GPT-5.6 Sol

// Overview

GPT-5.6 Sol is the flagship model in OpenAI's GPT-5.6 series, first announced on June 26, 2026 as a limited preview to around 20 vetted partners and publicly released across the API, Codex, and ChatGPT on July 9, 2026 alongside two smaller tiers, Terra and Luna. OpenAI describes Sol as its strongest model for complex reasoning, advanced coding, agent workflows, and demanding enterprise tasks.

Sol introduces a new max reasoning effort setting that gives the model more time to think, plus an ultra mode that delegates work to subagents to accelerate large, multi-step jobs. The model accepts text and image input, supports a 1,050,000-token context window with up to 128,000 output tokens, and carries a February 16, 2026 knowledge cutoff.

OpenAI positions Sol as the workhorse tier of the GPT-5.6 family — frontier intelligence for professional work — priced at $5.00 per 1M input tokens and $30.00 per 1M output tokens, the same headline rate as GPT-5.5.

On August 6, 2026 OpenAI shipped an updated Sol to ChatGPT for Plus and Pro users, adding a slider that lets the user choose how much effort ChatGPT spends on a response. OpenAI's system card distinguishes this snapshot as the 'August' version; the July version remains in use in Codex and ChatGPT Work. The August release reduces factual error rates by roughly 60% versus GPT-5.5 Instant across all three of OpenAI's hallucination prompt sets (factuality-heavy, user-flagged failures, and high-stakes medical, legal, and financial prompts), and improves on every HealthBench evaluation — most sharply on HealthBench Professional (+15.6 points) and HealthBench Hard (+8.5). Under OpenAI's Preparedness Framework the August release is treated as High capability in both Cybersecurity and Biological and Chemical domains.

On August 13, 2026 Cerebras announced Ultrafast, a service tier in the OpenAI API that serves GPT-5.6 Sol on Cerebras hardware at up to 750 output tokens per second — up to 14× faster than Standard processing — with, in Cerebras' words, 'the same intelligence as GPT-5.6 Sol Standard'. It opened in limited preview to a small group of OpenAI customers, with no published price. Cerebras reports the tier completing the full 2,500-question Humanity's Last Exam set in just over 11 hours and a 5.6× end-to-end speedup on GDP-Val with no loss in quality; the speed comes from its Wafer-Scale Engine keeping model weights in 44 GB of on-chip SRAM instead of off-chip memory. OpenAI's Sachin Katti framed the rollout as starting 'with a small group of customers to learn where that speed creates meaningful value'.

// Benchmarks

This model's scores

1. Terminal-Bench 2.188.8%
2. HealthBench (length-adjusted, August release)55
3. HealthBench Hard (length-adjusted, August release)31.4
4. HealthBench Consensus (length-adjusted, August release)95.5
5. HealthBench Professional (length-adjusted, August release)54

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Cached input reads receive a 90% discount; cache writes are billed at 1.25x the input rate, with a 30-minute minimum cache life.

Pricing source ↗

// Strengths

• Frontier reasoning with a new max reasoning effort tier
• Ultra mode that orchestrates subagents for long-horizon agentic work
• State-of-the-art Terminal-Bench 2.1 score for command-line coding workflows
• Stronger SecureBio biology and cybersecurity reasoning than GPT-5.5
• Improved factuality, with fewer reported hallucinations than the prior flagship
• August 2026 release cuts factual error rates by roughly 60% versus GPT-5.5 Instant on all three of OpenAI's hallucination prompt sets
• Best-in-family health performance: 54.0 length-adjusted on HealthBench Professional, up 15.6 points on GPT-5.5 Instant

// Best for

• Reach for it when an agentic coding task needs the most capable single model and the longest reasoning budget.
• Reach for it for multi-agent workflows where Sol's ultra mode can spin up subagents to parallelise work.
• Reach for it for high-stakes professional reasoning in scientific, biological, or cybersecurity domains.
• Reach for it when factuality matters and the prior GPT-5.5 generation produced too many hallucinated facts.

// How to access

// GPT (Flagship / Thinking) — every version

The full lineage of the GPT (Flagship / Thinking) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.