GPT-6.1 Sol matches Astra on DeepSWE at a fifth of the cost

A week after GPT-6 Sol reached the API, OpenAI replaced it on September 29 with GPT-6.1 Sol and kept the price at $2 per million input tokens and $10 per million output. The rate that moved is cached input. It is now $0.10 per million, half what GPT-6 Sol charged and 95% below the standard input price, according to OpenAI’s announcement.

OpenAI pitches the release as “near-Astra intelligence for a fifth of the price”. GPT-6 Astra, the flagship that launched on September 3, costs $10 in and $50 out, and its cached input runs $1 per million. So when an agent resends the same long context on every turn, the repeated part costs a tenth of what Astra charges for it. Every benchmark figure below comes from OpenAI’s own charts.

The model is live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, but not yet in Chat. In the API it is gpt-6.1-sol. OpenAI says a GPT-6.1 Sol Ultrafast variant arrives in the coming days, with token generation up to 8x faster than standard speed in Codex.

DeepSWE v1.1 runs complex software-engineering tasks in real codebases. On it, OpenAI says, GPT-6.1 Sol matches Astra at roughly one-fifth of the cost and beats GPT-6 Sol’s best score by 6.4 percentage points, at a lower reasoning effort and cost. In September, OpenAI’s chart put GPT-6 Sol at 68.8% at max effort. The News Stack reads the new model at about 75%.

GDP.pdf asks professional questions about complex PDFs, with tables, charts, diagrams and fine print, taken from finance, healthcare, legal and seven other fields. GPT-6.1 Sol scores higher than Anthropic’s Opus 5.5 with fallbacks at less than half the cost per task across the tested reasoning settings, and approaches Astra’s state-of-the-art score at roughly a fifth of the cost per task. The News Stack puts the scores at around 32% for Sol and about 29% for Opus 5.5.

AutomationBench 1.0.6 tests agents on end-to-end business workflows using 47 tools across sales, marketing, operations, support, finance and HR. At medium reasoning effort, GPT-6.1 Sol scores 2.2 points above Opus 5.5 at roughly a third of the cost, and 4.8 points above GPT-6 Sol at the same setting. Claude Fable 5.1 also appears on that chart, with a footnote saying its datapoint “understates its actual cost, as it omits the cost of fallbacks, which occurred on ~40% of tasks.”

OpenAI’s comparisons are mostly with Opus 5.5, which launched the same day as GPT-6 Sol. The News Stack instead put the new model next to Sonnet 5.5, which Anthropic released on Monday, September 28, at the same $2/$10. Few benchmarks cover both models, and the results it found were mixed. Sol leads on DeepSWE at about 75% to Sonnet’s 71%. On AutomationBench, Sonnet scores 44.7% to Sol’s roughly 36%. Sol’s only edge on that benchmark is the bill: $0.30 per task against $1.14 for Sonnet.

On computer use, OpenAI reports partial reward on the offline set of OSWorld 2.0, release v2026.08.08, which covers long computer-use workflows across everyday and professional tasks. At maximum reasoning effort, GPT-6.1 Sol beats GPT-6 Sol by seven points at less than half the cost, and comes within 2.1 points of Astra at roughly one-seventh of the cost per task. With results like these, The News Stack wrote, “it’ll be hard to justify using Astra for most use cases.”

Factuality was measured on de-identified ChatGPT conversations in which users had flagged a factual error from an earlier model. OpenAI calls these prompts deliberately difficult and says they don’t represent typical usage, where errors are rarer. The biggest gain comes at low reasoning effort: the share of responses with at least one factual error falls from 11.4% with GPT-6 Sol to 7.7%, a cut of about 32%. Across the tested settings, the new model’s error rate stays within 1.9 points of Astra’s at less than a fifth of the cost per task.

OpenAI also says alignment improved, bringing Sol closer to Astra. In tests built around hard situations, GPT-6.1 Sol fails less often than GPT-6 Sol at three things: saying when a search tool is broken, respecting explicit restrictions, and avoiding unauthorized outcomes during agentic tasks. The News Stack gives the search-tool figure: the model fails to disclose a broken tool 2.8% of the time, rather than guessing. OpenAI observed no attempts to get around an automated safety reviewer, the same result as Astra and GPT-6 Sol. The rest is in the GPT-6.1 Sol system card addendum.

Terminal-Bench Science 0.1 covers scientific workflows such as data analysis, simulation and theorem proving. There GPT-6.1 Sol more than doubles GPT-6 Sol’s score at maximum effort, at less than half the cost per task. At max effort it averages $5.47 per task, against $23.21 for Opus 5.5 and $23.80 for Astra. Astra still has the top score among the models tested, 68.1%, and OpenAI says it “should be used for the most difficult scientific research tasks.”