GPT-6.1 Sol Posts 75.2% on DeepSWE and 71.4% on OSWorld — Near-Astra Performance at One-Seventh the Cost
Benchmark data released by OpenAI alongside the GPT-6.1 Sol DevDay announcement quantifies what “near-Astra” actually means in numbers, and the cost-performance ratio is significant.
DeepSWE v1.1
On DeepSWE v1.1, a long-horizon software engineering benchmark, GPT-6.1 Sol achieves 75.2% at high reasoning effort. GPT-6 Sol’s best score on the same benchmark was 68.8% at maximum effort — a 6.4 percentage point gain. OpenAI’s figure of “approximately 76% lower cost per task” refers to the comparison between GPT-6.1 Sol at high effort and GPT-6 Sol at maximum effort.
For context, Gemini 4 Argon, announced the following day by Google, posts 77.9% on DeepSWE v1.1 — 2.7 points higher than Sol at its standard operating mode. GPT-6 Astra was included in OpenAI’s benchmark comparison table, scoring 74.1% to 74.8% on DeepSWE v1.1 — placing between Sol’s 75.2% and GPT-6 Sol’s 68.8%.
OSWorld 2.0
On OSWorld 2.0’s offline evaluation set, GPT-6.1 Sol scores 71.4% at maximum reasoning effort. GPT-6 Astra scores 73.5% on the same set at maximum reasoning effort. The 2.1-point gap translates to Astra costing roughly seven times more per task than Sol, based on published API pricing ($10/M vs $2/M input tokens; $50/M vs $10/M output tokens).
AutomationBench
On AutomationBench, Zapier’s benchmark for end-to-end execution across core business functions, GPT-6.1 Sol scores 31.7% at medium reasoning effort. This is 4.8 percentage points above GPT-6 Sol at the same setting. Gemini 4 Argon posts 51.3% on the same benchmark — 19.6 points higher — per Google’s own launch materials.
The Cost Arithmetic
GPT-6.1 Sol launched at $2 per million input tokens and $10 per million output tokens, unchanged from GPT-6 Sol’s pricing. GPT-6 Astra is priced at $10/M input and $50/M output. At the standard 25%/75% input-output token mix, Astra’s blended cost per million tokens is $40, versus Sol’s $8 — a 5x multiple. On OSWorld specifically, OpenAI’s calculation puts the per-task cost ratio at roughly 7x, reflecting that OSWorld tasks involve higher output token volumes.
For workloads where the 2-point gap between Sol and Astra on OSWorld is not material — the majority of production computer-use deployments — Sol is the economically rational choice. Where quality maximisation matters — complex enterprise workflows, one-shot high-stakes agent runs — Astra’s margin remains real.
Pricing Structure
GPT-6.1 Sol keeps the same $2/$10 input/output pricing as GPT-6 Sol, but cuts the cached input token price from $0.20 per million to $0.10 per million — a meaningful reduction for long-context applications that re-read system prompts and document bases across many turns.
The model is available through ChatGPT Work and Codex. It is not yet accessible through the standard ChatGPT interface.