OpenAI Announces GPT-5.6 Sol: Dominating Terminal Bench 3.0 (34.6%) and DeepSWE (72.7%) with Sol Architecture

Por Sarah Lin • Publicado el 2026-09-10 • 5 min read • Fuente: OpenAI Research

OpenAI unveils GPT-5.6 Sol, featuring autonomous agent test-time compute, setting all-time highs across SWE-Bench Verified, DeepSWE, and Terminal Bench 3.0.

OpenAI has officially launched **GPT-5.6 Sol**, the commercial flagship designed specifically for end-to-end software engineering agents and autonomous systems.

Sol Test-Time Compute Architecture GPT-5.6 Sol integrates deep search during code generation. When tasked with fixing an open-source issue, Sol simulates compiler outputs, generates test fixtures in ephemeral containers, and verifies solutions before returning final code diffs.

Record Scores across Benchmark Suites - DeepSWE (v1.1): 72.7% (World Record). - Terminal Bench 3.0: 34.6% (World Record). - ExploitGym (6h): 293 verified exploit chains, establishing unmatched proficiency in automated security patching.

Fuente y Verificación

Este informe técnico fue contrastado contra la documentación primaria publicada por OpenAI Research.

Leer Anuncio Original en OpenAI Research →

OpenAI Announces GPT-5.6 Sol: Dominating Terminal Bench 3.0 (34.6%) and DeepSWE (72.7%) with Sol Architecture

Sol Test-Time Compute Architecture GPT-5.6 Sol integrates deep search during code generation. When tasked with fixing an open-source issue, Sol simulates compiler outputs, generates test fixtures in ephemeral containers, and verifies solutions before returning final code diffs.

Record Scores across Benchmark Suites - **DeepSWE (v1.1)**: **72.7%** (World Record). - **Terminal Bench 3.0**: **34.6%** (World Record). - **ExploitGym (6h)**: 293 verified exploit chains, establishing unmatched proficiency in automated security patching.

Fuente y Verificación

Record Scores across Benchmark Suites - DeepSWE (v1.1): 72.7% (World Record). - Terminal Bench 3.0: 34.6% (World Record). - ExploitGym (6h): 293 verified exploit chains, establishing unmatched proficiency in automated security patching.