Paintings Subscribe Tiếng Việt

19 November 2026

AI News/OpenAI releases GPT-6.2 Astra and Sol: new models, old lesson

Two new models and cheaper API prices, but early testers are lukewarm. The lesson for builders: do not choose a model by its leaderboard, choose it by your real work.

On 17 November OpenAI released two new models: GPT-6.2 Astra and GPT-6.2 Sol.

Two models, two jobs

  • Astra is for demanding professional work: documents, presentations, spreadsheets, code and scientific research.
  • Sol is for everyday tasks and routine multi-step work.

Both are in ChatGPT, Codex and the API. Astra’s API price is $8 per million input tokens and $40 per million output tokens, about 20% cheaper than the previous Astra. Sol stays at $2 and $10.

One notable detail: according to OpenAI, Astra is the first model it has released externally that it rates “High” for AI self-improvement under its own risk framework. OpenAI also tightened safeguards around cybersecurity requests.

What early testers say

Not everyone is impressed. According to The Next Web, Epoch AI found Astra trailed Claude Fable 5.2 on realistic work tasks, and some well-known developers said they would still pick other models for daily coding.

The lesson for builders

I have seen this many times: every new model arrives with a beautiful benchmark table. But a benchmark is not your job.

1. Evaluate on your own work. Keep a set of 20–50 questions or tasks from your own product, with known answers. When a new model ships, run the set again. It takes an afternoon and is more reliable than any leaderboard.

2. Pick models per kind of task. There is no single “best” model. Use strong models for hard work and cheap ones for repetitive work. OpenAI’s Astra and Sol pairing admits exactly that.

3. Do not lock in to one provider. Design your product so changing models is a configuration change, not a rewrite. The market changes every month.

4. Price the whole system, not the token. A cheaper model that needs more retries because it gets things wrong ends up costing more.

New models will keep coming. What keeps you calm through each launch is a test set of your own.

Source: The Next Web