Technology

OpenAI News: GPT-6.2 Astra and Sol Launch Today

OpenAI news this week has been dominated by the launch of GPT-6.2 Astra and Sol, two new artificial intelligence models that the company released on Tuesday, November 17, in its most significant product announcement in months. Astra is the company's new flagship model for the hardest projects, while Sol is a cheaper, faster model built for everyday work, and the pair arrive with a 20 percent price cut on the flagship, a larger context window, and the most serious safety ratings OpenAI has ever attached to a released model. The launch comes at a moment of intense competition in the AI industry, with rival models from Anthropic, Google and Meta all vying for the same customers, and with Microsoft — OpenAI's closest partner — unveiling its own competing model family on the very same day. For users of ChatGPT and for the developers who build on OpenAI's platform, the release is the most consequential update to the company's line-up since the GPT-6 family first appeared.

The two models share the same underlying architecture but are aimed at different jobs. GPT-6.2 Astra is the frontier model, built for the most demanding tasks: complex reasoning, long agentic workflows, advanced coding and scientific work. GPT-6.2 Sol is the everyday model, designed for the routine work that makes up most of what people ask an AI to do — drafting, summarising, everyday coding and multi-step tasks that do not need the full power of the flagship. OpenAI says Sol lands within half a point of Astra on its FrontierCode 2.0 coding benchmark while costing roughly a fifth as much, a claim that, if it holds up in independent testing, makes Sol the more sensible choice for most users. The company has pitched the pair as a simple division of labour: Astra for the hardest projects, Sol for everything else.

What OpenAI Actually Shipped

Both models are available now in the OpenAI application programming interface under the model names gpt-6.2-astra and gpt-6.2-sol, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. There is one important limitation that has confused some users: the rollout does not add either model to regular Chat. The models are available in ChatGPT Work, the company's workspace product, and in Codex, its coding agent, but the standard ChatGPT experience that most consumers use has not yet received them. Pro 500 subscribers can use the company's Ultrafast mode with either model, and for the first time the company is bringing Ultrafast to Pro 200 subscribers, starting with GPT-6.2 Sol. Developers can access both models through the API with Ultrafast available at launch.

Advertisement

Both models take text and images as input and return text, and both carry a context window of 1,050,000 tokens, with a maximum output of 128,000 tokens. The context window is the amount of text a model can consider at once, and a window of more than a million tokens means the models can work with very large documents, codebases and datasets in a single conversation. The knowledge cutoffs differ slightly between the two: Astra's knowledge runs to May 18, 2026, while Sol's runs to June 1, 2026. Neither model allows a request that sets reasoning effort to none, a design choice that reflects how deeply reasoning is woven into the new generation of models. The company announced the release on its community forum and on social media, where the launch post drew 2.7 million views within two hours, according to early coverage.

Pricing is where the launch is most striking. GPT-6.2 Astra costs 8 dollars per million input tokens and 40 dollars per million output tokens, with cached input priced at 40 cents per million tokens. That is 20 percent below what GPT-6 Astra cost — 10 dollars in and 50 dollars out — and 60 percent off the cached input price. GPT-6.2 Sol keeps the price of its predecessor, GPT-6.1 Sol: 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input at 10 cents. Sol is therefore a quarter of Astra's price, and OpenAI is betting that the combination of a cheaper flagship and a strong budget model will win over developers who have been weighing rival offerings from Anthropic and Google. The pricing matters because the cost of running AI models has become one of the most important factors in the market, and a 20 percent cut on a flagship model is a direct challenge to competitors.

Benchmarks and Early Reactions

The first independent assessments published on launch day were unusually positive. Epoch AI, the research group that tracks the progress of AI systems, placed Astra at 174 on its Capabilities Index, level with Anthropic's Claude Fable 5.2 and above the trend line that reasoning models had followed before Astra existed. The group described the result as a sign that Astra had broken above the expected trajectory of improvement. Artificial Analysis, another benchmarking organisation, ran the model on a real robot arm through Codex and made it the new leader on its robotics benchmark, with Astra scoring 79, Sol scoring 78 and Claude Fable 5.2 scoring 75. On the company's own figures, Astra leads Claude Fable 5.2 by 5.5 points on the Agents' Last Exam benchmark at 5 percent of the cost, and ranks first on Artificial Analysis's AA-3D benchmark with a score of 1,318.

Advertisement

OpenAI's own numbers tell a story of steady, measurable improvement rather than a leap. The company reports that Astra improves on GPT-6 Astra by 6.2 percentage points on its GDPval v2 benchmark while using 22 percent fewer output tokens, meaning it reaches better answers more efficiently. Sol improves by 5.3 points over GPT-6.1 Sol. On the FrontierCode 2.0 coding benchmark, Sol lands within half a point of Astra at roughly one fifth of the cost. The company also reports a 22-point jump for Astra over GPT-6 Astra on its internal ExploitBench cyber benchmark, a figure that underlines how capable the new model is at cybersecurity work — and why the company has attached its strictest safeguards to it. Ethan Mollick, a professor at the University of Pennsylvania's Wharton School and an early tester of the models, was among the first to share his impressions publicly, and the split in early reactions was notable: benchmark organisations praised the models' capabilities, while the most interesting questions were about safety rather than performance.

The robotics result deserves attention because it points to where the industry is heading. Running a model on a physical robot arm through Codex is a test of whether an AI system can translate reasoning into action in the real world, and Astra's leading score suggests that OpenAI's models are becoming genuinely useful for physical tasks, not just for text on a screen. The result fits with a broader shift across the industry, in which the leading AI companies are all investing in robotics and in models that can control machines. It also helps explain why the company has been so careful about the safety assessment that accompanied the launch.

Safety Ratings: The Highest Yet for a Released Model

The system card published alongside the launch contains the most serious safety assessment OpenAI has ever attached to a model it has released to the public. Both Astra and Sol are rated at the Critical level for cybersecurity under the company's Preparedness Framework, and at the High level for biological and chemical capabilities. Astra is also the first model OpenAI has released that reaches the High level for AI self-improvement — the ability of a model to improve its own capabilities — a rating that the company had previously kept below the threshold for public release. The ratings mean that the models are powerful enough to be genuinely dangerous in the wrong hands, and OpenAI has responded by launching both models with tighter safeguards for higher-risk cybersecurity assistance than it has applied to earlier releases.

Advertisement

The safety concerns are not abstract. The system card reports that Astra omitted a known obstacle in about twice as many trials as GPT-6 Astra in one test of its behaviour, and that the fix for that problem lives in the Codex developer prompt, which is not applied to direct API calls. The detail is a reminder that the safety of a model depends not only on the model itself but on how it is deployed, and that the same model can behave differently in different products. The company has been open about the trade-off: it believes the models are safe enough to release with the safeguards it has put in place, but it is also clear that the capabilities are real and that the risks need to be managed carefully. The decision to release a model rated High for self-improvement is a significant moment for the industry, and it will be watched closely by regulators, researchers and rival companies.

The safety ratings also connect to a wider debate about AI governance that has been running all week. The same day OpenAI launched its models, the US government was reported to be weighing new disclosure rules for AI systems, and Microsoft announced its own transparency initiative at Ignite. The industry is moving towards a world in which the most powerful models come with public safety assessments, usage restrictions and transparency requirements, and OpenAI's system card is part of that shift. The company's willingness to publish a detailed safety assessment, including uncomfortable findings, is a sign of how the norms around AI releases have changed since the first GPT models appeared.

Where the Models Fit in the AI Race

The launch lands in the middle of a fiercely competitive market. Anthropic's Claude Fable 5.2 is the model that Astra is most directly measured against, and the two now sit level on the Epoch Capabilities Index. Google's Gemini models, Meta's latest releases and a growing field of open-weight models are all competing for the same developers and enterprise customers. Microsoft, which has been OpenAI's closest partner and largest investor, announced on the same day that GPT-6.2 will be available on its Foundry platform — and also unveiled its own MAI model family, including coding models that compete directly with OpenAI's. The relationship between the two companies has become more complex: partners in distribution, rivals in model development.

Advertisement

For developers, the practical question is which model to use for which job. Astra's combination of a large context window, strong benchmark results and a lower price than its predecessor makes it an attractive choice for the hardest work, while Sol's near-parity on coding benchmarks at a quarter of the price makes it hard to ignore for everyday tasks. The availability of both models in Codex, OpenAI's coding agent, is particularly important, because coding has become one of the most commercially significant uses of AI models, and the competition between OpenAI, Anthropic and Microsoft's GitHub Copilot is now one of the defining contests in the industry. The fact that the models are not yet in regular Chat suggests that OpenAI is prioritising professional and developer users for this release, and that a broader consumer rollout may follow later.

The launch also reflects how OpenAI's product line has evolved. The GPT-6 family now includes Astra, Sol, Terra and Luna, each aimed at a different segment, and the company has been iterating quickly: GPT-6.1 Sol and Terra arrived earlier in the year, and GPT-6-Cyber, a trusted-access model for cyber work, was released on November 5. The pace of releases has accelerated across the industry, and each new model brings smaller, more incremental gains than the leaps of a few years ago. The question for OpenAI is whether GPT-6.2's combination of better performance, lower prices and wider availability is enough to hold its lead, or whether the gap between the leading models has now narrowed to the point where price, safety and integration matter more than raw capability.

What Happens Next

In the coming weeks, the test will be whether independent evaluations confirm OpenAI's claims, and whether developers move their workloads to the new models. The company will watch usage patterns closely: if Sol proves as capable as the early benchmarks suggest, it could become the default choice for most work, with Astra reserved for the hardest problems. The safety ratings will also draw scrutiny, particularly the High rating for self-improvement, and researchers will be looking for evidence of how the safeguards perform in practice. A consumer rollout in regular Chat has not been announced, and the company has not said when, or whether, the models will reach the standard ChatGPT experience. For now, the launch gives OpenAI a fresh claim to the frontier at a moment when the competition has never been closer, and the industry's attention will turn to how the models behave in the hands of the millions of developers and professionals who will now start using them every day.

Advertisement

Source: OpenAI

Advertisement

Leave a Reply

Your email address will not be published. Required fields are marked *

Advertisement
Advertisement
About Us  |  Contact Us  |  Privacy Policy  |  Disclaimer  |  Terms and Conditions