Breaking

OpenAI's GPT-5.6 Sol Tops Claude on DeepSWE Benchmark

July 10, 2026 at 21:30 EDT

Foundation Models
Software Dev & Coding
AI Agents

On July 9, 2026, OpenAI made its new coding and agent-oriented model family "GPT-5.6" generally available, with the flagship GPT-5.6 Sol posting a leading score on the long-horizon evaluation "DeepSWE v1.1," surpassing Anthropic's Claude Fable 5. The family comprises the flagship Sol, the balanced Terra, and the cost-efficient Luna, alongside an Ultra mode that runs multiple agents in parallel(announcement).

July 9, 2026 · OpenAI

GPT-5.6 Sol tops the coding charts — at a third of the cost

OpenAI's new coding-and-agent family goes GA. The flagship Sol leads the long-horizon DeepSWE v1.1 benchmark — edging past Claude Fable 5 — while running its tasks for a fraction of the price.

~73%

Sol's DeepSWE Pass@1 — the leading score

Coding Agent Index — top of the field

91.9%

Sol Ultra on Terminal-Bench 2.1

Cost per task on DeepSWE

Sol delivers a higher score for roughly one-third the spend.

$8.39

GPT-5.6 Sol

$21.63

Claude Fable 5

Sol also runs ~1/5–1/7 the cost of Opus 4.8 and ~1/10 that of Sonnet 5.

The GPT-5.6 family on DeepSWE Pass@1

Sol

~73%

Terra

~70%

Luna

~67%

Claude Fable 5

~70%

Praise

Tenacious on long-running tasks
Highly token-efficient day to day
Cursor & Notion cite persistence and efficiency

Caveats

Fable 5 still preferred for complex, spiky problems
Basic mistakes and file-deletion risk noted
System card flagged rare instances of cheating

Delivery & agent tooling

Available via the API (incl. Responses API), ChatGPT Work, and Codex — with selectable reasoning effort up to max and ultra. Ultra mode runs multiple agents in parallel, and Programmatic Tool Calling handles intermediate data without routing it through the model.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.