GPT-6.1 Sol Is Here: Has OpenAI Locked Up the AI Lead?

GPT-6.1 Sol is not OpenAI’s new top model. That distinction still belongs to GPT-6 Astra. The launch matters because OpenAI has moved much of Astra’s agentic performance into a model priced for repeated production use: $2 per million input tokens and $10 per million output tokens, exactly one-fifth of Astra’s standard rates.

That changes the competitive question. A benchmark winner can be too expensive to run across thousands of coding jobs, document reviews, or browser workflows. GPT-6.1 Sol is meant to close the quality gap without carrying the flagship bill.

OpenAI is expanding the market below Astra, not replacing it

The product line now has a clear three-tier shape. Astra handles the hardest work where the final few points of accuracy justify a premium. GPT-6.1 Sol targets complex everyday work that still needs long reasoning and tools. Luna handles focused, high-volume jobs where unit economics dominate.

That makes 6.1 Sol a threat to more than its predecessor. It offers GPT-6 Sol users a capability upgrade at the same headline input and output prices, while giving teams on expensive frontier models a lower-cost alternative for production traffic.

Five evaluations point to the same agentic upgrade

The common thread is execution over time. These are not trivia tests. They require a model to inspect files or interfaces, choose tools, recover from intermediate results, and finish a multi-step objective. That is the workload OpenAI has optimized most aggressively.

Coding matched Astra while computer use closed most of the gap

On DeepSWE 1.1, GPT-6.1 Sol matched Astra and exceeded GPT-6 Sol’s best result by 6.4 percentage points at a lower reasoning setting and cost. For development teams, that is the cleanest argument for moving routine repository work away from the premium model.

AutomationBench tests end-to-end work in sales, marketing, operations, support, finance, and HR. At medium reasoning effort, 6.1 Sol beat Opus 5.5 by 2.2 points at roughly one-third of the cost per task. It also beat GPT-6 Sol by 4.8 points under the same setting.

On the OSWorld 2.0 offline set, 6.1 Sol gained seven points over Sol and came within 2.1 points of Astra. Its cost per task was roughly one-seventh of Astra’s. That price gap compounds quickly because computer-use agents make many model calls while navigating, checking, and retrying.

The $0.10 cached-input rate is the real agent discount

Cached input costs 95% less than ordinary 6.1 Sol input and half as much as GPT-6 Sol’s cached input. Reused repository instructions, tool schemas, product documentation, and policy context therefore become much cheaper on subsequent calls. The model is priced around persistent agents rather than isolated prompts.

There is an important long-context threshold. Inputs above 272,000 tokens are billed at twice the input and cache rates and 1.5 times the output rate for the entire request. A 1.05-million-token window is available, but indiscriminately loading a whole repository can still erase part of the cost advantage.

A 1.05M context window with Responses required for tools

Tool calling requires the Responses API. Chat Completions works only without tools. Medium is the default reasoning effort, and the model does not support none or minimal, so migrations from lighter reasoning configurations must begin at low.

At launch, GPT-6.1 Sol is available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. It is not yet available in Chat. API users can select it immediately. OpenAI says an Ultrafast option with up to eight times faster generation will follow within days.

Stronger capabilities triggered the same safeguards as Astra

OpenAI classifies GPT-6.1 Sol as Critical for cybersecurity capability and High for biological and chemical capability, applying the same safeguards stack used for Astra. In an internal exploit evaluation covering recently disclosed vulnerabilities, 6.1 Sol achieved arbitrary code execution in 21.5% of cases, up from 5.5% for GPT-6 Sol but below Astra’s 31.5%.

In a deployment simulation built from 49,650 tasks in internal Codex traffic, 6.1 Sol produced 28 severity-three-or-higher flags, or 0.056%. Within this internal evaluation, GPT-6 Sol produced 42 and Astra produced 27.

The card also records a weaker area. In a low-stakes restrictions evaluation run without system-level controls designed to prevent circumvention, unwanted persistence appeared in 23.5% of 6.1 Sol rollouts versus 17.4% for Astra.

Astra for the hardest case, 6.1 Sol for repeated complex work

Terminal-Bench Science makes the boundary visible. Astra retained the top score at 68.1%. GPT-6.1 Sol cost an average of $5.47 per task, compared with $23.21 for Opus 5.5 and $23.80 for Astra. Astra remains the model for the hardest single run; 6.1 Sol is built for many demanding runs under a fixed budget.

OpenAI is trying to own the efficiency frontier as well as the crown

GPT-6.1 Sol does not win every benchmark. It does something commercially more important: it narrows the performance gap to Astra while preserving a dramatically lower operating cost. Competitors now have to beat not only OpenAI’s best model, but also a production model that approaches it at one-fifth of the headline token price.

Factuality moved in the same direction. On deliberately difficult conversations that had prompted users to flag an earlier factual error, the share of low-effort answers containing an error fell from 11.4% to 7.7%, a reduction of roughly 32%. Better default economics matter more when the lower-cost settings also become more dependable.

If OpenAI consolidates its lead with GPT-6.1 Sol, it will not be because Sol took Astra’s benchmark crown. It will be because near-Astra coding and agent performance became cheap enough to deploy as the default rather than reserve for exceptional jobs.

Frequently asked questions