How is GPT-6 Sol different from GPT-6 Astra — and which one should you code with?

GPT-6 Sol is the lower-cost choice for long coding and agent runs; Astra remains the model to compare for the hardest reasoning. Benchmark both on your own workload before committing.

OpenAI released gpt-6-sol through the API on September 22 and began a gradual rollout in ChatGPT Work and Codex for eligible paid plans. Sol is positioned as the less expensive GPT-6 option for sustained coding and agent work: standard pricing starts at $2 per million input tokens and $10 per million output tokens for prompts up to 272,000 input tokens, with higher rates beyond that threshold. Its published limits are roughly a 1.05-million-token context window and 128,000 tokens of output, but capacity is not the same as accuracy.

Use Sol when a long repository, repeated tool calls, or many iterations make cost the constraint. Compare Astra for the hardest reasoning or when one bad decision would cost more than the model upgrade. Run the same representative task through both, measure usable results and total tool cost, and recheck the official price before production because processing tiers and long-context rules change the bill.