Brief IA
›

OpenAI Launches GPT-6.2 Astra with 20% Price Drop

Models & LLM·Tom Levy·

OpenAI Launches GPT-6.2 Astra with 20% Price Drop

OpenAI Launches GPT-6.2 Astra with 20% Price Drop
Key Takeaways
1OpenAI presented GPT-6.2 Astra and Sol on November 17, 2026.
2Astra achieves a High level in OpenAI's internal evaluation of AI self-improvement.
3OpenAI announces a gain of 6.2 points on GDPval v2 for Astra and 22% fewer tokens generated.
4Astra's rates decrease by 20% for both input and output compared to GPT-6 Astra.
Why it matters : These models are deployed in ChatGPT Work, Codex, and the API for professional use, promising increased efficiency and reduced costs.

OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, announcing a 20% reduction in entry and exit prices for Astra compared to GPT-6 Astra. The company promises better performance with lower token consumption and assigns Astra a new level of capability in its assessment of AI self-improvement.

OpenAI Ranks Astra at High Level

In the category dedicated to AI self-improvement, Astra achieves a High level according to OpenAI. The company claims that none of its public models had previously reached this ranking. The capability in question pertains to accelerating certain tasks aimed at designing future AIs.

This assessment, conducted under specific conditions, does not demonstrate an ability to freely transform its own architecture. It does not further establish that Astra could independently trigger training or choose its evolution. To monitor its sensitive research, OpenAI notably uses Control Room.

Sol Promises Lower Programming Costs

For demanding tasks, OpenAI targets Sol at users who do not wish to pay the price of its most powerful model. The company emphasizes programming: during the FrontierCode 2.0 evaluation, Sol would achieve a performance level close to that of GPT-6 Astra, while showing a cost per task approximately five times lower. Those who frequently use programming agents could benefit from this gap, but this advantage still needs to be confirmed in real-world use cases. The underlying model does not solely determine the quality of such an agent.

La lettre du soirGratuit · 19 h

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

For Astra, OpenAI also announces a 60% reduction on already cached entries. Off-cache, the announced standard API prices for one million tokens drop from $10 to $8 for input, and from $50 to $40 for output compared to GPT-6 Astra. Sol remains priced at $2 for inputs and $10 for outputs, on this same billing unit. These amounts may fall under specific tiers.

A company will not necessarily achieve a 20% savings on its total expenditure. The volume consumed, successive attempts, and the time required to complete the work also factor into the overall cost.

Astra Aims for Less Edited Deliverables

OpenAI designs Astra for complex professional missions that involve multiple steps, with the goal of limiting corrections before use. Presentations, financial documents, and software projects are among the targeted outputs. More broadly, the announced uses for the new models also cover document analysis, code writing, and computer manipulation.

On GDPval v2, the results reported by OpenAI give Astra a gain of 6.2 points over the previous model. The consumption of generated tokens would be 22% lower in this evaluation. These figures do not constitute a business test: the available data, tooling, and difficulty of tasks remain decisive.

The deployment concerns ChatGPT Work, Codex, as well as OpenAI's API. For developers, the announced accesses carry the identifiers gpt-6.2-astra and gpt-6.2-sol. OpenAI intends to integrate them into tasks already performed with its tools, without imposing reorganization.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.