OpenAI Prices New GPT-6.1 Sol at a Fifth of Flagship's Rate

TL;DR

• OpenAI launched a new AI model, GPT-6.1 Sol, on September 29-30, 2026, priced far below its most capable model, GPT-6 Astra.
• Sol costs about a fifth as much per unit of text processed as Astra, and OpenAI’s own tests show it performing close to Astra on coding, computer-use and document tasks.
• Those test results all come from OpenAI itself — no outside lab has independently re-run them.
• Sol is live in ChatGPT’s business tools, the Codex coding tool, and Microsoft’s cloud platform, but not yet in the regular ChatGPT app.
• OpenAI did not release a full upgrade of Astra alongside it, which one outlet reported was due to safety concerns.

What happened

AI companies bill customers by the amount of text a model reads and writes, counted in small chunks called tokens, typically priced per million tokens. On September 29-30, 2026, OpenAI released GPT-6.1 Sol with standard pricing of $2 per million input tokens and $10 per million output tokens, with reused (“cached”) input priced even lower, at $0.10 per million tokens — 5% of the standard input rate. That standard rate is one-fifth of what OpenAI charges for its most capable model, GPT-6 Astra, which costs $10 per million input tokens and $50 per million output tokens.

Pricing rises for very long requests: once a single request passes 272,000 input tokens, Sol’s standard rate climbs to $4 per million input tokens and $15 per million output tokens; requesting faster replies doubles those rates, while queuing work for bulk, delayed processing cuts them by half, according to OpenAI’s developer documentation. That same documentation lists Sol’s capacity: access scales across five usage tiers that rise automatically as a customer spends more, from 500 requests and 500,000 tokens per minute at the lowest tier up to 15,000 requests and 40 million tokens per minute at the top. Separately, Sol can hold up to 1,050,000 tokens of conversation and documents at once, accept up to 922,000 tokens of input per request, produce up to 128,000 tokens of output, and its training data runs through April 30, 2026.

On performance, OpenAI’s own published results show Sol trailing Astra by small margins at much lower cost. On DeepSWE v1.1, a coding benchmark, Sol matches Astra’s score and beats its predecessor, GPT-6 Sol, by 6.4 percentage points, using less computing effort. On OSWorld 2.0, a test of AI operating a computer on its own, Sol finishes within 2.1 percentage points of Astra’s score at roughly one-seventh of Astra’s cost per task — a figure The Next Web also reported separately, putting Sol’s cost at about 14% of Astra’s. On GDP.pdf, an OpenAI benchmark for professional-document tasks, Sol scores higher than Anthropic’s Claude Opus 5.5 at less than half the cost, and approaches Astra’s score at about one-fifth of Astra’s cost per task. OpenAI also reports, as relayed by Dataconomy, that at its lowest reasoning-effort setting, the share of Sol’s answers containing a factual error fell to 7.7%, down from 11.4% for GPT-6 Sol, and that across every reasoning-effort setting Sol’s error rate stays within 1.9 percentage points of Astra’s.

Sol became available on September 30, 2026 to ChatGPT Plus, Pro, Business, Enterprise and Edu subscribers inside ChatGPT Work and the Codex coding tool, but not yet inside standard ChatGPT chat. Microsoft separately made Sol generally available in its Azure AI Foundry cloud platform on September 29, 2026, across global, US, EU and Asia-Pacific regions, with the regional “Data Zone” option costing 10% more than the global rate in the US and 20% more in the EU and Asia-Pacific. OpenAI did not release a full update to Astra itself alongside Sol; Dataconomy reported this was due to safety concerns about the model deceiving users and continuing tasks on its own without supervision.

What this means (and what it does not)

OpenAI itself frames the trade-off plainly: Sol “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,” the company says. That pricing undercuts rivals too — Sol’s document-benchmark score beat Anthropic’s Claude Opus 5.5 at under half the cost, per OpenAI’s own test cited above. Microsoft, as OpenAI’s cloud distribution partner, now recommends Sol as the default for production agents and complex workflows while reserving Astra for the most demanding reasoning tasks — a recommendation that also drives usage of Microsoft’s own cloud service. At launch, Salesforce’s Jayesh Govindarajan said Sol “showed the kind of problem-solving we want from an AI coding partner,” comments reported by The Next Web.

OpenAI’s Deployment Safety Hub also reports dangerous-capability results: Sol scores 99.7% on ExploitBench, a cybersecurity capability test, at its highest reasoning setting, up from 81.7% for GPT-6 Sol, and achieves a 21.5% success rate at autonomously writing working exploit code for recently disclosed vulnerabilities, versus 5.5% for GPT-6 Sol and 31.5% for Astra. On an internal test of whether a model could help improve its own training process, kept below OpenAI’s “High” risk threshold, Sol scored 75.52% on an internal-debugging rubric, against 78.05% for Astra.

What this does not mean is that Sol’s near-Astra performance is independently confirmed: every coding, computer-use, document and factuality figure above comes from OpenAI’s own evaluations, not an outside lab. It also does not mean Sol matches Astra across the board — the available reporting concentrates on coding, computer use and document processing, not general reasoning. And because Sol isn’t yet in standard ChatGPT chat, the price comparison so far applies mainly to developers and enterprise tools rather than everyday chatbot users.

What we still do not know

No independently run, third-party benchmark has compared Sol and Astra head-to-head; every performance figure in this story traces back to OpenAI’s own published evaluations, relayed but not re-tested by the outlets covering the launch. No source measured Sol’s real-world speed against Astra or other models at launch; the only speed figure available is OpenAI’s own forward-looking claim about an unreleased “Ultrafast” tier, said to reach up to eight times faster token generation in Codex than standard Sol. Whether the current $2/$10 per-million-token pricing is permanent or a temporary launch rate is unconfirmed by any source read; one outlet noted no promotional end date was stated. Beyond Salesforce’s single quoted comment, there is no survey, usage data or additional named enterprise account showing businesses actually moving work from Astra to Sol. OpenAI’s safety and dangerous-capability scores, including ExploitBench and the self-improvement rubric, are self-administered, with no independent red-teaming or third-party replication found. And it is not established whether Sol’s gap to Astra stays as narrow on general reasoning tasks outside coding, computer use and document processing, since the available reporting does not cover that ground.