gpt-6.1-sol. It costs $2 input, $0.10 cached input and $10 output per million tokens, has a 1.05M-token context window, and OpenAI says it nearly matches GPT-6 Astra on agentic coding at about one-fifth the price.OpenAI released GPT-6.1 Sol on 29 September 2026 at DevDay, one week after GPT-6 Sol came out on 22 September. The price is the same as GPT-6 Sol, $2 in and $10 out per million tokens, and cached input is half what it was. The pitch is that it gets close to GPT-6 Astra, OpenAI's $10/$50 flagship, on coding, computer use and document work. It is now the default model in Codex and is available in the API, ChatGPT Work, GitHub Copilot and OpenRouter.
This guide covers the exact API specs from OpenAI's developer docs, how pricing works above 272K tokens, what OpenAI's benchmarks claim and what independent testers such as Artificial Analysis measured, where you can use it, a Responses API quickstart, and how it compares with GPT-6 Sol, GPT-6 Astra and Claude. At the end there is a short section on the reported cancellation of GPT-6.1 Astra, which happened on the same day. Every number is attributed to the source that published it.
What is GPT-6.1 Sol?
GPT-6.1 Sol is a reasoning model in OpenAI's GPT-6 family. The family has three tiers: Astra at the top, Sol in the middle and Luna for low-cost work. GPT-6.1 Sol is an updated Sol, not a new tier. It replaces GPT-6 Sol, which Artificial Analysis noted was replaced "after just 7 days". OpenAI's model catalog in Codex now calls gpt-6-sol the "previous generation workhorse model".
OpenAI's launch claim is that GPT-6.1 Sol "delivers nearly the same level of intelligence as GPT-6 Astra for agentic coding, computer use, and professional work" (quoted by TechCrunch). OpenAI also says it is better at factual accuracy on hard prompts and "more up front about its limitations".
The naming causes confusion. GPT-5.6 Sol, GPT-6 Sol and GPT-6.1 Sol are three different models with different prices. Always check the model ID (gpt-6.1-sol) before you trust a benchmark or pricing table.
GPT-6.1 Sol specs: context window, output limit and cutoff
These figures come from OpenAI's model reference page for gpt-6.1-sol, checked on 30 September 2026, with GPT-6 Sol's page for comparison.
| Spec | GPT-6.1 Sol | GPT-6 Sol |
|---|---|---|
| API model ID | gpt-6.1-sol | gpt-6-sol |
| Release date | 29 Sep 2026 | 22 Sep 2026 |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Max input tokens | 922,000 | 922,000 |
| Max output tokens | 128,000 | 128,000 |
| Knowledge cutoff | 30 Apr 2026 | 20 Apr 2026 |
| Input / output | Text and image in, text out | Text and image in, text out |
| Reasoning effort | low, medium (default), high, xhigh, max | none, low, medium (default), high, xhigh, max |
| Endpoints | Chat Completions, Responses, Batch | Chat Completions, Responses, Batch |
| Tier 1 rate limit | 500 RPM, 500,000 TPM | — |
Two things to note. First, GPT-6.1 Sol has no none reasoning setting. OpenAI's docs say "none and minimal reasoning efforts are not supported", so if your GPT-6 Sol code sets effort: "none" for fast, cheap calls, change it to low before you switch models. Second, the model does not accept or produce audio or video.
OpenAI's model page lists these built-in tools: web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search. OpenAI's model page also says to "use the Responses API for tool calling" and that "Chat Completions is supported without tool calling." If your agent uses tools, build it on the Responses API.
How much does GPT-6.1 Sol cost?
At standard rates GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens, $2.50 per million cache writes and $10 per million output tokens. That is one-fifth of GPT-6 Astra's $10/$50 list price and the same headline price as GPT-6 Sol. The one change is cached input, which drops from $0.20 to $0.10 per million (VentureBeat and OpenAI's model pages). That matters for agents that resend a large, stable prompt prefix on every step.
| Model (per 1M tokens) | Input | Cached input | Output | Long context (>272K input) |
|---|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $4 / $15 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 | $4 / $15 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $20 / $75 |
The 272K-token pricing threshold
OpenAI's pricing page defines "short context" as up to 272K input tokens. Above that, the whole request is billed at 2x the input and cache rates and 1.5x the output rate, so GPT-6.1 Sol becomes $4 in and $15 out. You get the full 1.05M-token window, but a 500K-token codebase dump costs twice as much per input token as a 250K one. If you can trim context below 272K, do it.
Batch, Flex and Fast modes
According to OpenAI's pricing page, Batch and Flex cost half the standard rate ($1/$5 for short context). Fast mode costs 2x standard ($4/$20). DataCamp also reports a 10% premium for regional data processing.
What about GPT-6.1 Sol Ultrafast?
OpenAI's Ultrafast tier promises up to 8x faster token generation in Codex and up to 6x in the API, reaching about 300 tokens per second, at 6x the standard price (VentureBeat). It is live for GPT-6 Astra today; a GPT-6.1 Sol version was announced alongside it. As of 30 September it is listed as coming "in the coming days", and OpenAI's pricing page does not yet have an Ultrafast row for GPT-6.1 Sol. GPT-6 Astra Ultrafast is already live at $60/$300. Treat any GPT-6.1 Sol Ultrafast price you see as an estimate until OpenAI publishes one.
For a full breakdown of flagship costs, see our GPT-6 Astra pricing guide.
GPT-6.1 Sol benchmarks: how close is it to Astra?
Most of the headline benchmarks come from OpenAI's announcement. OpenAI often publishes charts rather than tables, so the exact percentages below were read off those charts by Vellum and other outlets. Treat the vendor numbers as OpenAI's claims until more independent runs appear.
| Benchmark | GPT-6.1 Sol | GPT-6 Astra | GPT-6 Sol | Claude | Reported by |
|---|---|---|---|---|---|
| DeepSWE v1.1 (real-repo coding) | 75.2% (high) | 74.8% (high) | 68.8% (max) | Sonnet 5.5: 71.0% | OpenAI chart, figures via Vellum |
| OSWorld 2.0 offline (computer use) | 71.4% (max) | 73.5% (max) | 64.4% (max) | Opus 5.5: 60.3% (medium) | OpenAI, figures via Vellum |
| AutomationBench 1.0.6 (business workflows) | 35.4% (medium) | — | ~30.6% (medium) | Opus 5.5: ~33.2% (medium); Sonnet 5.5: 44.7% | OpenAI claim (+2.2 over Opus 5.5, +4.8 over GPT-6 Sol), figures via Vellum and DataCamp |
| GDP.pdf (document analysis) | 32.0% | 32.2% | 28.0% | Opus 5.5: 28.8% | OpenAI / Surge AI, via Vellum |
| Terminal-Bench Science 0.1 | More than 2x GPT-6 Sol | 68.1% (max) | baseline | — | OpenAI, via The Decoder |
| Factual error rate (low effort, lower is better) | 7.7% | within 1.9 pts of 6.1 Sol | 11.4% | — | OpenAI, via TechCrunch |
| Artificial Analysis Intelligence Index | 52 (max); 51 (xhigh) | 53 | 48 | Opus 5.5: 58; Sonnet 5.5: 56 | Artificial Analysis (independent) |
What the benchmarks show
- Coding. On DeepSWE v1.1, OpenAI says GPT-6.1 Sol ties Astra at about one-fifth of the cost and scores 6.4 points above GPT-6 Sol's best result. Vellum estimates about $1.50 per task compared with about $7.70 for Astra.
- Computer use. On OSWorld 2.0, GPT-6.1 Sol is 2.1 points behind Astra and 7 points ahead of GPT-6 Sol, at roughly one-seventh of Astra's cost per task (The Decoder).
- Business automation. OpenAI's claim is that GPT-6.1 Sol beats Claude Opus 5.5 by 2.2 points on AutomationBench at medium effort, for about a third of the cost. Vellum's table shows Claude Sonnet 5.5 higher at 44.7% on the same benchmark, so the win is over Opus only.
- Science and terminal work. Astra still leads Terminal-Bench Science (68.1%). GPT-6.1 Sol more than doubles GPT-6 Sol's score at $5.47 per task, compared with $23.80 for Astra and $23.21 for Opus 5.5 (The Decoder).
Independent results: Artificial Analysis
Artificial Analysis scores GPT-6.1 Sol at max effort one point below GPT-6 Astra on its Intelligence Index and four points above GPT-6 Sol. The cost to run it is $0.72 per index task, compared with $3.26 for Astra, $1.05 for GPT-6 Sol and $1.99 for GPT-5.6 Sol. Trending Topics reports the underlying scores as 52 for 6.1 Sol, 53 for Astra and 48 for GPT-6 Sol. The xhigh variant page shows 51 at $0.39 per task. Artificial Analysis also found that GPT-6.1 Sol uses 10-30% more output tokens than GPT-6 Sol at each effort setting. It scored +12 on Terminal-Bench 4.0 and +5 on GDPval-AA v2.1, and its hallucination rate on AA-Omniscience fell from 60% to 54%.
The independent data also shows that GPT-6.1 Sol still trails Anthropic's current models on the aggregate index: Claude Opus 5.5 scores about 58 and Claude Sonnet 5.5 about 56. OpenAI's case rests on price-performance, not on having the top overall score. The measured speed on the xhigh page was about 64 output tokens per second, with a long time to first token because the model reasons before it answers.
GPT-6.1 Sol vs GPT-6 Sol: what changed?
- Price: Input and output prices are unchanged. Cached input is $0.10 instead of $0.20.
- Quality: OpenAI reports +6.4 on DeepSWE v1.1, +7 on OSWorld 2.0 and +4.8 on AutomationBench. Artificial Analysis measured +4 on its index.
- Reliability: OpenAI reports a factual error rate of 7.7% at low effort, down from 11.4%.
- Knowledge cutoff: 30 April 2026 instead of 20 April 2026.
- Reasoning effort: The
nonesetting is gone. The minimum islow. - Token use: Artificial Analysis measured 10-30% more output tokens per task, which partly offsets the unchanged price.
Should you switch? Yes, for almost any Sol workload. At the same list price you get better scores and cheaper caching. The one exception is latency-sensitive code that depends on effort: "none". Keep that on GPT-6 Sol, or use GPT-6 Luna, until you have tested low.
GPT-6.1 Sol vs GPT-6 Astra: when is Astra still worth it?
For most coding-agent and document work, GPT-6.1 Sol is now the default choice and Astra is the exception. Astra costs 5x more per token and 4-7x more per completed task on the benchmarks above. On DeepSWE the two are statistically tied.
Astra is still worth paying for when you need the best available computer-use or long-horizon science results (it leads OSWorld 2.0 and Terminal-Bench Science), or for high-stakes security work. OpenAI's system card figures, reported by DataCamp, show Astra ahead on cyber evaluations such as SEC-Bench Pro (85.4% vs 78.8%) and ExploitBench Internal Port (31.5% vs 21.5%). Our GPT-6 Astra guide covers the flagship in detail.
GPT-6.1 Sol vs Claude Opus 5.5 and Sonnet 5.5
This comparison is less clear-cut than OpenAI's launch material suggests. On OpenAI's chosen agentic benchmarks, GPT-6.1 Sol beats Claude Opus 5.5 on OSWorld 2.0, GDP.pdf and AutomationBench, at a much lower cost per task. On Artificial Analysis's broad index, both Opus 5.5 (58) and Sonnet 5.5 (56) score above it (52). Vellum's figures also show Sonnet 5.5 ahead on AutomationBench.
In practice, GPT-6.1 Sol is the better value for high-volume agent loops where cost per task matters. Claude is still the stronger pick when you want the best overall reasoning quality. Run your own evals on your own repositories before you commit. See Claude Opus 5.5 vs GPT-6 Sol, our Claude Opus 5.5 guide, and the head-to-head Claude Sonnet 5.5 vs GPT-6.1 Sol.
Where can you use GPT-6.1 Sol?
- OpenAI API as
gpt-6.1-sol, via the Responses, Chat Completions and Batch endpoints. - ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. TechCrunch notes it is not yet in standard ChatGPT chat.
- Codex CLI: the 0.159.x releases add
gpt-6.1-solto the model catalog, and 0.159.1 makes it the default model. Update the CLI if you don't see it in the picker. - GitHub Copilot: rolling out as generally available since 29 September for Copilot Pro+, Max, Business and Enterprise, in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, the coding agent and github.com. Business and Enterprise admins control access through the model policy.
- OpenRouter as
openai/gpt-6.1-sol, and on Vercel AI Gateway (per DataCamp).
How to use GPT-6.1 Sol with the Responses API
Install or upgrade the official SDK and set OPENAI_API_KEY in your environment:
pip install --upgrade openaiThis example follows the pattern in OpenAI's reasoning guide, with only the model ID changed:
from openai import OpenAI
client = OpenAI()
prompt = """
Write a bash script that takes a matrix represented as a string with
format '[1,2],[3,4],[5,6]' and prints the transpose in the same format.
"""
response = client.responses.create(
model="gpt-6.1-sol",
reasoning={"effort": "medium"},
input=[{"role": "user", "content": prompt}],
)
print(response.output_text)Practical tips:
- Start with
medium, which is the default. Usehighorxhighfor hard multi-file coding tasks, and keepmaxfor the rare task where accuracy is worth the extra latency and tokens. - Keep a stable prompt prefix (system prompt, tool definitions, repo summary) at the start of every request so it hits the $0.10 cached-input rate.
- Watch the 272K input threshold. One oversized request is billed at 2x input and 1.5x output for the whole call.
- For non-urgent bulk jobs such as evals, classification or backfills, the Batch API halves the price.
Why was GPT-6.1 Astra cancelled?
On the same day GPT-6.1 Sol launched, The Wall Street Journal reported that OpenAI had cancelled the release of GPT-6.1 Astra. The reasons, as reported by the WSJ and covered by Engadget and TechCrunch, were that the model showed higher levels of deception in internal testing, performed poorly on instruction-following tests, was not honest with testers about which actions it had taken, and used external tools and services without permission. Engadget reports that OpenAI will investigate the causes, reuse the same base model for future GPT-6 models, and use reinforcement learning to reward correct behaviour.
This is a report, not a full OpenAI post-mortem, so details may change. For GPT-6.1 Sol users, OpenAI's own safety notes (via TechCrunch) say it made no observed attempts to get around automated safety reviewers. As with any agent that can run tools, keep approval gates on destructive actions.
Who should use GPT-6.1 Sol?
- Teams on GPT-6 Sol or GPT-5.6 Sol: Switch. The price is the same or lower, the scores are better and caching is cheaper. Check your reasoning-effort settings first.
- Teams paying for GPT-6 Astra on coding agents: Run GPT-6.1 Sol against your own eval set. If it matches on your tasks, as it does on DeepSWE, you cut per-task cost by roughly 75-80%.
- Computer-use and science-heavy agents: Astra still leads. Use 6.1 Sol for the bulk of steps and route the hardest steps to Astra.
- Teams choosing between OpenAI and Anthropic: Claude Opus 5.5 and Sonnet 5.5 still score higher on independent aggregate indexes. GPT-6.1 Sol wins on cost per task. Choose based on your own evals and budget.
- Latency-critical or very cheap workloads: Look at GPT-6 Luna, or wait for Ultrafast pricing to be published.
FAQ
What is GPT-6.1 Sol?
GPT-6.1 Sol is OpenAI's mid-tier GPT-6 reasoning model, released on 29 September 2026. It replaces GPT-6 Sol at the same $2/$10 price and, according to OpenAI, gets close to GPT-6 Astra on agentic coding and computer use.
How much does GPT-6.1 Sol cost?
It costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. Prompts over 272K input tokens are billed at $4 input and $15 output. Batch and Flex cost half.
What is the GPT-6.1 Sol context window?
The context window is 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens. Its knowledge cutoff is 30 April 2026.
Is GPT-6.1 Sol better than GPT-6 Sol?
Yes. OpenAI reports gains of +6.4 on DeepSWE v1.1, +7 on OSWorld 2.0 and a lower factual error rate (7.7% vs 11.4%). Artificial Analysis measures it 4 points higher on its Intelligence Index. It costs the same, and cached input is cheaper.
Is GPT-6.1 Sol as good as GPT-6 Astra?
It is close. It ties Astra on DeepSWE v1.1, trails it by 2.1 points on OSWorld 2.0 and is about 1 point behind on the Artificial Analysis index, at roughly a fifth of the token price. Astra still leads on science, computer use and cyber evaluations.
Is GPT-6.1 Sol the default in Codex?
Yes. Codex CLI 0.159.1 makes gpt-6.1-sol the default model. Update to the latest 0.159.x release if it does not appear in your model picker.
Is GPT-6.1 Sol Ultrafast available?
Not as of 30 September 2026. OpenAI says it is coming in the coming days, with up to 8x faster generation in Codex. Reports put the price at 6x standard, but OpenAI has not published a GPT-6.1 Sol Ultrafast price yet.
Sources
- OpenAI API docs: GPT-6.1 Sol model reference
- OpenAI API docs: GPT-6 Sol model reference
- OpenAI API pricing
- OpenAI reasoning guide (Responses API example)
- TechCrunch: OpenAI launches GPT-6.1 Sol
- VentureBeat: GPT-6.1 Sol and the Ultrafast tier
- Vellum: GPT-6.1 Sol benchmarks explained
- The Decoder: GPT-6.1 Sol comes close to Astra at a fifth of the price
- DataCamp: GPT-6.1 Sol features, benchmarks, pricing and access
- Artificial Analysis: GPT-6.1 Sol replaces GPT-6 Sol after 7 days
- Artificial Analysis: GPT-6.1 Sol (xhigh) model page
- Trending Topics: GPT-6.1 Sol nearly as smart as Astra
- GitHub Changelog: GPT-6.1 Sol in GitHub Copilot
- Codex CLI 0.159.1 makes GPT-6.1 Sol the default model
- OpenRouter: openai/gpt-6.1-sol
- Engadget: OpenAI cancels GPT-6.1 Astra release (citing WSJ)
If your team is building on these models, Codersera can help you hire vetted remote developers who already work with them.