OpenAI released GPT-6.1 Sol on 29 September 2026 as a mid-priced reasoning model that reads text and images and can write up to 128,000 tokens in one response. It fits teams running coding, document and computer-use agents at volume who want results near Astra, though the model lacks audio, video and fine-tuning support.
GPT-6.1 Sol is OpenAI's reasoning model, released in September 2026, that scores 52 in Artificial Analysis's composite Intelligence Index when run at maximum effort. Its inputs are text and images, its output is text, and it exposes five reasoning effort levels. OpenAI positions it as the mid-cost GPT-6 tier, built to sit closer to Astra than GPT-6 Sol does.
Where it sits
- $4.00/M$ per 1M tokensBlended price (3:1)Lower is better#48 / 73peer median $1.71/Mvendor price, checked by HokAI
- 67 tok/stokens/sOutput speedHigher is better#30 / 45peer median 85 tok/scited: Artificial Analysis
Pricier than 68% of the 73 GA models with a published price, rank 30 of 45 on output speed as cited from Artificial Analysis, and one of 28 that document a zero-data-retention option.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-6
Context window: 1,050,000 tokens · Max output: 128,000
Input modalities: text, image · Output: text
About GPT-6.1 Sol
GPT-6.1 Sol is a reasoning model from OpenAI, released on 29 September 2026 at DevDay as an upgrade to GPT-6 Sol, which shipped a week earlier next to GPT-6 Luna. It is the middle tier of the GPT-6 family: GPT-6 Astra is the flagship, Sol is the mid-cost option and Luna is the cheapest. OpenAI has not disclosed the architecture or parameter count. The API id is gpt-6.1-sol, and the model accepts text and images and returns text. You can browse the rest of the lineup on the OpenAI models page.
OpenAI's launch post frames the model as near-Astra performance on agentic coding, computer use and professional work. In its own evaluations, run in its research environment, GPT-6.1 Sol matches Astra on DeepSWE v1.1 and beats GPT-6 Sol's best score there by 6.4 percentage points at a lower reasoning effort. On AutomationBench it scores 2.2 percentage points above Claude Opus 5.5 at medium effort, and on the OSWorld 2.0 offline set it lands 2.1 points behind Astra at maximum effort. On Terminal-Bench Science 0.1 it costs $5.47 per task at maximum effort, against $23.21 for Opus 5.5 and $23.80 for Astra, but OpenAI says Astra keeps the top score there at 68.1% and should take the hardest research tasks. All of these are vendor-run results, not independent tests.
The independent view comes from Artificial Analysis, which measured 66.8 output tokens per second on OpenAI's API and an average cost of $0.72 per Intelligence Index task at maximum effort, both checked on 30 September 2026. That speed is below the median it reports for comparable reasoning models, and its composite index places the model one point under Astra and below Opus 5.5. The gap between OpenAI's selective charts and a single composite score is the main thing to weigh before switching.
Its context window holds 1,050,000 tokens, of which up to 922,000 can be input. Through the Responses API it can call functions, search the web and files, run code, use a hosted shell, apply patches, drive a computer, reach MCP servers and delegate to subagents in beta. Chat Completions works only without tools, and audio, video, Realtime and fine-tuning are not supported. In ChatGPT it is available in ChatGPT Work and Codex on paid plans, and not yet in standard Chat. Read ChatGPT Work for what that plan surface adds.
On safety, OpenAI treats the model as Critical for cybersecurity and High for biological and chemical capability under its Preparedness Framework, and applies the same safeguards stack as Astra. In its broken-search-tool test, where an agent must tell the user that a search tool failed, GPT-6.1 Sol failed to say so in 2.1% of cases, compared with 4.9% for GPT-6 Sol and 1.5% for Astra. OpenAI notes that its safety tasks are chosen to elicit failures and do not describe typical use.
The practical read: it suits teams that run coding, document and computer-use agents many times a day and can accept a lower composite score than the leaders in exchange for a lower bill. It is the wrong choice for audio or video work, for interactive paths at maximum effort, and for anyone who needs weights or fine-tuning: it is one of the proprietary models with closed weights, and open-weights options such as Kimi K3 exist. Compare it against Claude Fable 5.1 and GPT-5.6 Sol if you are choosing across vendors. HokAI's guide to choosing an LLM in 2026 and its Astra comparison guide cover the wider field, and Smart Match can shortlist models for a specific job.
Pricing
Standard billing is $2 per 1M tokens for input, $10 for output and $0.10 for cached input, with cache writes at $2.50. Batch and Flex processing are 50% lower, Fast mode is 2x Standard, regional processing adds 10% where offered, and a request over 272K input tokens pays 2x on input and cache rates and 1.5x on output for the whole request. OpenAI has announced an Ultrafast tier for this model but had not released it at launch, and it has not published a separate Ultrafast rate. GPT-6 Astra and Claude Opus 5.5 list higher standard rates.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.060 | $0.010 | $0.070 |
| Support reply | $0.0040 | $0.0030 | $0.0070 |
| One coding agent run | $0.400 | $0.200 | $0.600 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Five reasoning effort levels: The reasoning.effort setting takes low, medium (the default), high, xhigh and max, while none and minimal are not supported.
- Large input and output limits: Up to 922,000 input tokens fit inside the context window, and one response can run to 128,000 output tokens.
- Text and image input: The model reads images alongside text and returns text, and it does not accept audio or video.
- Built-in tools in the Responses API: Web search, file search, code interpreter, hosted shell, apply patch, computer use, MCP, skills and tool search all work through Responses.
- Cheap cached input: Cached input is billed at 5% of the uncached rate, and cache writes cost 1.25x the uncached rate.
- Multi-agent delegation (beta): OpenAI's changelog says a Responses request can let the model hand work to subagents, in beta at launch.
Pros
- Repeated agent runs stay affordable: OpenAI reports per-task costs far below Astra's in its own tests.
- Speed and cost also come from an independent measurement, not only vendor claims, so a team has two sets of numbers to check against its own tests.
- US and EU data residency cover Standard, Flex and Batch processing, and OpenAI holds SOC 2 Type 2 and ISO 27001 audits for its API.
- OpenAI reports the failure-to-disclose rate in its broken-search test at 2.1%, down from 4.9% for GPT-6 Sol.
- OpenAI's published deprecation policy promises at least six months' notice before it retires a generally available model, which gives teams time to plan a migration.
Cons
- Sits below Claude Opus 5.5 and GPT-6 Astra in Artificial Analysis's composite scoring of maximum-effort runs.
- No audio, video, Realtime or fine-tuning support, and the weights are closed.
- At launch it is in ChatGPT Work and Codex but not standard Chat, and Ultrafast was not yet available for it.
Benchmarks
- AA Intelligence Index: 52 cited: Artificial Analysis · 30 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- AA blended price: $1.47/M cited: Artificial Analysis · 30 Sep 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 67 tok/s cited: Artificial Analysis · 30 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does GPT-6.1 Sol cost in 2026?
On the standard API, input costs $2 per million tokens and output costs $10, while cached input is $0.10 and cache writes are $2.50. Batch and Flex are 50% lower, Fast mode is 2x, and any prompt over 272K input tokens reprices the entire request. GPT-6 Astra lists $10 and $50 on the same basis, and Claude Opus 5.5 lists $4 and $20. OpenAI says an Ultrafast option for this model is coming within days but has not published its rate.
Is GPT-6.1 Sol better than Claude Opus 5.5?
Artificial Analysis's composite (index v4.3.2, maximum effort, checked 30 September 2026) has Claude Opus 5.5 at 58 and GPT-6.1 Sol at 52, so Opus leads that comparison. OpenAI's own evaluations point the other way on selected agent tasks: it reports GPT-6.1 Sol 2.2 points ahead on AutomationBench at medium effort for about a third of the cost, and ahead on GDP.pdf at under half the cost per task. Those are vendor-run figures. Pick Opus 5.5 when the single highest score matters most, and pick GPT-6.1 Sol when you run many agent tasks and the bill decides.
Can you download or self-host GPT-6.1 Sol?
No. It is a proprietary model: OpenAI publishes no weights and no parameter count, so you reach it through OpenAI's API (Responses, Chat Completions and Batch endpoints) or inside ChatGPT Work and Codex. Fine-tuning is not supported. If you need weights you can host yourself, open-weights models such as Kimi K3 are the alternative.
What happens to the data you send to GPT-6.1 Sol?
OpenAI says data sent to its API is not used for training unless you opt in, and abuse-monitoring logs are kept for up to 30 days by default. Approved customers can move to Zero Data Retention or Modified Abuse Monitoring. Data residency in the United States or Europe (EEA and Switzerland) is offered through OpenAI's sales team, adds a 10% price uplift, and is not available with Fast mode in the EU.
Should you use GPT-6.1 Sol or GPT-6 Astra?
Use Astra when a wrong answer costs more than the tokens: OpenAI says it should take the hardest scientific research tasks, where Astra's 68.1% Terminal-Bench Science score is the highest OpenAI tested. Use GPT-6.1 Sol for coding, document and computer-use agents run at volume, where Artificial Analysis measured $0.72 per Intelligence Index task. Skip it for audio, video or Realtime work, which it does not support.
Top Alternatives
- GPT-6 Astra: Pick GPT-6 Astra when the hardest research tasks justify the spend; pick GPT-6.1 Sol for volume agent work at one-fifth of Astra's standard token price.
- GPT-6 Sol: Pick GPT-6.1 Sol over GPT-6 Sol unless you need the none reasoning effort; standard rates match and cached input costs half as much.
- Claude Opus 5.5: Pick Claude Opus 5.5 for the higher Artificial Analysis composite score; pick GPT-6.1 Sol when per-token cost decides.
- GPT-6 Luna: Pick GPT-6 Luna for scoped edits and simple extraction; pick GPT-6.1 Sol for complex work you expect to revise.