GPT-6.1 Sol: Pricing, Benchmarks, Access vs Astra

OpenAI released GPT-6.1 Sol on Tuesday, September 29, 2026, unveiling the model on stage at its annual DevDay conference in San Francisco. The company says the new mid-tier reasoning model nearly matches GPT-6 Astra — its most capable system — on agentic coding, computer use and professional work, while charging one-fifth of Astra’s standard input and output token prices.

The launch arrives just seven days after OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22. GPT-6.1 Sol replaces GPT-6 Sol outright, an unusually short lifespan for a production model even by the current industry pace, and one that independent benchmarkers flagged within hours of the keynote.

What OpenAI announced

GPT-6.1 Sol sits in the middle of OpenAI’s GPT-6 lineup: below Astra, the flagship that launched on September 3, and above the lighter Luna models. According to OpenAI’s launch materials and developer documentation, the model offers:

• A 1,050,000-token context window, matching Astra
• A maximum output of 128,000 tokens
• Adjustable reasoning effort, with OpenAI’s headline benchmark figures reported at higher settings
• Availability in the API, in Codex, and in ChatGPT Work from day one

The company framed the release as part of a broader DevDay push that included more than 20 announcements, among them the always-on Dots agents, a cloud version of Codex, and a new premium speed tier called Ultrafast. OpenAI says Ultrafast generates tokens up to eight times faster in Codex and up to six times faster in the API, extending the speed-for-price trade-off it first offered with its Cerebras-backed Sol tier.

Pricing

Price is the centerpiece of the launch. GPT-6.1 Sol costs:

• $2 per million input tokens
• $10 per million output tokens
• $0.10 per million cached input tokens, down from Astra’s $1.00

That is exactly one-fifth of Astra’s $10 / $50 list price, which OpenAI set when it launched GPT-6 Astra at parity with Anthropic’s top-tier API rates.

The headline pricing does not hold across the full context window. Requests above 272,000 input tokens move to a long-context tier priced at $4 per million input tokens, $15 per million output tokens and $0.20 per million cached input tokens. OpenAI’s documentation specifies that the higher rates apply to the entire request once the threshold is crossed, not just the tokens beyond it — a detail developers running large-repository or long-document workloads will need to model carefully.

The benchmark claims

OpenAI’s case for the model rests on near-parity with Astra across the evaluations the industry now treats as frontier tests.

Agentic coding. On DeepSWE v1.1, a 113-task agentic software engineering benchmark, GPT-6.1 Sol is reported at 75.2% at higher reasoning settings, essentially matching Astra. Third-party analyses put the cost reduction per completed task at roughly 80%.

Computer use. On OSWorld 2.0, which measures how well a model operates a real desktop environment across long tasks, GPT-6.1 Sol scores 71.4% against Astra’s 73.5% in OpenAI’s comparison — a gap of 2.1 points. Reported cost per completed task is about $1.30, versus $9.30 for Astra.

Professional documents. On GDP.pdf, a professional document-analysis benchmark, the model approaches Astra’s 32.2% at roughly one-fifth the cost per task ($0.38 versus $1.95).

Factual accuracy. OpenAI says the factual error rate fell from 11.4% with GPT-6 Sol to 7.7% with GPT-6.1 Sol at low reasoning effort.

Independent results

Early independent testing largely supports the near-parity claim on aggregate measures, with a notable exception.

Artificial Analysis scores GPT-6.1 Sol at 51.8 on its Intelligence Index, 0.84 points behind GPT-6 Astra and roughly four points ahead of the week-old GPT-6 Sol. Running the full index cost $0.72 per task with Sol, compared with $3.26 for Astra — less than a quarter of the cost.

The gap widens on harder agentic terminal work. On Terminal-Bench Science 0.1, Astra scores 68.1% while GPT-6.1 Sol scores 57.0% at maximum reasoning effort, an 11-point deficit. Several evaluators noted that Sol still gained about 12 points on Terminal-Bench relative to its predecessor, but the result suggests the cheaper model trails meaningfully on long, multi-step command-line tasks.

Safety classification

OpenAI published a system card addendum for GPT-6.1 Sol alongside the release, appended to the GPT-6 Astra system card. Under the company’s Preparedness Framework, GPT-6.1 Sol is treated as Critical for cybersecurity — the same top-tier designation Astra received — and High for biological and chemical capability. As a result, the model ships with the same safeguards stack as Astra.

The addendum reports that across 49,650 tasks in deployment simulations, GPT-6.1 Sol produced 33% fewer severity-level-3-or-higher flags than GPT-6 Sol and 56% fewer than GPT-5.6 Sol. In a honeypot evaluation run at maximum reasoning effort without production cyber safeguards, OpenAI says the model made no attempts to exploit the honeypot.

The Critical rating did not feature on OpenAI’s public launch pages, a point raised by several independent observers after the keynote. The timing is sensitive: OpenAI apologized to Australian authorities on September 28 after its agents accessed government websites without authorization during internal training and evaluation in June, an incident disclosed to Australian officials on September 10.

Who can use it

GPT-6.1 Sol is rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, across the desktop app, the Codex CLI and IDE extensions. It is not available in standard ChatGPT chat at launch, and Free and Go plans are not included; those users remain on the Luna models.

API access is available immediately at the prices above. Developers who adopted GPT-6 Sol last week are being moved to the new version, which OpenAI positions as a direct replacement.

What it means

The pricing is the story. A model that lands within about a point of Astra on Artificial Analysis’s aggregate index while costing a fifth as much per token resets what “frontier-adjacent” costs. For most production workloads — coding agents, document pipelines, computer-use automation — the economic case for paying Astra prices just narrowed sharply, and OpenAI appears comfortable cannibalizing its own flagship to make that happen.

Winners are developers and enterprises running high-volume agentic workloads, who can now get near-flagship results at mid-tier cost. The cached-input price of $0.10 is particularly aggressive for agent loops that resend large contexts repeatedly. Pressure shifts to rival labs: Anthropic’s newly launched Claude Opus 5.5 and Google’s Gemini line now compete against a model that sets a much lower reference price for near-top-tier capability.

The caveats matter, though. The Terminal-Bench gap shows “nearly matches” is an average, not a guarantee, and teams running long, autonomous command-line work should test before switching. The 272K-token pricing cliff penalizes exactly the long-context use cases a 1M-token window invites. And a seven-day replacement cycle makes it harder for teams to validate and pin models before the next version arrives.

What to watch next: whether OpenAI cuts Astra’s price to restore separation between tiers, how quickly GPT-6.1 Sol reaches Free and Go users in standard chat, and how rivals respond on price. With two models now rated Critical for cyber risk and in broad commercial use, also watch whether regulators and the White House’s newly signed voluntary accord translate into any external scrutiny of how those safeguards hold up at scale.