GPT-5.6 Sol Benchmarks: Terminal-Bench and BrowseComp SOTA at GA

On July 9, 2026, OpenAI launched the GPT-5.6 family for general availability, following its June limited preview: the flagship Sol, Terra for balanced everyday work, and Luna, the fastest and most affordable tier. GPT-5.6 is available now across ChatGPT, Codex, and the OpenAI API. With GA, OpenAI published a much broader eval suite than the preview — so this post replaces our preview coverage with every headline number Sol now posts, plus the recurring theme OpenAI leans on hardest: more work per token, and per dollar.

You can see GPT-5.6 Sol slotted into the live benchmark comparison alongside Claude Fable 5, Opus 4.8, and Sonnet 5. For a refresher on how these evals work, start with the complete guide to LLM benchmarks.

A tiered lineup: Sol, Terra, and Luna

GPT-5.6 keeps the naming system worth understanding. The number (5.6) marks the generation, while Sol, Terra, and Luna are durable capability tiers that advance on their own cadence — roughly flagship, balanced, and fast. OpenAI positions Terra as competitive with GPT-5.5 at a lower cost, and Luna as strong capability at the lowest cost, often outperforming older flagships at a fraction of the token count. Our comparison tracks the flagship, GPT-5.6 Sol.

Two reasoning modes: max and ultra

GPT-5.6 runs efficiently by default but can push harder on demand. max gives the model more time than xhigh to reason, run checks, and revise within a single agent. ultra goes further, coordinating four agents in parallel by default (16 in some configurations) to trade higher token use for stronger results and faster time-to-result. These modes explain the split scores below: the standard Sol number and, where OpenAI reported it, the higher ultra ceiling.

Coding: Terminal-Bench and a new SOTA coding index

Coding is where GPT-5.6 Sol makes its loudest case. On the Artificial Analysis Coding Agent Index, Sol sets a new state of the art at 80 (2.8 points above Fable 5) while using less than half the output tokens and taking less than half the time. The gains continue on Terminal-Bench 2.1, where Sol posts 88.8% and reaches 91.9% with ultra mode — the top of our table, ahead of Claude Fable 5 (88.0%), Opus 4.8 (82.7%), and Sonnet 5 (80.4%).

The GA suite also fills in numbers the preview withheld. On SWE-bench Pro, Sol scores 64.6% (with Terra at 63.4% and Luna at 62.7% close behind), and it sets fresh records on DeepSWE v1.1 (72.7%). Terminal-Bench was already GPT-5.5's strongest category (see the GPT-5.5 benchmark breakdown), so a generational jump here lands squarely in OpenAI's wheelhouse.

Agentic browsing, computer use, and tool use

Sol sets a new state of the art on BrowseComp, the agentic web-browsing eval: 90.4% single-agent, rising to 92.2% with ultra — the top of our table. On OSWorld 2.0 it reaches 62.6%, surpassing Opus 4.8 while using 85% fewer output tokens (note this is a different, harder revision than the OSWorld-Verified row in our table, so the numbers are not directly comparable). On AutomationBench, Sol tops the row at 18.1%, and it makes a large jump on BenchCAD design tasks (70.6%, versus GPT-5.5's 44.4%).

Cybersecurity: a real step change

OpenAI frames GPT-5.6 Sol as its most capable cybersecurity model yet, with a focus on defensive work — vulnerability research, patching, and code review. On ExploitBench, Sol scores 73.5% versus GPT-5.5's 47.9% at a comparable token budget. On ExploitGym, which asks agents to turn real vulnerabilities into working exploits, it nearly doubles GPT-5.5's peak (24.9% under a two-hour cap, 33.7% with six hours), and on SEC-Bench Pro it scores 71.2% versus 45.8%. It also tops CyberGym at 84.5% and hits 96.7% on capture-the-flag challenges.

Critically, OpenAI says the GPT-5.6 models do not cross the Critical threshold in cybersecurity or biology. GA ships with OpenAI's most robust safeguards to date — layered protections, a reasoning monitor, and Trusted Access gating for the most sensitive capabilities — validated with roughly 700,000 A100-equivalent GPU hours of automated red-teaming. Sol's cyber safeguards now block about ten times more potentially harmful activity than previous models.

Science, health, and academics

On GPQA Diamond, Sol scores 94.6%, tied for the top of our table, and it reaches 86% on FrontierMath Tier 1–3. On HealthBench Professional, the physician-graded tier, Sol scores 60.5% — up roughly 11 points over GPT-5.5 and within reach of Fable 5 (measured with a different scoring method). Life-sciences gains show up on GeneBench Pro (28.7%, more than double GPT-5.5) and LifeSciBench (59.9%).

The through-line: performance per dollar

Across nearly every category, OpenAI pairs each score with an efficiency claim: comparable or better results using fewer output tokens, less time, and lower estimated cost. Sol reaches its coding and browsing records while roughly halving output tokens and latency, and the smaller tiers extend the story — Terra and Luna outperform prior flagships at a fraction of the cost. GPT-5.6 also introduces more predictable prompt caching, with explicit cache breakpoints and a 30-minute minimum cache life.

Pricing and availability

GPT-5.6 is priced per million tokens across the three tiers: Sol at $5 input / $30 output (matching GPT-5.5), Terra at $2.50 / $15, and Luna at $1 / $6. Cache writes are billed at 1.25x the uncached input rate, while cache reads keep the 90% cached-input discount. In the Responses API, Programmatic Tool Calling lets Sol run in-memory programs that coordinate tools (Zero Data Retention compatible), and a multi-agent beta powers ultra-style parallel workstreams.

GPT-5.6 is rolling out now across ChatGPT, Codex, and the OpenAI API, reaching full availability over the following day. For head-to-head context, see Claude Opus 4.8 vs GPT-5.6 Sol and Claude Fable 5 vs GPT-5.6 Sol.

Key takeaways

Now generally available: Sol, Terra, and Luna ship across ChatGPT, Codex, and the API, with a much broader eval suite than the June preview.
Coding SOTA: 80 on the AA Coding Agent Index, 88.8% (91.9% ultra) on Terminal-Bench 2.1, and 64.6% on SWE-bench Pro — mostly at far fewer tokens.
Browsing SOTA: 90.4% on BrowseComp, rising to 92.2% with ultra, plus a jump on OSWorld 2.0 and BenchCAD.
Cybersecurity leap: 73.5% on ExploitBench (vs 47.9% for GPT-5.5) and 84.5% on CyberGym, without crossing OpenAI's Critical threshold.
Efficiency is the pitch: comparable or better results at fewer tokens, less time, and lower estimated cost across the family.
Explore the full profile on the GPT-5.6 Sol hub page or browse every score in the live benchmark comparison table.