GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

GPT-5.5 Leads Terminal-Bench 2.1 at 78.2% — The One Benchmark Opus 4.8 Still Trails

Claude Opus 4.8 leads on 12 of 13 major agentic benchmarks. Terminal-Bench 2.1 is the exception. GPT-5.5 scores 78.2% on the CLI automation benchmark, 3.6 points ahead of Opus 4.8’s 74.6%. Anthropic’s own system card includes the comparison without burying it.

The previous leader on Terminal-Bench 2.1 was Gemini 3.5 Flash at 76.2%, the same model that held top spot earlier this year on price-performance grounds. Both GPT-5.5 and Opus 4.8 have since cleared that number — GPT-5.5 by a wider margin.

Current Leaderboard

ModelTerminal-Bench 2.1
GPT-5.578.2%
Gemini 3.5 Flash76.2% (prior leader)
Claude Opus 4.874.6%
Claude Opus 4.766.1%
DeepSeek V4 Pro67.9%
MiniMax M366.0%
Gemini 3.1 Pro68.1%

Terminal-Bench 2.1 measures autonomous multi-step shell execution: a model gets a terminal, a task, and no hand-holding. It has to issue commands, read error output, recover from failures, and complete the task without human intervention. Scores represent the percentage of tasks completed end-to-end.

What the Gap Means

GPT-5.5’s 78.2% advantage is consistent across runs per multiple independent analyses. The model handles specific error patterns — particularly correction loops on shell execution failures — more robustly than Opus 4.8. Opus 4.8 gets stuck in those loops at a measurably higher rate.

Against every other major agentic benchmark, Opus 4.8 is ahead:

  • SWE-bench Pro: 69.2% (Opus 4.8) vs 58.6% (GPT-5.5)
  • OSWorld-Verified computer use: 83.4% vs ~79%
  • GDPval-AA knowledge work: 1890 Elo vs 1769
  • Humanity’s Last Exam with tools: 57.9% vs ~52%

The pattern is clear: Opus 4.8 is the stronger general-purpose and coding agent. GPT-5.5 is the stronger terminal-native agent. Neither gap is large enough that scaffolding can’t close it — Terminal-Bench results are harness-sensitive, and the vix scaffold that currently holds the overall Terminal-Bench 2.0 lead runs Claude Opus 4.7.

Selection Guidance

Teams building overnight CLI agents that run autonomously on infrastructure tasks have a genuine reason to favor GPT-5.5. Teams building coding agents, multi-turn research tasks, or computer-use pipelines should favor Opus 4.8 by a larger margin.

The frontier at mid-2026 has fractured by workload type. Single-model stacks optimized around one benchmark are increasingly wrong.