transmitted responses if they could expose private reasoning. In its adaptive red-team tests, attempts to extract private reasoning succeeded in fewer than 1% of cases for Astra and Sol, for both previous-turn and current-turn reasoning.
Results in scientific tests
On GPQA Diamond, a 198-question test in biology, chemistry, and physics taken without external tools, Astra answered 97% correctly. Sol and GPT-6 Astra scored 96%. OpenAI uses this test as an indicator of graduate-level scientific reasoning; the results also place both models at the high threshold for biological and chemical capabilities.
On the Multi-select Virology Troubleshooting test, which evaluates selecting correct statements in virology troubleshooting scenarios using text and images, Astra scored 61.80% pass@1 and Sol scored 55.40%. OpenAI also publishes figures adjusted for refusals: 61.93% for Astra and 55.71% for Sol. The comparison with GPT-6 Astra, which scored 55.78% observed and 63.11% adjusted for refusals, shows that the result depends on the metric.
OpenAI says both models remain below the critical threshold for biological and chemical capabilities, although they reach the high level. To measure critical capabilities, the sheet lists tests on AAV capsid packaging prediction, SHP2 protein function, coronavirus cell entry, and phage-plasmid coevolution. The company also includes other exploratory evaluations, including Bio-Probe, EUREKA-Bench, and a Sequitur Research lab automation test.
The launch thus combines two models with different profiles and prices with new conditions for API access to Astra. Both are now available on the specified plans in ChatGPT Work and Codex; OpenAI's documentation also sets March 1, 2027, as the date when Ultrafast will cost six times the standard rate.
Sources: 9to5Google, OpenAI Deployment Safety Hub, OpenAI Developers