AT A GLANCE

OpenAI reports that Astra beat Claude Fable 5.2 on one multistep work benchmark. Use a lower-cost model for routine work if it meets your requirements, and confirm where your data is processed and stored.

Last week I covered Anthropic’s Claude Fable 5.2. Eight days after that launch, OpenAI introduced GPT‑6.2 Astra for more complex projects and Sol for faster, lower-cost work. Both are available in ChatGPT Work and Codex on eligible plans and through OpenAI’s developer interface. OpenAI’s release (opens a new tab)

On Agents’ Last Exam, a benchmark for multistep professional tasks, OpenAI reports that Astra scored 5.5 points ahead of Fable 5.2 at about 5% of the estimated cost. Results vary by benchmark and by task. OpenAI also shows a walkable 3D Paris built with Astra in Codex using street maps and building data.

Fit for purpose still comes first. Try a lower-cost option to extract details from invoices. Pay for more reasoning if it improves a real result enough to justify the extra cost. Include the checking and corrections people have to do.

When a law, customer contract or internal policy requires data to stay in South Africa, check where the service processes prompts and files. Confirm where results, logs and backups are stored, and who can access them. Choose a deployment that meets those requirements.

THE SOURCE / FURTHER READING

Want the full feature?

You’ve read the TOP STORIES interpretation. Continue to the publisher for the source feature or detailed guidance.

Read the original feature at OpenAI (opens a new tab)
TOP STORIES #194, edition dated 23 November 2026. Reporting and commentary reflect that edition; external source material may have changed.
← Back to the edition