OpenAI’s GPT-6.1 Sol brings near-flagship performance at a fifth of the price

OpenAI used its DevDay event on September 29 to launch GPT-6.1 Sol, a cheaper model it says gets within touching distance of its flagship, GPT-6 Astra, on the work developers actually pay for: agentic coding, computer use and professional tasks.

The headline is price. In the API, Sol costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million — roughly one-fifth of Astra’s standard rates, according to The Next Web.

What actually changed with GPT-6.1 Sol?

OpenAI’s published comparisons (its own numbers, not independent tests) put Sol close to Astra at a fraction of the cost:

• DeepSWE v1.1: matches Astra’s score at roughly one-fifth of the cost.
• OSWorld 2.0: within 2.1 points of Astra at about one-seventh of the per-task cost.
• AutomationBench: 2.2 points above Anthropic’s Opus 5.5 at about a third of the cost.

Sol is available now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and to developers as gpt-6.1-sol. It is not yet in the main ChatGPT chat, and OpenAI says a faster “Ultrafast” tier is on the way.

The fine print

Sol is not a free lunch. OpenAI’s own published results show some safety evaluations with slightly higher failure rates than Astra — for example 2.1% versus 1.5% on detecting a broken search tool — and Astra keeps the lead on the hardest scientific tasks. A model that is “nearly as good” on average can still be meaningfully worse on your specific edge cases.

What to build next

1. Re-run your own evals. Public benchmarks set expectations; your traffic decides. Fifty real examples beat any leaderboard.
2. Route by difficulty. Send routine calls to Sol and escalate only low-confidence or high-stakes requests to Astra.
3. Lean on caching. At $0.10 per million cached input tokens, long, stable system prompts and shared context become very cheap.

The price of “good enough” fell again. The teams that win this round will be the ones that measure, route and cache — not the ones that simply pick the biggest model.

Cheaper tokens are only half of inference cost — idle GPUs are the other half. See why Nebius bought Inferize to fix AI cold starts.