OpenAI released GPT-6.1 Sol at its DevDay conference in San Francisco on 29 September, one week after GPT-6 Sol. The list price is unchanged at $2 per million input tokens and $10 per million output tokens, but cached input drops to $0.10 per million, which Unite.AI reports is half GPT-6 Sol's cached rate. Alongside it came a Decisions API for classification and routing, and computer use in the Agents API. If your product makes LLM calls, all three are worth an hour this week.
OpenAI's own announcement pages blocked automated access when we checked, so the figures below are as reported by Unite.AI, Apidog and the DEV Community recap, which cite OpenAI's posts.
GPT-6.1 Sol: same price, cheaper repeats
According to Apidog, the model ID is gpt-6.1-sol, with a 922,000-token maximum input, 128,000-token maximum output and an April 2026 knowledge cutoff. It is live in the API and in ChatGPT Plus, Pro, Business, Enterprise and Edu.
OpenAI's headline claim, per Unite.AI, is that it matches GPT-6 Astra on the DeepSWE v1.1 coding benchmark at roughly one-fifth of the cost. Other vendor figures: seven percentage points ahead of GPT-6 Sol on OSWorld 2.0, 2.2 points ahead of Claude Opus 5.5 on AutomationBench at medium effort, and a factual error rate down from 11.4% to 7.7%. The independent view is more modest. The DEV Community recap cites Artificial Analysis putting Sol at 52 on its Intelligence Index against 58 for Opus 5.5.
One breaking change: Apidog reports that the "none" effort level has gone, and the minimum is now "low". If you set effort to none to keep latency down on simple calls, those requests need updating before you switch model strings.
The price lands exactly on Claude Sonnet 5.5, which Anthropic released the day before at $2/$10. Sonnet 5.5's cache reads cost $0.20 per million, so for workloads dominated by a long, repeated system prompt, Sol's cached rate is now half Anthropic's.
The Decisions API
This one is aimed squarely at the "which bucket does this go in?" calls that litter most SaaS backends. Apidog describes it as handling "a specific set of user-defined questions with finite pre-defined answers", for content classification and request routing. It runs on GPT-6 Luna, OpenAI's small model, and entered limited preview on 29 September with broad availability promised "in the coming days".
Pricing is unclear. The DEV Community recap lists $0.10 input and $0.50 output per million tokens, which is Luna's standard rate. Apidog says dedicated pricing and docs have not been published. Assume Luna rates until OpenAI says otherwise.
This is the same idea as TypeSafe's Jev and the open-source Jeff models we covered yesterday: return a probability over fixed options instead of generating text you have to parse. OpenAI entering the space makes it a mainstream pattern.
Computer use in the Agents API
Computer use is now in public beta in the Agents API, letting agents drive a hosted browser. Per Apidog, it asks for approval before visiting a new website origin and before handling credentials. There is no surcharge beyond tokens, but the sandboxed environment costs $0.03 to $0.48 per 20-minute session.
Two constraints matter for anyone selling to European or regulated customers: it is US data residency only, and Zero Data Retention is not available. If your contracts promise either, you cannot use it for those customers yet.
What else was announced
Dots are always-on agents running on GPT-6 Astra, each with its own cloud computer and access to more than 4,000 apps, reachable from ChatGPT, Slack and Teams. They are for Pro and Business Premium plans in eligible markets. Pro 500 is a new $500-a-month ChatGPT tier with 25 times the Plus allowance and the Ultrafast mode, which OpenAI says reaches 300 tokens per second. SQ Magazine reports that Pro 200's Codex and Work allowance falls from 20x to 10x Plus, with existing subscribers keeping the old allowance until 29 October.
Bloomberg reported the same day that OpenAI has held back a version of its Astra model over safety concerns. OpenAI has not said when, or whether, it will ship.
What a small SaaS team should do
1. Re-run your model comparison. Sol and Sonnet 5.5 are now the same list price, so the deciding factors are quality on your prompts and tokens per task. Take 50 to 100 real prompts from your logs and run them through GPT-6.1 Sol, Sonnet 5.5 and whatever you use today. Promptfoo or Braintrust will do it in an afternoon.
2. Restructure prompts for caching. At $0.10 per million cached tokens, a long static system prompt is close to free on repeat calls. Put instructions, examples and reference material first and user-specific content last so the prefix stays identical.
3. Find your decision calls. Search your codebase for LLM calls whose output is really a label or a yes/no. Log inputs and accepted answers now, so you have a test set ready when the Decisions API opens up.
4. Check effort settings before migrating. Grep for effort "none" and change it to "low", then measure the latency difference.
5. Treat computer use as a prototype tool. It is useful for internal automation against sites without APIs. Keep it away from customer data you have promised to keep in the EU or out of logs.
The bottom line
GPT-6.1 Sol does not cut the headline price, but halving cached input makes it the cheapest frontier-class option for prompt-heavy workloads, and it now ties Sonnet 5.5 on list price. The Decisions API is the more interesting launch for SaaS builders: once pricing is published, many small classification calls could move to it. Benchmark on your own traffic, fix the effort setting, and wait for published Decisions pricing before rewriting anything.