Flagship of the GPT-5.6 family: agentic coding and near-frontier reasoning
Sol is the top tier of three; it matches GPT-5.5 on short-context pricing.
Overview
GPT-5.6 Sol is the flagship model of OpenAI's GPT-5.6 family, announced June 26, 2026 as a limited preview and opened to broad API access on July 9, 2026. It is the top tier of a three-model lineup — Sol, Terra, and Luna — aimed at complex reasoning, coding, and agentic workflows, sitting above GPT-5.5 in capability while matching its short-context pricing.
Sol keeps the same architecture-level shape as GPT-5.5 — a 1.05M-token context window, image input, and configurable reasoning effort — but OpenAI reports it reaching new highs on command-line and agentic coding benchmarks while using fewer output tokens per task than several larger competing models.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Input modalities | Text, image |
| Output modalities | Text |
| Reasoning effort | none, low, medium, high, xhigh, max (default: medium) |
| Knowledge cutoff | February 2026 |
Inputs above 272K tokens enter a higher long-context tier (2x input, 1.5x output). Cached input reads get the standard ~90% discount. See live pricing in the model catalog.
Benchmarks
Intelligence and cost efficiency
Near-frontier intelligence at a fraction of the cost per task
Source: Artificial Analysis Intelligence Index (max reasoning effort).
On Artificial Analysis's Intelligence Index (max reasoning effort), Sol scores 59 — one point behind Claude Fable 5's 60 — while costing roughly one-third as much per task (about $1.04 versus Fable 5's per-task cost). Sol also uses fewer output tokens per task (about 15K) than Claude Opus 4.8, GLM-5.2, and Gemini 3.5 Flash despite matching or beating their intelligence scores, a sign of more efficient reasoning rather than longer chains of thought.
Coding and terminal work
Top score on command-line and agentic coding evaluations
Source: OpenAI and Artificial Analysis published GPT-5.6 results.
Sol leads Terminal-Bench 2.1 at 91.9% in ultra mode (88.8% base), ahead of Claude Mythos 5 and GPT-5.5. On Artificial Analysis's Coding Agent Index inside Codex, Sol scores 80 — the top score among evaluated models — while running about 40% cheaper than Claude Fable 5 and about 10% cheaper than Claude Opus 4.8 in comparable harnesses.
Office and knowledge work
On Artificial Analysis's AA-Briefcase office-task suite, Sol's Presentation Elo ranks second only to Claude Fable 5, though its rubric score (42%) still trails Fable 5's 56% — worth knowing if your workload leans toward long-form document and deck generation rather than coding.
When to use it
- Agentic coding and terminal workflows: multi-step command-line tasks, repository-scale changes, and CI-style iteration loops.
- Cost-sensitive frontier reasoning: tasks that need near-frontier intelligence without the token or dollar cost of the very largest models.
- Long-context engineering: large repos, logs, and design docs inside a single 1.05M-token window.
- General agentic tool use: multi-tool workflows where reasoning effort can be tuned per request.
CrossModel exposes GPT-5.6 Sol through an OpenAI-compatible /v1/chat/completions API. Current pricing is available in the model catalog.