GPT-6 Sol is OpenAI's mid-tier GPT-6 model for coding and agent work — 50% cheaper than GPT-5.6 Sol, built for high-volume iteration.
What Changed
Released September 22, 2026, GPT-6 Sol sits one step below OpenAI's flagship GPT-6 Astra: same training approach and much of the same alignment work, packed into a faster, cheaper model meant for everyday coding and agent work rather than the hardest, most demanding tasks. The pitch is capability-per-dollar rather than a new capability ceiling — Astra stays OpenAI's top pick when you want the best possible result regardless of cost; Sol is for the work you run often enough that cost and speed start to matter as much as raw intelligence.
What It's Actually Good At
- On AutomationBench (business workflows across 47 tools), Sol at its highest reasoning setting scores 33.2%, ahead of Claude Opus 5 at max effort (26.9%) — and OpenAI says it does so at about 9% of Opus 5's cost per task.
- On Agents' Last Exam (long, economically-valuable tasks across 55 industries), Sol scores 56.4% at max effort, above Claude Opus 5's best score at roughly 60% lower cost per task.
- On DeepSWE (real-codebase software engineering tasks), Sol scores 68.8% at max effort — within 1.1 points of Claude Fable 5's top score, at about 80% lower cost.
- On OSWorld 2.0 (computer-use tasks), Sol scores 60.5%, essentially matching Claude Opus 5's 60.3%, again at roughly 80% lower cost per task.
These are all OpenAI's own benchmark numbers, run with OpenAI's chosen settings and cost assumptions — a genuinely useful signal, but a vendor's comparison of itself to competitors, not an independently verified result.
A Concrete Example of the Change
OpenAI's own release notes include a side-by-side: asked to restyle a website into a bento-box layout with a sliding page switcher, the older GPT-5.6 Sol wrote a longer reply that explained its plan in vague terms ("get the bento feel," "reshaping... around that system") before confirming what it built. GPT-6 Sol's reply was shorter, stated what it checked (desktop and narrow mobile, including back-button navigation) and what it decided (React wasn't needed) without restating obvious context back to the user. OpenAI frames this as the general shift in Sol and Luna's writing style: less jargon, fewer hedges, more direct confirmation of what was actually done.
Pricing and Access
GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens — half of GPT-5.6 Sol's promotional pricing. Cached input reads (repeated context) cost $0.20 per million, a 90% discount. It has a 1.05-million-token context window and supports up to 128,000 tokens of output per response. It's rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts, and is available via the API as gpt-6-sol. It is not available to free ChatGPT accounts or in regular ChatGPT chat yet — Codex, Work, and the API are where it lives today.