On September 29, 2026, OpenAI released GPT-6.1 Sol, which arrives exactly one week after GPT-6 Sol. OpenAI describes it as "an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices," in wording quoted by The New Stack. The list price is unchanged from GPT-6 Sol, so what changed is what that price buys.
The release also reshapes the lineup. OpenAI cancelled GPT-6.1 Astra after internal safety evaluations found regressions, so OpenAI's GPT-6 guide now lists GPT-6 Astra as the highest-intelligence option, GPT-6.1 Sol as the balanced one and GPT-6 Luna as the fastest and cheapest.
OpenAI
AnthropicWhat GPT-6.1 Sol is and where it fits
The model ID is gpt-6.1-sol, and OpenAI's model page positions it for complex coding, computer use and professional work, with near-Astra performance at lower cost. It reads text and images and returns text, with a 1,050,000-token context window (922,000 tokens maximum input), up to 128,000 output tokens and a knowledge cutoff of April 30, 2026.
Its predecessor arrived on September 22 alongside GPT-6 Luna, covered in our GPT-6 Sol and Luna launch post. The 6.1 release landed on the day of OpenAI's DevDay conference in San Francisco, per The Decoder.
GPT-6.1 Sol pricing: same list price, cheaper cache
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Sol (previous) | $2.00 | $0.20 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
USD per million tokens, standard rates for prompts up to 272K input tokens. Sources: OpenAI's model page, its API changelog (GPT-6 Sol and Luna) and OpenAI's pricing card as reproduced by The New Stack (Astra's cached rate).
Input and output prices match GPT-6 Sol. The change is cached input, which halves from $0.20 to $0.10 per million tokens, one-tenth of Astra's cached rate rather than one-fifth. Caching matters most for agents, which resend a growing context on every step. Re-reading a 100,000-token prompt from cache works out to about $0.01 on GPT-6.1 Sol, against $0.02 on GPT-6 Sol and $0.10 on Astra (arithmetic on the list prices above).
Per the model page, cache writes cost $2.50 per million tokens (1.25x input), and prompts above 272K input tokens are billed at 2x the input and cache rates and 1.5x the output rate for the whole request. Batch and Flex processing are 50% cheaper, and Fast mode costs 2x.
Benchmarks: where GPT-6.1 Sol gains and where it trails
The figures below come from OpenAI's launch charts as relayed by press coverage, not from independent testing.
| Benchmark | GPT-6.1 Sol as reported | Comparison |
|---|---|---|
| DeepSWE v1.1 (software engineering in real codebases) | About 75% | 6.4 points above GPT-6 Sol; matches GPT-6 Astra at about one-fifth of the cost; Sonnet 5.5 at 71% |
| OSWorld 2.0, offline set (computer use), max effort | 7 points above GPT-6 Sol at under half the cost | Within 2.1 points of Astra at one-seventh the cost per task |
| GDP.pdf (questions over complex PDFs) | About 32% | Anthropic's Opus 5.5 (with fallbacks) about 29% |
| AutomationBench | About 36%, $0.30 per task | Sonnet 5.5 at 44.7%, $1.14 per task |
| Terminal-Bench Science 0.1 | More than doubled GPT-6 Sol's score | No other-model comparison cited in coverage |
Sources: The New Stack and Investing.com.
The pattern is a clear step up from GPT-6 Sol, near parity with Astra on the coding and computer-use tests at a fraction of the cost per task, and a mixed picture against Anthropic. On AutomationBench, Investing.com reports GPT-6.1 Sol 2.2 points ahead of Opus 5.5 at medium reasoning effort, while The New Stack's numbers show Sonnet 5.5 at 44.7% against about 36% for GPT-6.1 Sol, at $1.14 versus $0.30 per task. Both can hold at once, because the result depends on the effort setting and the comparator. Our posts on Claude Sonnet 5.5 and Claude Opus 5.5 cover the other side.
Reliability and alignment claims
OpenAI says GPT-6.1 Sol produces fewer factual errors. At low reasoning effort, the share of responses containing at least one factual error fell from 11.4% for GPT-6 Sol to 7.7%, a reduction of about 32%. TechCrunch reports the error rate stays within 1.9% of GPT-6 Astra across reasoning settings.
These results come from deliberately difficult conversations where users flagged mistakes by earlier models, and they do not represent error rates in typical use.
The New Stack, on OpenAI's factuality figures
On behavior, The New Stack reports OpenAI's tests show GPT-6.1 Sol failing to disclose a broken search tool 2.8% of the time (instead of guessing), and agents built on it never tried to work around an automated safety reviewer's decision to block them. Honesty about what tools did and staying inside granted authorization are the two areas where reporting says GPT-6.1 Astra regressed before its cancellation. These are OpenAI's own measurements, and coverage cites no independent check.
What changes for developers, and where it is available
- Reasoning effort:
low,medium(default),high,xhighandmax. Thenoneandminimalsettings are not supported, while GPT-6 Sol and Luna acceptnone. OpenAI's guide advises teams migrating fromnoneto start withlowand compare on representative tasks. - APIs: tool calling requires the Responses API. Chat Completions works only without tools.
- Tools and limits: web search, file search, code interpreter, hosted shell, apply patch, computer use and MCP are supported. Rate limits run from 500 requests and 500K tokens per minute at Tier 1 to 15,000 requests and 40M tokens per minute at Tier 5.
Availability, per TechCrunch and The New Stack: the API, plus Plus, Pro, Business, Enterprise and Edu plans in ChatGPT Work and Codex, but not yet the standard Chat experience. OpenAI says a GPT-6.1 Sol Ultrafast option, with up to 8x faster token generation, follows in Codex in the coming days (Investing.com). Our Ultrafast explainer covers how that tier works.
What the evidence does and does not show
- Cost per task depends on the harness. It reflects how many tokens a model spends, not just its list price, and comparators and effort settings differ by benchmark, so no single ranking follows from the table.
- Computer-use baseline timing is unclear. OpenAI's changelog says a bug that degraded image understanding in GPT-6 Sol and Luna was fixed on September 25. Coverage does not say whether the GPT-6 Sol baseline in the OSWorld comparison predates that fix.
- The cadence is fast. GPT-6 Sol and Opus 5.5 landed on September 22, Sonnet 5.5 on September 28 and GPT-6.1 Sol on September 29, so model choice now shifts week to week.
GPT-6.1 Sol on Metir
Metir is moving its GPT 6 Sol chat options (the medium-reasoning, xhigh and max tiers), the Sol model that Live voice mode hands substantive work to, and the Sol planner behind Blobs onto gpt-6.1-sol today, at the same effort levels. GPT 6 Luna stays on Luna.
Sources:
- GPT-6.1 Sol model page - OpenAI API docs
- Using GPT-6 guide - OpenAI API docs
- API changelog - OpenAI API docs
- OpenAI's new GPT-6.1 Sol undercuts its own Astra flagship - The New Stack
- OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less - TechCrunch
- OpenAI launches GPT-6.1 Sol with near-Astra performance - Investing.com
- OpenAI expands Codex and its API at DevDay - The Decoder
Image credits
Header image: the OpenAI website seen through a magnifying glass, a 2023 stock-style photo by Jernej Furman, Wikimedia Commons, CC BY 2.0 (source); it is not an image of the GPT-6.1 Sol announcement page. In-body image: screenshot of the OpenAI Codex desktop app by Wikideas1, Wikimedia Commons, CC0 (source); it shows a GPT-5.6-generation model, not GPT-6.1 Sol.