GPT-6.1 Sol: Near-Astra Work at One-Fifth the List

GPT-6.1 Sol is the Sol-class upgrade that approaches Astra on coding, computer use, and professional work at one-fifth of Astra's Standard list: $2 / $10 versus $10 / $50. Hold OpenAI's announcement charts as self-reported, not as LLM Stats verified.

The pitch is cost and throughput for near-Astra agentic jobs, not a claim that Sol beats Astra overall. Astra still leads hardest science (TB-Science Astra 68.1% vs Sol Max 57.02%). Prefer Astra when peak score matters; prefer 6.1 Sol when the Sol sticker and cache math matter.

At a Glance

• Catalog id: gpt-6.1-sol
• Organization: OpenAI
• Status: API + ChatGPT Work / Codex released September 29, 2026 (not yet in Chat)
• Model id: gpt-6.1-sol
• Pricing: $2 / $10 Standard; cache $0.10; writes $2.50
• vs Astra: one-fifth Standard input and output ($2 / $10 vs $10 / $50)
• Long prompts:>272K input → 2× input/cache, 1.5× output for the full request
• Context: 1,050,000 / 128K max output
• Cutoff: April 30, 2026
• Effort: low / medium (default) / high / xhigh / max; no none or minimal
• Modalities: text + image in, text out
• Self-reported: DeepSWE 75.22% High · OSWorld 71.42% Max · TB-Science 57.02% Max · AutomationBench 36.10% Max · GDP.pdf 32.0% High

What's New

OpenAI positions GPT-6.1 Sol as the successor to GPT-6 Sol that closes much of the gap to Astra on agentic coding, computer use, and professional work while staying in the Sol price class. Context stays at 1,050,000 with 128,000 max output. reasoning.effort runs low through max, defaulting to medium. There is no none or minimal rung.

Shipping today through the OpenAI API as gpt-6.1-sol, plus ChatGPT Work / Codex for Plus, Pro, Business, Enterprise, and Edu. ChatGPT Chat access is not included yet. An Ultrafast variant is mentioned as coming in days (OpenAI says up to 8× faster token generation); Ultrafast pricing is not published.

What the card supports

Streaming, function calling, and structured outputs. Responses is preferred for tools; Chat Completions is supported without tool calling. Batch is supported; fine-tuning is not. Responses API tools include web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. Input is text and image; output is text only.

Benchmarks

OpenAI published self-reported announcement charts on the GPT-6.1 Sol announcement page. Numbers below are theirs, not LLM Stats verified. Read by job: coding, computer use, science, automation, then professional docs. Effort ladders matter: DeepSWE and GDP.pdf peak at High on these charts, not Max.

Catalog domains

Ten capability benches from the catalog sync, grouped by job. Coding / agents and documents echo the announcement peaks with effort notes. Health is the main added depth from the system card addendum (length-adjusted preferred). Absolute bars only; no peer inventions.

Coding

DeepSWE v1.1 headlines at 75.22%on High ($0.65/task on OpenAI's chart), ahead of Max at 71.90% ($1.57). The Full ladder: Low 64.38% / $0.17, Medium 73.01% / $0.42, High 75.22% / $0.65, Xhigh 71.90% / $0.79, Max 71.90% / $1.57. OpenAI says Sol matches Astra at roughly one-fifth the cost and beats GPT-6 Sol's best by 6.4 pp at lower effort and cost. Prefer High when citing a single DeepSWE number.

Computer use

OSWorld 2.0 offline partial lands at 71.42% on Max ($1.27/task). High is 69.56% / $0.96; Xhigh is 69.38% / $1.05. OpenAI says that is +7 pp vs GPT-6 Sol at Max at less than half the cost, and within 2.1 pp of Astra Max at about one-seventh the cost.

Science

Terminal-Bench Science 0.1 reaches 57.02% on Max at about $5.47per task (High 51.14% / $2.76; Xhigh 53.71% / $2.89). OpenAI's announcement peers: Opus 5.5 about $23.21/task, Astra about $23.80/task. Astra's peak score remains 68.1%. Sol more than doubles GPT-6 Sol Max at less than half of that prior Sol cost. This is the clearest reminder that Astra still leads hardest science.

Automation and professional work

AutomationBench Max is 36.10% / $0.30 (Medium 31.70% / $0.19; High 33.20% / $0.23; Xhigh 35.50% / $0.25). OpenAI says +2.2 pp vs Opus 5.5 at medium at about one-third the cost, and +4.8 pp vs GPT-6 Sol at the same setting.

GDP.pdf peaks at High: 32.0% / $0.35 (Xhigh 31.8% / $0.37; Max 31.0% / $0.42). OpenAI says higher than Opus 5.5 with fallbacks at less than half the cost, and approaches Astra at about one-fifth the cost.

Health (system card)

From the system card addendum, prefer length-adjusted cites. HealthBench Professional LA 64.2% (raw 67.2%) sits within 0.5 pp of Astra 64.7. HealthBench Hard LA is 36.2%. MentalHealthBench overall is 57.9% at max (non-acute 57.3; high acuity 59.1; emergent 58.0; Astra 58.7; GPT-6 Sol 54.2). HealthBench LA 58.5% and Consensus LA 96.0% (near-ceiling) round out the rail above.

Factual error rate

Optional callout, lower is better. At low effort, GPT-6 Sol 11.4% moves to GPT-6.1 Sol 7.7% (about a 32% relative reduction). Across settings, OpenAI says Sol stays within 1.9 pp of Astra at less than one-fifth the cost. The eval is on flagged error conversations, not typical traffic.

Pricing

First-party OpenAI Standard, USD per million tokens. The headline is Sol-class list at one-fifth of Astra on both input and output, with a cheaper cached-input line than GPT-6 Sol.

When to Use It

• Good fit:shops that want near-Astra coding, computer use, or professional-doc work through the OpenAI API without paying Astra's $10 / $50 sticker; 1.05M context jobs; teams that care about cache hits at $0.10.
• Prefer Astra when peak score on the hardest science or agentic jobs matters (TB-Science Astra 68.1% still leads Sol Max 57.02%).
• Default effort is medium. Set reasoning.effort explicitly for High (DeepSWE / GDP.pdf peaks) or Max (OSWorld / TB-Science / AutomationBench peaks on these charts).
• Upgrade path from GPT-6 Sol: same price class, stronger agentic coding and computer-use tables on OpenAI's charts, cheaper cached input.
• Budget the 272K multiplier if prompts are huge. Fast is 2×. Regional endpoints add about 10% where available.

Caveats

• Self-reported charts.Scores are OpenAI's announcement numbers, not LLM Stats verified. DeepSWE High beats Max on their chart; cite the effort rung with the score.
• Not stronger than Astra overall. Astra still leads hardest science. Prefer Astra for peak score; Sol for cost and throughput.
• Not yet in Chat. API and ChatGPT Work / Codex for paid tiers today; ChatGPT Chat comes later.
• Default effort is medium. No none or minimal. Set effort explicitly when the chart peak matters.
• Preparedness. OpenAI treats 6.1 Sol as Critical cybersecurity and High biological/chemical, with the same safeguard stack as Astra (Trusted Access / Daybreak for advanced cyber).
• Knowledge cutoff April 30, 2026.
• No fine-tuning on this card. Ultrafast pricing is not published yet.

Outlook

This post ages when LLM Stats verifies DeepSWE, OSWorld, or Terminal-Bench Science, when Ultrafast pricing lands, or when ChatGPT Chat access opens. Through-line: near-Astra agentic work in the Sol price class at $2 / $10, with self-reported charts that still leave Astra ahead on hardest science.

Sources: OpenAI GPT-6.1 Sol announcement, system card addendum, model docs.