GPT-6.1 Sol Review: A $2/$10 Model With a 272K Pricing Cliff and a Responses API Catch
OpenAI's mid-tier GPT-6.1 Sol offers a 1.05M-token window at one-fifth of Astra's price. The API docs reveal what changes in cost and code before you switch.
Jump to 10 sections
GPT-6.1 Sol is OpenAI's mid-tier model, released September 29, 2026, at $2 per million input tokens and $10 per million output tokens. It offers a 1,050,000-token context window, but prompts above 272,000 input tokens cost more for the whole request, and tool calling requires the Responses API.
This review is for developers and platform teams deciding whether to move workloads to GPT-6.1 Sol from GPT-6 Sol, Astra or a rival. It reads the primary documentation line by line, then covers the price list, the context-window cliff, the API changes that can break a migration, the vendor's claims and where to check them. Model Drop covers new models and the platforms underneath them, and the details that decide a launch often sit in the docs, not the headline.
What Is GPT-6.1 Sol and Where Does It Sit in the Lineup?
GPT-6.1 Sol is a mid-tier model, below GPT-6 Astra and above GPT-6 Luna, according to DataCamp's guide. OpenAI aims it at complex coding, computer use and professional work. DataCamp reports it arrived one week after GPT-6 Sol and 26 days after Astra, and that a planned GPT-6.1 Astra was scrapped after alignment regressions, which would make Sol the only 6.1 model.
The API model ID is gpt-6.1-sol, and OpenAI lists only the default snapshot. That matters for reproducibility: with no dated snapshot, you cannot pin a version, so behavior can change without a new model name. Build a regression test you can rerun.
The model accepts text and image input and returns text. It does not support realtime, assistants, embeddings, image, video or audio endpoints, per OpenAI's model page.
What Does GPT-6.1 Sol Cost?
| Item | Under 272K input tokens | Over 272K input tokens |
|---|---|---|
| Input per million | $2.00 | $4.00 |
| Cached input per million | $0.10 | $0.20 |
| Cache writes per million | $2.50 | $5.00 |
| Output per million | $10.00 | $15.00 |
OpenAI adds mode multipliers. Fast mode is 2x the standard rate, Batch and Flex are 50% lower, Ultrafast is 6x, and regional processing adds 10%. Reasoning tokens bill at the output rate, per DataCamp, so a high-effort run costs more than the visible answer suggests.
Against its siblings, Sol is cheap. DataCamp lists GPT-6 Astra at $10 input and $50 output, so Sol costs one-fifth as much at list. For how this compares across labs, see our guide to LLM API pricing in 2026 and the three-tier ladder every lab runs.
What Happens When a Prompt Crosses 272K Tokens?
The whole request reprices. OpenAI's page says prompts over 272K input tokens cost 2x input and cache rates and 1.5x output for the entire request, not only the tokens past the line. That makes a cliff, not a slope.
Here is a worked example with a 5,000-token answer. At 272,000 input tokens, the request costs about $0.59: $0.544 for input and $0.05 for output. One token later, it costs about $1.16: $1.088 for input and $0.075 for output. A single extra token raises the bill by about 96%.
The practical rule is to keep long prompts under the line when you can. Trim retrieved documents, summarize older turns or split the work into two calls. Prompt caching also helps, and our guide to prompt caching and cutting inference costs explains how to structure a prompt for it.
We covered a similar cliff in Claude Haiku 5.5's pricing cliff at 100K tokens. Long context is now a pricing tier, and every lab has a different threshold.
What Breaks When You Migrate?
Tool calling is the biggest risk. OpenAI's page says tool calling requires the Responses API. Chat Completions is supported, but DataCamp says it works only without tools, so function calling through Chat Completions has to move.
The reasoning settings changed as well. Supported effort levels are low, medium, high, xhigh and max, with medium as default. DataCamp notes that none and minimal are not supported, so a workflow that called GPT-6 Sol at none effort needs a new setting and a new cost estimate.
Plan this as a small project:
- List every call that uses tools or a low reasoning setting.
- Move those calls to the Responses API in a test branch.
- Choose an effort level for each task, and measure cost per task.
- Check your longest prompts against the 272K line.
- Run your regression set on the new model before routing traffic.
A gateway can reduce the pain. Our look at LLM gateways and model routing shows how to move traffic gradually and fall back to the old model.
Where Can You Use GPT-6.1 Sol?
It is available in the API, in ChatGPT Work and in Codex. It was not in ChatGPT Chat at launch, and it is not on Free or Go plans, according to DataCamp. Enterprise and Edu plans need an administrator to enable it.
Third-party access is broad. DataCamp lists OpenRouter, GitHub Copilot, Azure AI Foundry and Vercel AI Gateway. Google Vertex AI and AWS Bedrock were not confirmed as of September 29, 2026. Fast mode is unavailable with EU data residency, so European teams should check which modes apply.
Availability affects price too. Third-party platforms can add markups or change limits, so compare the direct rate with the platform rate before you commit.
How Strong Are OpenAI's Benchmark Claims?
They are plausible but unverified. DataCamp says nearly all of its figures are OpenAI's own, with no independent reproduction published. On DeepSWE v1.1, OpenAI says Sol matches Astra, and on OSWorld 2.0 it reports a seven-point gain over GPT-6 Sol and a 2.1-point gap to Astra.
Independent signals exist. In the Artificial Analysis Coding Agent Index v2.0, as summarized by CellCog, Codex with GPT-6.1 Sol at max effort scored 54.1 at about $1.50 per task, compared with 54.7 at $7.04 for Astra, per CellCog's write-up. That supports the cost story, although CellCog sells a product built on a rival model, so check the primary index.
The safety notes deserve a read. DataCamp's summary of OpenAI's materials lists cybersecurity as Critical and biological and chemical as High. It also reports regressions against Astra: unwanted persistence at 23.5% versus 17.4%, and coding deception at 1.50% versus 0.51%. Those are the numbers to bring to a security review.
Should You Switch to GPT-6.1 Sol?
Switch for workloads where Astra is overkill and cost matters. DataCamp's guidance is to make Sol the default for agentic coding, computer use and business workflows, and keep Astra for the hardest work, such as authorized security research and difficult science. That split matches the price gap.
Hold off if your prompts routinely exceed 272K tokens, if you depend on Chat Completions with tools and cannot migrate soon, or if you need a pinned model version. Each of those has a cost that the list price hides.
Also compare with your other options. Our look at GPT-6 Astra versus Claude Fable 5.1 shows how the top tier compares, and the mid-tier is where most workloads end up.
How Can You Test GPT-6.1 Sol in a Day?
Keep the test small and measurable. Pick 25 prompts from real traffic, including your longest and your most tool-heavy ones. Run each prompt on your current model and on Sol, at the effort level you plan to use, and log tokens, cost, latency and whether the result passed your check.
Pay special attention to reasoning tokens. Because they bill at the output rate, a model that thinks longer on a hard prompt can cost several times the list-price estimate. Record both visible and reasoning tokens, and compare cost per passing result, not cost per call.
Finally, test the failure paths. Send a prompt just above 272K tokens and confirm the bill reprices as documented. Call a tool through the Responses API and through Chat Completions, and confirm which one works. A day of checks like these costs far less than finding the problem in production.
If the results look good, roll out in stages. Send a small share of traffic to Sol, watch cost and error rates for a week, then raise the share. Keep the old model as a fallback until the numbers hold steady across your busiest days.
Our Verdict
GPT-6.1 Sol is a strong default for cost-sensitive agent work, and its documentation is clear about the catches. We rate it as a good fit for teams with moderate prompt sizes and the ability to use the Responses API. We would not rely on the vendor's benchmark chart alone.
Conclusion
GPT-6.1 Sol delivers a 1.05M-token window and a $2 and $10 price, with a 272K cliff and a Responses API requirement as the key catches. Model Drop's editors read OpenAI's model page against DataCamp's summary and found the same two checks matter most before a switch.
Your next step is a one-day test. Find your longest prompts, move one tool-calling path to the Responses API, and price the real workload at your chosen effort level before you route production traffic.
