GPT-5.6 Performance Milestone: What Sol's Benchmark Scores Actually Mean

What is the GPT-5.6 performance milestone

The GPT-5.6 performance milestone refers to a set of new record benchmark scores that OpenAI's Sol model posted after GPT-5.6 reached general availability on July 9, 2026. The headline result is a 92.2% score on BrowseComp, a benchmark for finding hard-to-locate information on the web, and 62.6% on OSWorld 2.0, which measures how well a model can operate a real computer over long, multi-step tasks.

Both numbers beat the previous best scores from any model, including OpenAI's own GPT-5.5. What makes the milestone notable isn't just the top-line percentage. It's that Sol reached these scores while using far fewer output tokens than the models it surpassed, which changes how much a completed task actually costs.

Three tiers, three jobs

GPT-5.6 does not ship as a single model. OpenAI split it into three tiers, each aimed at a different workload:

• Sol, the flagship, built for complex reasoning, coding, and long agentic sessions
• Terra, a mid-tier model priced near half of Sol, competitive with it on several tasks
• Luna, the budget tier for high-volume requests where raw capability matters less than cost

This tiered approach mirrors a broader pattern across the industry, where labs increasingly ship a family of models instead of one, letting developers route requests based on task complexity rather than defaulting to the most expensive option every time.

The benchmark numbers

Three benchmarks anchor the GPT-5.6 performance milestone: BrowseComp, OSWorld 2.0, and Terminal-Bench 2.1. Each tests a different kind of agentic skill, from web research to operating a desktop to running commands in a terminal.

On OSWorld 2.0 specifically, Sol surpassed Claude Opus 4.8 while using 85% fewer output tokens to get there. That gap matters more than it looks. A model that scores slightly higher but burns several times the tokens per task can end up costing more per completed job, even at a lower headline price.

Why token efficiency is the real story

Benchmark percentages get the headlines, but they don't tell you what a task will cost. A model that completes a browsing task in 4,000 output tokens is cheaper to run than one that needs 20,000 tokens, even if the second model's list price per token is lower. This is the same tradeoff developers already navigate when comparing domain-specific language models against generic LLMs for narrower tasks.

Pricing published after launch reflects the tiering: Sol runs $5 per million input tokens and $30 per million output tokens, Terra is roughly $2 and $12, and Luna drops to $0.20 and $1.20. For teams running agentic workflows at scale, the combination of tier pricing and per-task token count is what actually shows up on the invoice, not the benchmark table.

What this means for developers choosing a model

If you're building tools that browse the web, operate a virtual desktop, or run long multi-step agent sessions, Sol's OSWorld 2.0 result is directly relevant, since that benchmark simulates exactly those tasks. Teams building simpler coding assistants or chat features may find Terra's price-to-performance ratio a better fit, especially given how it tracks close to Sol on several benchmarks at roughly half the cost.

Developers comparing AI coding tools day to day, such as those evaluating options in the Cursor versus Claude Code versus Copilot comparison, should expect GPT-5.6 to show up as a backend option across several of these products in the coming months, since IDE vendors typically add new frontier models within weeks of general availability.