Strategy5 minNewsroom

OpenAI Debuts GPT-6.2 Astra at 20% Lower Price than GPT-6

GPT-6.2 Astra costs $8 and $40 per million tokens versus $10 and $50 from its predecessor, ranking second in the Artificial Analysis index, behind Claude Fable 5.2.

Arte de lançamento do GPT-6 Astra: galáxia em espiral no formato do número 6 sobre o céu estrelado.
© OpenAI

OpenAI launched GPT-6.2 Astra on Tuesday (17), its most capable model for complex professional work, along with GPT-6.2 Sol, the intermediate version. Astra arrives at the API priced at $8 per million input tokens and $40 per million output tokens, 20% lower than the $10 and $50 charged by GPT-6 Astra. The cached input price drops from $1.00 to $0.40, a 60% cut on the item that has the most weight in applications with repeated context.

The Sol model maintained the price of GPT-6.1 Sol: $2 for input and $10 for output. According to Bloomberg, the launch of Astra comes almost two months after OpenAI discarded a previous model upgrade, which makes version 6.2 a testing ground for the company as well.

What the New Pricing Means

The discount represents not a reduction of the existing model but rather a new model launched below its predecessor. This distinction matters for ongoing contracts: applications built on GPT-6 Astra will only capture the savings after migrating and revalidating, and in regulated environments, such revalidation can require weeks of testing.

The cache is where the real change occurs. Corporate agents who re-read the same policy manual, the same codebase, or the same contract with each call primarily consume cached input. For these cases, the drop from $1.00 to $0.40 carries more weight on budget than the 20% headline.

Second Place in Independent Ranking

In the intelligence index from Artificial Analysis, one of the few independent assessments published on the day, GPT-6.2 Astra came in second, behind Claude Fable 5.2 from Anthropic. Artificial Analysis itself noted that the launch delivers "most of the Pareto frontier of cost per task," meaning that for a large portion of budget ranges, one of OpenAI's two new models offers the best results for the money.

Improvements over its predecessor exist but are incremental across most metrics, and standout numbers come from the release charts of OpenAI itself. Those expecting a clear return to leadership after the abandonment of the previous upgrade found a competitive model, not a dominant one.

The Same Day as Microsoft

OpenAI chose the day Microsoft, its main business partner, opened Ignite, presenting seven proprietary models from the MAI line. The MAI Code 2 was compared by Microsoft to GPT-6.1 Sol, with task costs of $0.309 versus $0.761 for the OpenAI model, according to graphs from the Redmond company. The competition between the two has ceased to be implicit and is now appearing in pricing slides.

A skeptical reading deserves to be recorded. List price per token is becoming a less useful metric for comparing vendors, as different models consume very different amounts of tokens to resolve the same task. A model that is 20% cheaper per token but requires more reasoning steps may end up being more expensive per problem solved. This is why cost per task, rather than price per million tokens, has become the number that vendors highlight and that buyers should demand in their own testing.

GPT-6.2 Astra is now available on the API and in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers, according to the launch documentation circulated by specialized media.

The Calculation in Two Markets

In the European Union, switching models carries regulatory costs. The transparency obligations of Article 50 of the AI Act take effect from August 2, 2026, and the transition period for machine-readable labeling of generated content ends on December 2, 2026, for systems already in operation. European businesses that migrate from GPT-6 Astra to 6.2 over the coming weeks will do so just as this deadline expires, and must document the model change within the same compliance cycle.

In India, the effect comes through global capacity centers and integrators operating agents for American and European clients at fixed prices. For these contracts, a 60% drop in cached input costs is direct margin, as long as the client doesn't require passing the savings through. Those purchasing managed AI services now have a public number to bring to the negotiation table.

The most revealing aspect of the launch is the cadence, more than the price. OpenAI took nearly two months to replace a discarded update, while Anthropic leads the independent ranking and Microsoft has begun to offer its own alternatives in the Copilot menu. For buyers, this reinforces the argument for architectures that switch models without re-writing the application.

Sources

  1. bloomberg.comhttps://www.bloomberg.com/news/articles/2026-11-17/openai-unveils-new-astra-ai-model-after-scrapping-prior-release
  2. x.comhttps://x.com/ArtificialAnlys/status/2122750634210529569
  3. the-decoder.comhttps://the-decoder.com/openai-launches-gpt-6-2-astra-and-sol-with-lower-api-prices-and-some-benchmark-gains/
  4. mlq.aihttps://mlq.ai/news/openai-launches-gpt-62-astra-and-sol-cutting-astras-api-price-by-20/
  5. datastudios.orghttps://www.datastudios.org/post/openai-gpt-6-2-astra-sol-launch-api-pricing-chatgpt-codex

The week's analysis, by email

One weekly edition with what matters to people who decide. No ads, no sponsorship.

One-click cancellation, at any time.

Strategy →