OpenAI announced GPT‑6.1 Sol, an upgraded variant of GPT‑6 Sol intended to offer a new balance between capability and cost for professional, agentic tasks. The company positions the model between GPT‑6 Sol and GPT‑6 Astra: on many complex evaluations it approaches Astra’s performance while charging substantially lower token prices.
Pricing and availability
According to the announcement, GPT‑6.1 Sol’s standard API prices are: $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. The cached input price is highlighted as being 95% cheaper than standard input pricing and 50% cheaper than GPT‑6 Sol’s cached input pricing, which can lower costs when agents reuse context across requests.
GPT‑6.1 Sol is available immediately to Plus, Pro, Business, Enterprise, and Edu customers in ChatGPT Work and Codex; it is not yet available in Chat. Developers can access it through the OpenAI API as gpt-6.1-sol. OpenAI also plans to offer a GPT‑6.1 Sol Ultrafast variant with up to 8x faster token generation in Codex in the coming days.
Benchmark results
OpenAI published comparisons across several benchmarks showing cost‑performance improvements for GPT‑6.1 Sol:
-
DeepSWE v1.1 (complex software‑engineering tasks in real codebases): GPT‑6.1 Sol matches GPT‑6 Astra at roughly one‑fifth of the cost and exceeds GPT‑6 Sol’s best score by 6.4 percentage points while requiring lower reasoning effort and cost.
-
GDP.pdf (answering professional questions from complex PDFs with tables, charts, diagrams and fine print): GPT‑6.1 Sol scores higher than Opus 5.5 with fallbacks at less than half the cost per task across tested reasoning settings, and approaches GPT‑6 Astra’s performance at roughly one‑fifth the cost.
-
AutomationBench (multi‑step business workflows): GPT‑6.1 Sol scores 2.2 percentage points above Opus 5.5 at medium reasoning effort at about one‑third the cost, and is 4.8 percentage points higher than GPT‑6 Sol at the same setting.
-
OSWorld 2.0 (computer‑use workflows): On the offline set, GPT‑6.1 Sol outperforms GPT‑6 Sol by seven percentage points at maximum reasoning effort at less than half the cost, and comes within 2.1 percentage points of GPT‑6 Astra’s score at maximum effort while costing roughly one‑seventh per task.
-
Terminal‑Bench Science 0.1 (scientific workflows using code and terminal tools): GPT‑6.1 Sol more than doubles GPT‑6 Sol’s score at maximum reasoning effort while costing less than half per task. At maximum effort, OpenAI reports an average cost of $5.47 per task for GPT‑6.1 Sol, compared with $23.21 for Opus 5.5 and $23.80 for Astra — a roughly 75% lower cost than those models on this benchmark. OpenAI notes that GPT‑6 Astra still achieved the highest score among tested models (68.1%) and should be used for the most demanding scientific research tasks.
Factuality and alignment improvements
OpenAI reports factuality gains for GPT‑6.1 Sol on difficult, deliberately error‑inducing prompts. At low reasoning effort, the share of responses containing at least one factual error fell from 11.4% (GPT‑6 Sol) to 7.7% (GPT‑6.1 Sol), a reduction of about 32%. Across tested reasoning settings, GPT‑6.1 Sol’s error rate stays within 1.9 percentage points of GPT‑6 Astra’s, at less than one‑fifth the cost per task.
In alignment evaluations, GPT‑6.1 Sol is said to be more transparent about limitations and more reliable at respecting user intent and safety constraints than GPT‑6 Sol. The company reports lower failure rates on tests of transparency (e.g., indicating broken search tools), respecting explicit restrictions, and avoiding unauthorized outcomes in agentic tasks. No attempts to bypass an automated safety reviewer were observed, matching GPT‑6 Astra and GPT‑6 Sol. Full details are available in the GPT‑6.1 Sol system card addendum.
OpenAI emphasizes these evaluations deliberately target challenging scenarios and do not reflect failure rates in typical use.
Methodology notes and caveats
OpenAI ran evaluations in its research environment or via its API, which it says can produce slightly different outputs from production ChatGPT due to differences in system prompts, available tools, and other factors. Competitor model results were taken from publicly available reports. The announcement notes instances where reported costs understate actual cost (for example, Claude Fable 5.1’s datapoint omitted fallback costs, and fallbacks occurred in roughly 40% of tasks).
Summary
GPT‑6.1 Sol is positioned as a cost‑efficient option for many professional and agentic use cases, offering significant improvements over GPT‑6 Sol and approaching GPT‑6 Astra performance in several benchmarks at a fraction of the cost. It may be particularly attractive where cached input pricing and multi‑request context reuse reduce operational expense; for the most demanding scientific research tasks, OpenAI still recommends GPT‑6 Astra.
