GPT-6.1 Sol took the stage at DevDay on September 29th with a price list that overturns the developers’ accounts: $2 per million input tokens and $10 per million outputa fifth of what OpenAI is asking for GPT-6 Astra. The company declared performance almost in line with the flagship model on agent coding, computer use and professional activities. The debut came exactly one week after the release of GPT-6 Sol.
The numbers of DeepSWE v1.1which measures software engineering interventions on real codebases, agree with the statement: 75.2% with reasoning on high, against 74.8% of Astra and 68.8% of GPT-6 Sol pushed to maximum effort. The average bill per task drops to approx $1.50whereas Astra consumes 7.70.
Overtaking is not generalized. On OSWorld 2.0 the model stops at 71.4% against 73.5% for Astra, on GDP.pdf for the documentary analysis the gap is statistically irrelevant (32.0% against 32.2%). On Terminal-Bench Science 0.1 the advantage is all economic: $5.47 per task versus Astra’s $23.80.
The million token window is not free
The context window comes to 1,050,000 tokens input with output up to 128,000, but the rate is not linear. Passed i 272,000 input tokens,OpenAI bills double on input and cache and 1.5x output for the entire request. Those who work below that threshold instead find the input in cache a $0.10 per million, half of its predecessor, with write caching at $2.50.
The control over reasoning also changes: five levels of reasoning effort (low, default medium, high, xhigh and max), without minimal or none modes. The low-effort factual error rate drops from 11.4% to 7.7%while adherence to instructions on the chain of thought went from 23.2% to 44.8%. Tool calling, however, only works through the Responses API.
It will arrive in the next few days GPT-6.1 Sol Ultrafasta variant designed for those who have latency as a constraint: up to eight times the generation speed in Codex, with a peak of 300 tokens per secondand up to six times in the API. The price follows, because the rate rises to six times that of the standard bracket.
The security chapter is interesting, where OpenAI classifies the model as Critical in the cybersecurity field and High on biological and chemical capabilities, applying the same countermeasures stack as Astra. On recently disclosed exploits the arbitrary code execution rate stands at 21.5%well above the 5.5% of GPT-6 Sol and below the 31.5% of the flagship.
The distribution starts from Plus, Pro, Business, Enterprise and Edu users on ChatGPT Work and Codexwith the identifier gpt-6.1-sol for API and availability on OpenRouter, Vercel AI Gateway, and GitHub Copilot Pro+. It’s not there in the standard chat.
The void left by GPT-6.1 Astra remains, which OpenAI did not release: according to press rumors the project stopped due to higher levels of deception and the tendency to carry out tasks without asking the user for permission. An economic model that touches the top is convenient, but the real test will be to understand how much of this price cut depends on efficiency and how much on the absence of the successor who should have occupied that step.
