OpenAI Launches GPT-6.2 Astra and Sol, Astra Costs 20% Less

OpenAI made GPT-6.2 Astra and GPT-6.2 Sol available on Tuesday in ChatGPT Work, Codex, and its API, with a 20% price reduction for Astra. The company claims advancements in handling complex tasks, but tests from Epoch AI and developer Theo Browne yield more mixed results.
Epoch AI and Theo Browne Raise Concerns
The research group Epoch AI subjected the models to a manual evaluation based on its scientific activities. While Astra ranks approximately with Fable 5.2 in achieving a new record on the Epoch Capabilities Index, its average on realistic tasks reaches 65%, 17 points behind Anthropic's model. None of the tested models meet the standards that Epoch applies to its staff. For the analysis of an AI data center, the scores are 98% for Fable 5.2, 88% for Opus 5.5, and 65% for Astra. Epoch criticizes the latter for excessive caution and missing estimates despite sufficient elements, while praising its more concise writing and superior graphic design compared to GPT-6 Astra.
Developer Theo Browne, who tried Astra before its release, published a letter he claims to have sent to OpenAI. He appreciates its code review capabilities, as well as its applications in 3D, gaming, and computing, but does not recommend it despite some improvements made by OpenAI. According to him, seven days spent rewriting his TypeScript compiler yielded no improvements and even resulted in a 56% slowdown. He prefers Anthropic's Opus 5.5 for his development work and describes Astra as very intelligent but cumbersome to use on real projects.
OpenAI, for its part, claims that Astra holds the top position globally among scientists. In the final exam for agents, a multi-step professional test, the company reports a 5.2-point advantage over GPT-6 Astra for less than a third of the cost. It also claims a 5.5-point lead over Claude Fable 5.2, with a cost per task equivalent to 5% of that competitor. Anthropic launched Fable 5.2 on November 9.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
More Frequent Omissions and Cybersecurity Restrictions
OpenAI's security assessments reveal setbacks: both models more frequently omit certain blocks, particularly when a file is missing. The company indicates it is investigating the causes of these omissions. Astra also shows more than GPT-6 Astra the suspicion of being subject to evaluation, in less than 3% of cases. In some adversarial tests, it also performs better at circumventing the monitoring of its thought chain.
OpenAI further claims that Astra is the first model it has launched to reach the high threshold of AI self-improvement defined by its Preparedness Framework. It compares this contribution to that of an additional highly skilled research engineer for each of its researchers. According to the company, Astra and Sol also meet its critical threshold in cybersecurity, like GPT-6 Astra, and come with enhanced restrictions for high-risk cyber requests. Defenders retain access to GPT-6 Cyber in the Daybreak program for their authorized testing.
Pricing and Access to the Two Models
Astra is priced at $8 per million incoming tokens and $40 per million outgoing tokens in the API. The price for cached inputs drops to $0.40, a 60% reduction. Sol remains at the rates of GPT-6.1 Sol: $2 for input and $10 for output per million tokens. OpenAI also states that both new models require fewer outgoing tokens to perform demanding tasks.
The company targets Astra for the most challenging projects, while the less expensive Sol aims at everyday use. This generation marks the first update of Astra since the abandonment of the launch of GPT-6.1 Astra in September. The Plus, Pro, Enterprise, and Edu plans include access to both models. The Pro 200 subscription grants access to the Ultrafast mode for the first time, starting with Sol.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.