CONFIRMED
OpenAI launched GPT-6.1 Sol on Tuesday, September 29, and the central change is economic: the company says the new model comes close to GPT-6 Astra on agentic coding, computer use and professional work, while costing one-fifth of the flagship model’s standard price. This is not a completely new Sol, but an upgrade to GPT-6 Sol with a much clearer ambition to occupy the upper middle of production workloads.
What changes in practice
In the API, GPT-6.1 Sol costs $2 per million input tokens, $10 per million output tokens and just $0.10 per million cached input tokens. OpenAI says cached input is 95% cheaper than standard input and 50% cheaper than GPT-6 Sol’s cached input. For agents that repeat context, tools, rules and documents across many turns, that detail can matter more to the bill than the full token price.
The numbers that matter
In OpenAI’s own evaluations, the model matched GPT-6 Astra on DeepSWE v1.1, which measures software engineering in real repositories, and scored 6.4 percentage points above the previous GPT-6 Sol. On AutomationBench, it beat Opus 5.5 by 2.2 percentage points at medium effort and cost about one-third as much per task. On OSWorld 2.0, it came within 2.1 points of Astra on computer use at roughly one-seventh of the cost. On Terminal-Bench Science, OpenAI calculated $5.47 per task versus $23.80 for Astra.
The caveat that keeps the hype in check
These results are not a neutral leaderboard. OpenAI ran its own tests and used public competitor results, while noting that API and ChatGPT outputs can differ because of system prompts, tools and reasoning effort. Independent coverage from Investing.com repeated the key figures and prices, but also relied on OpenAI’s announcement. That corroborates the news, not a controlled test in every team’s environment.
Impact on small and midsize businesses
For a small business, the opportunity is not to replace everything with Sol. It is to lower the cost of repetitive tasks that still need real capability: document analysis, context-rich support, spreadsheet automation, code review and agents with many calls. Cheap caching favors persistent workflows. The right calculation still is cost per verified delivery, including failed calls, human supervision, tools, storage and rework. A pretty benchmark does not pay the bill.
MaxAssistant’s read
GPT-6.1 Sol is the most interesting GPT-6 family launch so far because it attacks the real adoption bottleneck, the cost of keeping agents working for a long time. My read is that it deserves immediate testing, but not a default switch by decree. The contest is no longer who gets close to Astra on a chart, but who completes real tasks with less intervention and lower total cost.
Sources
OpenAI, official announcement
OpenAI, system card
Investing.com, independent coverage
