OpenAI strengthens its models with GPT-6.2 Astra and Sol

OpenAI unveils GPT-6.2 Astra and GPT-6.2 Sol, two models that use fewer tokens and outperform Claude Fable 5.2. The company also adjusts its API pricing and opens new access in ChatGPT Work and Codex.

On Tuesday, November 17, OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol, two AI models designed for different uses but launched simultaneously. Astra targets complex professional tasks, while Sol addresses everyday needs. The company mainly highlights improved efficiency, with fewer tokens consumed per task and revised API pricing, at a time when Anthropic is preparing for its IPO. The timing is no coincidence.

Astra climbs the ranks, and OpenAI targets heavy-duty tasks

GPT-6.2 Astra replaces GPT-6 Astra for the most demanding work, such as building a financial model or interpreting experimental data. OpenAI claims the generated texts "seem less written by an AI," a point that has become almost as strategic as raw performance. On Agents’ Last Exam, a multi-step professional task test, the model would outperform Claude Fable 5.2 by 5.5 points for "only 5% of its cost."

The comparison with Anthropic is direct. Astra would also gain 5.2 points over GPT-6 Astra on the same benchmark, with a cost per task less than a third of its predecessor’s. In science, OpenAI announces a record on Terminal-Bench Science 0.1, with 9 points more than GPT-6 Astra. On Harvey’s legal test Legal Agent Benchmark, Astra would complete more than three times as many tasks as GPT-6 Astra, with 34.5% fewer output tokens.

Sol seeks the right balance between price and performance

GPT-6.2 Sol succeeds GPT-6.1 Sol only seven weeks after its release. OpenAI presents it as a cheaper model but very close to Astra on several uses. On FrontierCode 2.0 by Cognition, it finishes less than 0.5 points behind Astra at about one-fifth the cost. On LifeVal, a daily task test published by OpenAI, Sol even surpasses GPT-6 Astra for about one-seventh of the cost.

The logic is clear: Astra for heavy missions, Sol for more common uses. Both models have been tuned to maximum effort, which reminds us that a benchmark never tells the whole story, but it still loves to be cited publicly. For OpenAI, the interest is twofold: to show solid results and to push users toward a more economical model where the performance difference remains small.

Aligned API prices, but different real usage costs

OpenAI also lowers Astra’s API price to $8 per million input tokens and $40 per million output tokens, down from $10 and $50 previously. The pricing now aligns GPT-6.2 Astra with Claude Fable 5.2. Sol retains GPT-6.1 Sol’s prices, at $2 per million input tokens and $10 per million output tokens. However, the cost difference per task highlighted by OpenAI mainly depends on the number of tokens consumed and the settings used in tests, not just the listed price.

Who can access it now?

Both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers, as well as in the API under the names gpt-6.2-astra and gpt-6.2-sol. Pro 200 subscribers get first-time access to Ultrafast mode, initially with Sol. Pro 500 subscribers and developers can use it with both models.

In this context, OpenAI is not just announcing a new version. The company wants to show it can still improve accuracy, speed, and efficiency while maintaining a commercial lead over its rival Anthropic, which is preparing for its IPO. In this market, power still matters, but so does the price tag.

Key points

  • November 17: OpenAI launches GPT-6.2 Astra and Sol.
  • Astra would outperform Claude Fable 5.2 by 5.5 points.
  • Sol costs about five times less on FrontierCode 2.0.
  • Anthropic sets its IPO price before November 19.
  • Pro 200 subscribers gain access to Ultrafast mode.

By the numbers

  • 5% — Astra’s cost on Agents’ Last Exam, according to OpenAI.
  • $8 — GPT-6.2 Astra’s API input price per million tokens.
  • $40 — Astra’s API output price per million tokens.
  • 34.5% — fewer output tokens on Harvey’s legal test.
  • 7 weeks — time between GPT-6.1 Sol and GPT-6.2 Sol.

To read