Morning Edition · Friday, July 10, 2026Published at 1:31 AM EDT · New York

OpenAI Ships GPT-5.6 Family to the Public and Claims the Top Coding Benchmark

Sol Ultra scores 91.9 percent on Terminal-Bench 2.1 by spawning subagents, but its margin over Anthropic's Mythos 5 is smaller than the variation between two runs of the same model.

Save

OpenAI Ships GPT-5.6 Family to the Public and Claims the Top Coding Benchmark

OpenAI released GPT-5.6 as a three-model family: Sol at the top, Terra in the middle, and Luna as the low-cost tier. The central claim is about agentic coding. OpenAI and early testers report that a new Sol Ultra mode, which spawns subagents to run parts of a complex task in parallel, reached 91.9 percent on Terminal-Bench 2.1, with base Sol at 88.8 percent. Both figures sit above Anthropic's Mythos 5 at 88.0 percent and Fable 5 at 84.3 percent.

The skeptical reading matters here. Terminal-Bench is an agentic, terminal-driven evaluation whose scores shift with random seeds, harness configuration, tool timeouts, and retry policy. A 0.8-point gap between base Sol and Mythos 5 falls within the range where two runs of the same model routinely disagree, so the practical claim is parity at the top, not a decisive lead. Ultra mode's larger margin comes from using more inference compute, not from a cheaper capability.

Pricing is where the release is more concrete. Sol lists at $5 per million input tokens and $30 output, Terra at $2.50 and $15, and Luna at $1 and $6, positioning the family below Anthropic's Fable 5 on cost for comparable coding work. OpenAI also confirmed that GPT-5.6 is now the preferred model in Microsoft 365 Copilot, placing the model directly inside Word, Excel, and PowerPoint for enterprise users. The Russian-language developer channel AI ML Big Data separately flagged Sol's record Terminal-Bench number and strong first-tester coding reports.

Veracity: Plausible

71/100

If true, who benefits: OpenAI's enterprise pricing power and Microsoft's Copilot distribution gain from a "top coding model" headline that firms cite when they standardize procurement.
The nuance: The 91.9 percent figure is self-reported on a benchmark OpenAI selected (Terminal-Bench 2.1) and run in its own harness, and the lead over Mythos 5 sits inside normal run-to-run variance, so it is corroborated as a claim but not as an independently reproduced lead.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

The competition at the frontier has narrowed to coding agents and cost per token rather than raw benchmark bragging rights. OpenAI is matching Anthropic on Terminal-Bench while undercutting it on price, and the Microsoft Copilot default gives it distribution Anthropic cannot easily match. The exposed party is Anthropic, whose coding lead was its clearest differentiator, and any lab whose pricing sits above Sol's tiers now competes on a worse cost curve for the same measured capability.

What to watch

Independent reproductions of the 91.9 percent Terminal-Bench figure outside OpenAI's own harness, which would confirm or shrink the claimed lead over Mythos 5.
Whether Sol Ultra's subagent mode raises real inference cost enough to erase the per-token price advantage over Anthropic on long agentic tasks.

Observations to monitor, not financial advice.

3 sources

Synthesized from: OpenAI · OpenAI (Microsoft 365 Copilot) · Polylog editors

Part of a tracked trend

Frontier Labs Race on AI Coding Capability

Coding is becoming a primary competitive battleground among frontier labs, with incumbents standing up permanent coding teams and investing in new training stages (e.g. midtraining) to match leaders like Anthropic; expect recurring reorganizations, benchmarks, and model releases aimed specifically at code.