GPT-6.2 Astra and Sol: OpenAI’s New Models Do More Work for Less Money
It was past midnight in a London consulting office when the analyst finally gave up on her spreadsheet. The client deck was due at eight in the morning, the numbers would not reconcile, and the AI assistant she had been leaning on all evening kept burning through her plan’s allowance before the work was done.
She is exactly the kind of user OpenAI had in mind on Monday, when it released two new models, GPT-6.2 Astra and GPT-6.2 Sol. The company says both get more done before a subscriber reaches their limit, and the bigger one, Astra, costs a fifth less than the model it replaces.
GPT-6.2 Astra is OpenAI’s new frontier model, built for the hardest professional projects. GPT-6.2 Sol is its efficient sibling, designed for everyday coding and agent work. Both are available from today in ChatGPT Work and Codex, and through the API as gpt-6.2-astra and gpt-6.2-sol.
There is one catch for casual users. The rollout covers ChatGPT Work and Codex only, for Plus, Pro, Enterprise and Edu subscribers. Regular Chat is not getting either model yet, so the millions of people on the free tier will not see them appear in their model picker.
What GPT-6.2 Astra Brings to Professional Work
Astra is the headliner of the pair, and OpenAI is pitching it on a simple idea: judge a model by what it costs to finish a job, not by what it costs per token.
On its own GDPval v2 test, built from multi-step consulting, finance and design assignments, Astra scores 6.2 points higher than GPT-6 Astra while using 22 percent fewer output tokens.
On Agents’ Last Exam, a test of demanding professional work, Astra beats its predecessor by 5.2 points at less than a third of the cost.
OpenAI says it also scores 5.5 points ahead of Anthropic’s Claude Fable 5.2 at about five percent of the cost, a comparison that will draw plenty of attention from rival labs.
Independent testing tells a slightly more modest story. Artificial Analysis, which runs its own model tests, ranks Astra second on its Intelligence Index, behind Claude Fable 5.2 and ahead of Google’s Gemini 4 Carbon and Anthropic’s Claude Opus 5.5. Second place, in other words, but at a price its rivals cannot match.
The price is where Astra makes its boldest move. Through the API it costs eight dollars per million input tokens, forty cents for cached input and forty dollars per million output tokens. That is twenty percent below GPT-6 Astra, and the cached-input price has fallen by sixty percent.
At those rates, Astra now costs exactly the same per token as Claude Fable 5.2, which Anthropic cut by twenty percent only a week earlier. The two labs are now charging identical prices for their flagship models, and the fight has moved squarely onto who can finish more work per dollar.
GPT-6.2 Sol: the Everyday Workhorse
Sol is the quieter half of the launch, but for many users it may matter more. OpenAI describes it as the efficient model for everyday coding and agent workflows, and it keeps exactly the same price as GPT-6.1 Sol: two dollars per million input tokens, ten cents cached, and ten dollars per million output tokens.
What changed is the capability underneath that price. Sol’s knowledge now runs to June 1, 2026, up from April 30, and it keeps the same 1,050,000-token context window with a 128,000-token output limit. On OpenAI’s own launch charts it beats its predecessor on all five evaluations.
The most striking of those is ExploitBench Internal Port, a cybersecurity test, where Sol scores about 51 percent against 22 percent for GPT-6.1 Sol. On Cognition’s FrontierCode 2.0, Sol lands within half a point of Astra itself at roughly a fifth of the cost, and on OpenAI’s LifeVal test it beats GPT-6 Astra at about a seventh of the cost.
There is also a small mystery attached to Sol. Users reported from November 8 that ChatGPT seemed to be routing them to a model called GPT-6.2 Sol, nine days before OpenAI named it. The company has not said whether those sessions were running this model.
The Safety Report Card: High for Self-Improvement
The most talked-about part of the launch is not a benchmark or a price. It is the safety report card that came with it. According to Unite.AI, Astra is the first model OpenAI has ever released externally that meets the High threshold for AI self-improvement under its Preparedness Framework.
That category measures whether a model can meaningfully help build better AI systems, including by improving itself. OpenAI says Astra cleared that bar, while Sol did not. Both models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities.
The numbers behind those labels are stark. In safeguard-off tests on ExploitBench Internal Port, Astra achieved a 53.5 percent mean arbitrary-code-execution rate at maximum reasoning, and Sol reached 50.5 percent.
GPT-6 Astra managed 31.5 percent on the same test.
Because of those increased capabilities, OpenAI says it is launching both models with tighter safeguards for higher-risk cybersecurity assistance. The company has not published the full details of those safeguards, but the message is clear: these models are more capable, and OpenAI knows they are more dangerous in the wrong hands.
That tension is not new. OpenAI has been building safety controls into its workplace products for months, including the Control Room kill switch for rogue AI models it announced earlier this month. GPT-6.2 arrives as those controls are being tested in the real world.
What Changes in December
The announcement carries a change that subscribers may feel more than any benchmark. From December, Plus, Pro and Business plans will share a single allowance across ChatGPT and Codex, and that allowance will be measured in tokens rather than by message counts or separate limits.
That is why OpenAI is suddenly talking about tokens per task. If your allowance is counted in tokens, a model that finishes the same job with 22 percent fewer output tokens is not just faster. It is literally more work for the same money, and it stretches the same plan further.
Ultrafast, OpenAI’s speed tier, is also widening. Pro 500 subscribers get Ultrafast with either new model, and Pro 200 subscribers get it for the first time, starting with GPT-6.2 Sol. Developers get Ultrafast in the API for both models at launch.
What Happens Next
The timing of this launch is no accident. Anthropic is pricing its own flagship IPO this week, with trading expected to begin on November 19, and Google’s Gemini 4 Carbon only arrived a week ago.
OpenAI has chosen to answer both rivals on the same day, with a cheaper frontier model and a safety disclosure that guarantees headlines.
Whether Astra lives up to the pitch will depend on what happens outside OpenAI’s own test suite. Independent benchmarks, real developer feedback and the first enterprise deployments will show whether the token savings hold up on messy, real-world work, or only on the tests OpenAI chose.
For now, the direction of travel is unmistakable. Models are getting cheaper per task, the allowance is becoming the product, and the safety questions are getting harder.
GPT-6.2 Astra and Sol are OpenAI’s bet that it can win on all three fronts at once, and the next few weeks will show whether that bet pays off.



