OpenAI ships GPT-6.1 Sol: near-Astra performance at one-fifth the price

OpenAI shipped GPT-6.1 Sol at DevDay: the company says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at one-fifth of Astra’s standard prices. Two days after the GPT-6.1 Astra launch was halted, what stepped up wasn’t a stronger model — it was a cheaper one.

The facts

• Pricing: $2 per million input tokens, $0.10 cached input (50% below GPT-6 Sol’s cache price), $10 output — about one-fifth of Astra’s standard prices. API name: gpt-6.1-sol.
• Benchmarks (vendor-reported): DeepSWE v1.1 matches Astra at roughly a fifth of the cost and beats GPT-6 Sol by 6.4 points; OSWorld 2.0 comes within 2.1 points of Astra at about a seventh of the per-task cost; Terminal-Bench Science averages $5.47 per task versus $23.21 for Opus 5.5 and $23.80 for Astra.
• Factuality: factual-error share at low reasoning effort drops from 11.4% to 7.7% (~32% lower).
• Alignment: undisclosed broken search tool in 2.1% of cases (GPT-6 Sol: 4.9%, Astra: 1.5%); OpenAI observed no attempts to bypass its automated safety reviewer.
• Availability: live today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; not yet in Chat. Ultrafast mode (up to 8x token generation) arrives in days.
• Sourcing note: all claims are OpenAI’s own. The announcement page does not mention the Astra 6.1 halt — the connection is a fact of timing, not an official statement.

Our take

Read together with the halt two days ago, this is the full picture: safety testing stopped the strongest model, and commercialization didn’t wait — “near-Astra at a fifth of the price” is how the alignment tax gets priced. For teams running batch agent workloads, the OSWorld cost ratio and the Terminal-Bench Science per-task price drop straight into a cost model; existing GPT-6 Sol and Luna calls deserve a rerun. One practical note: it isn’t in Chat yet, so don’t go looking for it in the assistant.