OpenAI shipped GPT-6.1 Sol on September 30, and the pitch is simple: results close to the company’s flagship GPT-6 Astra, at the price of its mid-tier model. If you pay for AI coding tools or run agents through the API, this is the release most likely to change your bill this month.
Here is what actually changed, what it costs, and who should switch.
What GPT-6.1 Sol Is
OpenAI’s GPT-6 family has three tiers. Astra is the flagship, which we covered in our breakdown of the GPT-6 Astra safety rating. Sol sits in the middle, and Luna is the low-cost option. Sol and Luna arrived after Astra with API prices cut roughly in half compared with the previous GPT-5.6 generation, according to The Next Web.
GPT-6.1 Sol is the first update to that middle tier. Per OpenAI’s release notes as tracked by Releasebot, it improves agentic coding, computer use, and general professional work, with results OpenAI describes as near-Astra at a much lower cost.
The Numbers Behind the GPT-6.1 Sol Upgrade
The headline benchmark is DeepSWE v1.1, a software engineering test. At high reasoning effort, GPT-6.1 Sol scores 75.2%, up from 68.8% for the original GPT-6 Sol. That is a 6.4-point jump inside one model tier.
Pricing did not move on the main rates:
- Input: $2 per million tokens, unchanged from GPT-6 Sol
- Output: $10 per million tokens, unchanged
- Cached input: $0.10 per million tokens, cut from $0.20
- Cost per task on DeepSWE: roughly 76% lower than GPT-6 Sol, according to OpenAI’s figures
That last line deserves a caveat. A lower cost per task alongside the same token price means the model finishes the work using fewer tokens, which is a claim from the vendor rather than something independently verified. Treat it as a direction, and test it on your own workload before budgeting around it.
For scale, GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens. Sol’s rates are one fifth of that on both sides.
Who Can Use It Right Now
GPT-6.1 Sol is rolling out in ChatGPT Work and Codex, starting with Pro users. It is also available through the OpenAI API, and it is included on the Pro 500, Business, Enterprise, and Edu plans.
Pro 500 subscribers get one extra: an Ultrafast tier in Codex that OpenAI says generates tokens up to eight times faster. If you are on a lower plan, expect a staggered rollout over the coming days rather than instant access.
Should You Switch From Astra?
The answer depends on what you run it for. OpenAI itself positions Astra for work where maximum quality matters more than cost, and Sol for complex work you run constantly.
- Stay on Astra for one-off, high-stakes tasks where the last few points of quality justify a 5x price.
- Move to GPT-6.1 Sol for coding agents, bulk document processing, and anything that loops through hundreds of calls a day.
- Consider Luna for simple classification or drafting where even Sol is overkill.
The agent case is where the saving compounds. Teams running coding agents burn through tokens quickly, which is part of why usage limits have been tightening across Google, Anthropic, and OpenAI. A model that scores higher on coding and costs less per task eases that pressure directly. It also fits the broader trend we tracked when AI coding agent adoption doubled in seven months: as more developers lean on agents, price per completed task starts to matter more than price per token.
How It Compares to the Competition
The nearest rival in this price bracket is Anthropic’s Claude Sonnet 5.5, which launched days earlier. We compared it head to head with Astra in our Claude vs GPT-6 Astra analysis, and the pattern there was the same: mid-tier models are closing in on flagships for most real work. GPT-6.1 Sol puts OpenAI back in that race on price. No independent head-to-head against Sonnet 5.5 exists yet for the 6.1 version, so wait for third-party benchmarks before declaring a winner.
What the Release Does Not Tell You
OpenAI’s announcement covers coding and computer use, so there is little public data on how GPT-6.1 Sol handles long writing, research, or multilingual work compared with the original Sol. A strong DeepSWE score also does not guarantee a strong result on your own codebase, where conventions, legacy code, and test setups differ from any benchmark. Cached input pricing only helps if your prompts repeat a large shared prefix, which suits agents with long system instructions but does little for one-off questions.
What to Do This Week
Pick one workflow you already run on Astra or the original Sol, such as a code review prompt or a document extraction job. Run the same 20 or 30 inputs through GPT-6.1 Sol, then compare the output quality and the actual token usage. If quality holds and tokens drop, you have found a real saving. If it does not, you lost an afternoon and learned where the tier boundary sits for your work.
Watch for independent benchmark results over the next two weeks. Vendor numbers on a launch day are a starting point, and the first outside tests usually show where the gaps are.
Sharing Is Caring
One thought on “GPT-6.1 Sol: Near-Astra Coding Power at One Fifth the Token Price”