GPT-5.6 Sol: The Agent-First Flagship (Full Benchmarks)
OpenAI's GPT-5.6 Sol is GA. #1 on the Coding Agent Index, 91.9 on Terminal-Bench, one point behind Fable 5 on intelligence — at a price that bites. Every number, and the catch.
On 9 July 2026 OpenAI moved GPT-5.6 to general availability and, with it, its flagship agentic mode: Sol. The GPT-5.6 family ships in three tiers — Sol (the heavy agentic model), Terra, and Luna — and Sol is the one every coding-agent team spent the week benchmarking. I pulled the numbers from the trackers and the model card so you can see exactly where it lands.
What Sol actually is
GPT-5.6 is OpenAI's mid-cycle bump on the GPT-5 line, and Sol is its top reasoning-and-tools tier — built for "difficult professional, coding, research, computer-use, and tool-heavy work." It carries a 1M-token context window, and it's the engine behind ChatGPT Work, OpenAI's new agent-based product that pairs Sol with Codex to actually do multi-step jobs rather than just answer.
It landed in limited preview on 26 June and went GA across ChatGPT, Codex, and the API on 9 July 2026. There's also an Ultra setting that trades latency and cost for a few more points on the hardest benchmarks.
The benchmarks
Sol's whole pitch is agentic and tool-use performance, and that's where the headline numbers are. These are the ones doing the talking:
A 94.6 on GPQA Diamond (PhD-level science) and a 91.9 on Terminal-Bench 2.0 — can the model drive a shell to finish real tasks — put Sol squarely at the frontier. On Artificial Analysis's Coding Agent Index it takes the #1 spot at 80 points, and on their overall Intelligence Index it sits at 59 — one single point behind Claude Fable 5, at roughly a third of Fable's price.
Agentic
| Benchmark | GPT-5.6 Sol |
|---|---|
| BrowseComp | 92.2 |
| Terminal-Bench 2.0 | 91.9 |
| CyberGym | 84.5 |
| OSWorld 2.0 | 62.6 |
| Toolathlon | 58.0 |
| ExploitGym | 33.7 |
Coding
| Benchmark | GPT-5.6 Sol |
|---|---|
| DeepSWE v1.1 | 72.7 |
| CursorBench 3.2 | 67.2 |
| SWE-bench Pro | 64.6 |
| Terminal-Bench 2.1 (Ultra) | 91.9 |
| Terminal-Bench 2.1 (standard) | 88.8 |
Knowledge, math & multimodal
| Benchmark | GPT-5.6 Sol |
|---|---|
| GPQA Diamond | 94.6 |
| FrontierMath v2 (Tiers 1–3) | 89.0 |
| FrontierMath v2 (Tier 4) | 83.0 |
| MMMU-Pro (with Python) | 84.6 |
| MMMU-Pro | 83.0 |
| HealthBench Professional | 60.5 |
Where it sits in the market
Sol's positioning is unusually clean: near-Fable-5 intelligence, best-in-class coding-agent scores, at $5 in / $30 out per 1M tokens. That output price is the one wrinkle — at $30/M out it's more expensive on output than Claude Opus 4.8 ($25) and far above the new wave of cheap frontier models. OpenAI is betting that if Sol finishes agentic tasks in fewer, better steps, the higher per-token price still wins on total job cost. For tool-heavy pipelines that's a real argument; for chatty, high-output workloads it's a harder sell.
The honest caveats
- Terminal-Bench "91.9" is the Ultra setting. Standard GPT-5.6 lands at 88.8 on Terminal-Bench 2.1. Both are excellent — just make sure you're comparing the tier you'll actually run.
- Output is pricey. $30/M out is the highest of the mainstream frontier tier right now. Budget by task completion cost, not sticker price.
- Benchmarks aren't your pipeline. A #1 Coding Agent Index is a strong signal, not a guarantee your agents will behave on your codebase. Run it against your own evals.
My take
Sol is the most convincing "agent-first" flagship yet — the Terminal-Bench and Coding Agent Index numbers are exactly what you want if you're building autonomous coding or computer-use pipelines. But 2026's real story isn't any one model; it's that the frontier got crowded in a single fortnight — Grok 4.5 on the 8th, Sol on the 9th, Kimi K3 on the 16th. I keep everything behind one interface and let models compete for my traffic on merit. Right now Sol earns the agentic slots; the cheaper models earn the high-volume ones. Wire for portability and you get to say that instead of being locked into whoever's on top this week.
Sources
- OpenAI — GPT-5.6
- BenchLM — GPT-5.6 Sol benchmarks
- Artificial Analysis — GPT-5.6 has landed
- ExplainX — GPT-5.6 release, features & benchmarks
- QCode — GPT-5.6 Sol, Terra & Luna guide
Numbers are as reported around GA (9 July 2026) and will move as independent boards settle. Spot a corrected figure? Tell me and I'll update.
