
AI-generated illustration
GPT-6.2 Astra Launches With a 20% API Price Cut
GPT-6.2 Astra is live. OpenAI released it on 17 November 2026 together with a cheaper sibling, GPT-6.2 Sol. Astra now costs 20% less through the API than the GPT-6 model it replaces. Sol keeps the price of the GPT-6.1 model it replaces.
Both models are rolling out to paid ChatGPT plans and the API. Here is what changed, what it costs and what to check before you switch.
What Is New in GPT-6.2 Astra and Sol?
OpenAI built Astra for hard professional work in science, law and finance. It calls Astra the best model for scientists. Sol is the fast, cheaper model for everyday tasks and multistep agent work.
Astra also gets better at matching a slide template’s layout, colors and fonts, even from a screenshot. It picks tools more sensibly too, such as choosing a plugin over computer use. OpenAI says it scores 6.2 points higher on its GDPval v2 test while using 22% fewer output tokens.
Astra and Sol Pricing and Availability
Astra costs $8 per million input tokens and $40 per million output tokens. Sol costs $2 and $10. Cached input is $0.40 for Astra and $0.10 for Sol.
|
Model |
Input |
Cached input |
Output |
Change |
|
Astra |
$8.00 |
$0.40 |
$40.00 |
20% cheaper than GPT-6 Astra |
|
Sol |
$2.00 |
$0.10 |
$10.00 |
Same as GPT-6.1 Sol |
Both models arrive in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Enterprise and Edu workspaces need an admin to switch them on. Pro 500 subscribers get Ultrafast mode with both. Pro 200 subscribers get it with Sol first.
Developers can call gpt-6.2-astra and gpt-6.2-sol in the API. An OpenAI staff reply says the models do not come to regular Chat. The announcement does not mention the free plan.
How Do the Benchmarks Look?
Astra improves on its predecessor, but it does not win everywhere. These figures come from OpenAI unless noted.
- Agents’ Last Exam: Astra scores 78.3%, up from 73.1% for GPT-6 Astra. Google’s Gemini 4 Carbon scores 77.4%.
- GDPval v2: Astra reaches 50.4%, against 44.2% before.
- Independent index: Artificial Analysis ranks Claude Fable 5.2 first at 64.6. Astra comes second at 62.1, and Sol scores 57.3.
- Coding: Gains stay small. Astra scores 44.2% on FrontierCode 2.0, behind Fable 5.2 at 49.1%.
Treat the vendor numbers with some care. Independent testers back the cost savings more than the raw performance jump.
Tighter Cyber Safeguards
Both models hit the Critical cybersecurity level in OpenAI’s Preparedness Framework. Astra is also the first OpenAI model released to the public that reaches the High level for AI self-improvement.
OpenAI added stricter safeguards. That means more refusals, and some of them will hit legitimate security research. Approved defenders can apply for access through OpenAI Daybreak.
Washington is tightening the screws on AI access from the hardware side too. The US recently tightened AI chip export rules to cover remote access to compute.
What You Should Do
- Developers: Swap in the new model IDs on a test branch first. Coding gains look small, so measure your own workload.
- Everyday users and teams: Start with Sol. It costs a quarter of Astra and handles most routine work.
- Admins: Turn on access for Enterprise and Edu workspaces.
- Security researchers: Expect refusals and apply for Daybreak if you need fewer.
Some users say the models have not appeared in their accounts yet. Access depends on plan, workspace permissions and rollout progress.
Sources: OpenAI announcement, The Decoder