OpenAI Sol Model: GPT-5.6 Sol Explained — Benchmarks, Pricing & API Guide
GPT-5.6 Sol is the flagship tier of the OpenAI GPT-5.6 model family, engineered for the most demanding professional tasks — from long-horizon coding agents to cybersecurity research and biology workflows. Built by OpenAI and released publicly on July 9, 2026 as part of the chat gpt 5.6 generation, it scores 88.8% on the key agentic benchmark Terminal-Bench 2.1 — and its compute-intensive Sol Ultra configuration pushes that to 91.9%.
What Is the OpenAI Sol Model?
Sol is OpenAI’s most capable frontier model for complex professional work. Its API model ID is gpt-5.6-sol, with the alias gpt-5.6 routing automatically to it. The context window is 1,050,000 tokens with a maximum output of 128,000 tokens; the knowledge cutoff is February 16, 2026. It accepts text and image input, outputs text only. OpenAI shipped it in two stages — a limited preview on June 26, 2026 restricted to vetted partners at the White House’s request over safety concerns, then general availability on July 9 after safeguards were confirmed. OpenAI stated such access restrictions “should not become the norm.” The GPT-5.6 naming replaced legacy suffixes (Turbo, Preview, Mini) with a three-tier system: the number identifies the generation, the suffix the capability tier. Sol is the “sun” — the highest-intensity tier, for work where correctness matters more than cost per token.
Sol vs Terra vs Luna: Which Tier to Choose
| Tier | Input (per 1M) | Output (per 1M) | Best For |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | Complex agents, coding, cybersecurity, biology |
| GPT-5.6 Terra | $2.50 | $15.00 | Bounded implementation, first-pass review |
| GPT-5.6 Luna | $1.00 | $6.00 | High-volume chatbots, classification, summaries |
Sol at $5/$30 per 1M tokens delivers the most capability at the same price point as GPT-5.5. Terra delivers roughly GPT-5.5-level quality at about half the cost (see the GPT-5.6 Terra comparison); Luna is the fastest, cheapest tier. The right tier depends on what happens when the model fails: use Sol for correctness-critical, multi-file, multi-step jobs; Terra for scoped implementation and first-pass review; Luna for high-volume preprocessing and labeling. Sol Ultra allocates more compute per request — 91.9% versus 88.8% on Terminal-Bench 2.1 — at the same $5/$30 rate (more tokens, not a higher rate), accessed via reasoning.effort: "max" in the Responses API.
Benchmarks vs Competitors
| Benchmark | GPT-5.6 Sol | Sol Ultra | GPT-5.5 | Claude Fable 5 |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.8% | 91.9% | 88.0% | 84.3% |
| DeepSWE v1.1 | 72.7% | — | 67.0% | 69.7% |
| AA Coding Agent v1.1 | 80.0 | — | 76.4 | — |
| BrowseComp | 90.4% | 92.2% | 84.4% | 84.3% |
| OSWorld 2.0 | 62.6% | — | 47.5% | 54.8% |
| Agents’ Last Exam | 52.7% | — | 46.9% | 40.5% |
| SWE-Bench Pro | 64.6% | — | — | 80.0% |
| AA Intelligence v4.1 | 58.9 | — | — | 59.9 |
Sol’s profile is uneven in ways that matter for routing: it dominates agentic and computer-use tasks but trails Claude Fable 5 on pure repository coding. On the Artificial Analysis Coding Agent Index v1.1, Sol scored 80/100 — 2.8 points above Fable 5, a new state-of-the-art at lower cost and faster speed. OSWorld 2.0 (computer use) is the most striking jump: Sol at 62.6% is 15.1 points above GPT-5.5’s 47.5%. Sol also leads on BrowseComp deep web research (90.4%, Ultra 92.2%) and Agents’ Last Exam (52.7%). The clearest counterpoint is SWE-Bench Pro, where Fable 5’s 80.0% beats Sol’s 64.6% by 15.4 points — concentrated in open-ended architectural judgment and cold-start code generation, not multi-step agentic persistence.
“We want the GPT-5.6 family to give developers a stable, durable vocabulary for routing — not a model you configure with sliders, but a tier you choose based on task depth.”
OpenAI, GPT-5.6 Launch Post, July 2026
The practical takeaway: test both Sol and Fable 5 on actual repository tasks before committing — the 15-point SWE-Bench Pro gap does not hold uniformly across real-world codebases and languages.
API Pricing: Complete Rate Card
Pricing has two regimes based on context length. Crossing 272,000 input tokens reprices the entire request at long-context rates — not just the tokens above the threshold — so it must be modeled explicitly.
| Rate Type | Input | Cached Input | Cache Write | Output |
|---|---|---|---|---|
| Standard (≤272K) | $5.00/M | $0.50/M | $6.25/M | $30.00/M |
| Long Context (>272K) | $10.00/M | $1.00/M | $12.50/M | $45.00/M |
| Batch/Flex | $2.50/M | $0.25/M | $3.13/M | $15.00/M |
| Priority | $10.00/M | $1.00/M | $12.50/M | $60.00/M |
A standard 200K-input / 20K-output request runs about $1.60; a long-context 500K / 50K request runs $7.25, with long-context rates applying to the whole request. Prompt caching (30-minute minimum, explicit breakpoints) drops cache reads to $0.50/M — 90% cheaper than uncached input — but cache writes cost $6.25/M (1.25× uncached), so repeatedly writing context that is never reused increases cost. Measure cache hit rate directly, not just whether caching is enabled.
How to Access GPT-5.6 Sol
Sol reached general availability July 9, 2026 across the OpenAI API and ChatGPT. Via the API, use model ID gpt-5.6-sol (or the gpt-5.6 alias) on Chat Completions or the Responses API; set reasoning effort (medium, high, or max for Sol Ultra). Batch pricing applies at $2.50/$15 per 1M. In ChatGPT, Free and Go have no Sol access; Plus offers Medium and High reasoning; Pro, Business, and Enterprise add Extra High and Pro. Sol also powers the ChatGPT Work context on Plus and above — see the ChatGPT Work overview.
Programmatic Tool Calling and Coding Persistence
Programmatic Tool Calling is Sol’s most significant new API capability: instead of calling one tool at a time and routing each result back through context, Sol writes a JavaScript orchestration script in a hosted runtime that coordinates multiple tool calls in a single round trip. One launch customer running the same agent on the same tools consumed 63.5% fewer tokens — from ~150,000 tokens per question to under 55,000 — making previously cost-prohibitive pipelines viable. It is best for filtering, joining, ranking, and aggregation over structured tool results; direct tool calls stay clearer where each result needs fresh semantic interpretation. Sol also adds persisted reasoning across turns, explicit cache breakpoints, and a multi-agent beta with parallel subagents.
Sol’s operational differentiator in coding is persistence — following through on step 8 of a 10-step task where most models stop, running verification and catching unrelated warnings before submitting rather than stopping at one passing happy-path test. On a 100-task run across TypeScript, Go, Python, JavaScript, and Rust, Sol hit a 63.7% pass rate at ~20,968 output tokens/task versus Terra’s 40.7% at 55,594 tokens/task. In code review it passed 69.7% of actionable cases but generated more noise (31.6% actionable precision), so production review with Sol needs a filtering layer.
Cybersecurity and Biology
Per OpenAI’s system card, Sol assists effectively with threat modeling, code review, patching, and blue teaming, but cannot carry out autonomous cyberattacks; OpenAI classified both Sol and Terra below the “Cyber Critical” threshold. Real-time cyber and biology misuse classifiers may block requests near capability boundaries, including some legitimate work — validate refusal behavior before deploying. Sol shows meaningful gains on SecureBio biology benchmarks over GPT-5.5 (exact scores unpublished) and is positioned as a research assistant, not a standalone autonomous tool.