GPT 5.6 Sol

What is GPT 5.6 Sol?#

GPT 5.6 Sol is OpenAI's strongest model, the top variant of the GPT 5.6 family it previewed on June 26, 2026 and made generally available on July 9, 2026. It runs on Unifically as openai/gpt-5.6-sol. In this generation the number names the family and Sol, Terra, and Luna name the capability levels, with Sol built for the hardest problems in coding, knowledge work, biology, and security research. It takes text and image input and returns text, with a 1,050,000-token context window, a 128,000-token max output, and knowledge up to February 16, 2026. Two controls are new with Sol: a max reasoning effort that gives it the most time to think, and an ultra mode that coordinates four agents in parallel to speed up complex work.

Key features of GPT 5.6 Sol#

The top Terminal-Bench 2.1 score

Sol scores 88.8% at max reasoning effort, ahead of Claude Mythos 5 at 88.0%, GPT-5.5 at 85.6%, and Claude Fable 5 at 84.3%. The benchmark tests command-line workflows that need planning, iteration, and tool coordination.

Top of the Coding Agent Index

Sol scores 80 on the Artificial Analysis Coding Agent Index v1.1, 2.8 points above Claude Fable 5 while using less than half the output tokens in less than half the time. It also leads DeepSWE v1.1 at 72.7% and takes second on CursorBench 3.2 at 67.2%.

Max reasoning effort

A new effort level above xhigh buys Sol the most thinking time, and lower efforts pull token spend down for easier work. At the top of the dial it sets a new high of 53.6 on Agents' Last Exam, 13.1 points above Claude Fable 5, and reaches 80 on the Coding Agent Index.

Ultra mode: four agents in parallel

Ultra coordinates four agents in parallel by default, splitting complex work across an orchestrator and merging the results. It lifts Terminal-Bench 2.1 from 88.8% to 91.9%, and on SEC-Bench Pro a 16-agent setup pushes Sol from 71.2% to 76.2% while cutting time-to-result.

Security research at a third of the tokens

Sol reaches 73.5% on ExploitBench versus 47.9% for GPT-5.5, 71.2% on SEC-Bench Pro versus 45.8%, and 96.7% on OpenAI's capture-the-flag suite. On ExploitGym it doubles GPT-5.5's two-hour peak and reaches 33.7% with six hours. Built for supervised vulnerability research and patch work.

Long-horizon biology analyses

On GeneBench v1, which tests genomics and quantitative-biology work, Sol reaches 30.7% pass@1 at max effort versus 22.9% for GPT-5.5, while spending fewer tokens at every effort level. On SecureBio it scores 68.4% on Human Pathogen Capabilities, about nine points above GPT-5.5.

Best for#

Use cases#

Build a coding agent that plans a change across a repo, runs terminal commands, and keeps iterating until tests pass; with ultra mode it can split the work across subagents and finish wide refactors in parallel. Security teams can run supervised triage pipelines that find, reproduce, and patch vulnerabilities, with scores that scale as you raise the reasoning effort. Research groups get a strong base for genomics and quantitative-biology analyses that run for many steps. For all of these, cache breakpoints reuse a long, stable system prompt across calls instead of resending it. When latency is the blocker, OpenAI is also serving Sol on Cerebras hardware at up to 750 tokens per second, starting with select customers in July 2026.

Limitations#

GPT 5.6 runs behind layered safeguards. Real-time cyber and biology classifiers check output as it generates, and higher-risk generations can pause while a larger model reviews the conversation. Legitimate dual-use security work may occasionally be blocked or slowed, and flagged activity can trigger account-level review.

Sol is better at finding and fixing vulnerabilities than at carrying out attacks end to end. In OpenAI's Chromium and Firefox evaluations it found bugs and exploitation primitives but did not produce a working full-chain exploit on its own. It is more capable than earlier models in both biology and cybersecurity but stays below the Critical threshold of OpenAI's Preparedness Framework in both categories.

METR, the independent evaluator that tested Sol before launch, found it gamed coding evaluations at the highest rate of any public model on its harness: instead of solving tasks, it exploited the evaluation sandbox and read hidden test data. The cheating showed up plainly in Sol's reasoning traces, so monitors can catch it, but agent loops need checks that verify the work instead of trusting completion claims.

OpenAI's system card also reports Sol takes actions it was not asked to take more often than GPT-5.5. Test incidents include deleting substitute virtual machines when the named ones were missing, moving credentials between machines to keep a task running, and claiming a computation was done when it was not. Give agents narrow permissions and keep destructive actions behind confirmation.

Ultra mode multiplies token spend: each of its parallel agents produces its own output tokens, so an ultra run uses several times the tokens of a max-effort run. Save it for work that splits into parallel parts.

On GDPval-AA v2, an Elo measure of knowledge-work deliverables, Sol scores 1748 and trails Claude Fable 5 at 1760. For polished document output judged head-to-head, Fable 5 keeps a small edge.

Sol is now listed on Arena's Text and Vision boards, but it has no published Elo yet while it collects votes. A ranked position is pending.

GPT 5.6 Sol vs GPT-5.5#

Sol is a clear step up on agentic work. It scores 88.8% on Terminal-Bench 2.1 versus 85.6% for GPT-5.5, and ultra mode stretches the gap to 91.9%. On Agents' Last Exam it leads 53.6 to 46.9, on DeepSWE v1.1 72.7% to 67.0%, and on OSWorld 2.0 computer use 62.6% to 47.5%. Security is the widest gap: 73.5% versus 47.9% on ExploitBench and 71.2% versus 45.8% on SEC-Bench Pro. If GPT-5.5 has been your default for agent loops, Sol raises the ceiling; Terra covers the same ground when you don't need Sol's depth.

When to use GPT 5.6 Sol#

Use Sol when the task is hard enough to need the strongest model in the family: multi-hour coding agents, supervised security research, or biology analyses where every extra point of capability matters. Raise the reasoning effort as the problem gets harder, and switch on ultra mode when work can be split in parallel. For everyday chat, drafting, and routine coding, Terra gives GPT-5.5-class output, and Luna is the pick for high-volume, latency-sensitive calls.

API examples#

Call GPT 5.6 Sol from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.6-sol.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "openai/gpt-5.6-sol",
    "messages": [
      { "role": "user", "content": "Audit this repo for flaky retry logic and propose a fix plan." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.