High
Claude Opus 5 Benchmark Consolidation: Frontier-Bench #1 (43.3%), ARC-AGI-3 SOTA (30.2%), SWE-bench Pro 79.2%
What changed
Claude Opus 5's SWE-bench Pro entry appeared July 25 (79.2%), completing the benchmark picture after the July 24 launch. Across four key developer benchmarks, Opus 5 leads the frontier on reasoning and agentic coding while matching Fable 5 on software engineering tasks — at half the price.
TL;DR
Claude Opus 5 scores 43.3% on Frontier-Bench (agentic coding, leads all models), 30.2% on ARC-AGI-3 (>3× the prior record of 7.8% set by GPT-5.6 Sol), 96% on SWE-bench Verified, and 79.2% on SWE-bench Pro (vs Fable 5 at 80.3%) — at $5/$25 per MTok vs Fable 5's $10/$50.
Developer signal
Four benchmarks, one clear takeaway: Opus 5 is optimized for agency and reasoning, not general-purpose breadth. (1) Frontier-Bench v0.1 (74-task agentic coding benchmark, successor to Terminal-Bench 2.1): Opus 5 at 43.3% max effort leads GPT-5.6 Sol (37.5%) and Fable 5 (33.7%) by a 6–10 point margin. If you are building agentic coding pipelines or computer-use workflows, Opus 5 at $5/$25 is now the primary option to benchmark — it leads the frontier and costs half of Fable 5. (2) ARC-AGI-3 (novel reasoning; no memorization possible by design): Opus 5 at 30.2% (High effort) is more than three times the previous best of 7.8% by GPT-5.6 Sol running at Max effort. ARC Prize confirmed it solved five previously unbeaten environments. The caveat: ARC-AGI-3 was evaluated at High effort only, not Max, due to the short testing window — the Max-effort score is not yet available. (3) SWE-bench Pro (harder version of SWE-bench Verified; leaderboard updated July 25): Opus 5 at 79.2% is third behind Mythos 5 (80.3%) and Fable 5 (80.0%), and ahead of its predecessor Opus 4.8 (69.2%) by a full 10 percentage points. The Fable 5 vs Opus 5 gap on SWE-bench Pro (1.1 pp) is narrower than on Frontier-Bench (9.6 pp in Opus 5's favor) — suggesting different optimization targets. (4) SWE-bench Verified: Opus 5 at 96%, Mythos 5 at 95.5%, Fable 5 at 95% — the three Anthropic models cluster at the top. Practical guidance: for long-running agentic coding agents and computer-use tasks, switch to Opus 5; for knowledge-work breadth, Fable 5 remains the default. The price differential ($5/$25 vs $10/$50) is now decisively in favor of Opus 5 for agentic use cases.
Affects you ifYou are choosing between Claude Opus 5 and Claude Fable 5 for an agentic coding, computer-use, or reasoning-heavy workload; you are benchmarking models for production agent selection.EffortQuick — Opus 5 is a drop-in swap from Fable 5 (
claude-fable-5 → claude-opus-5); pricing favors Opus 5 for agentic workloads. Verify output quality on your specific task distribution before committing.SWE-bench Pro Leaderboard / ARC Prize (official) / Anthropic launch materials | Date: July 25, 2026 (SWE-bench Pro leaderboard entry; ARC Prize official X post July 24) | Link: https://llm-stats.com/benchmarks/swe-bench-proARC Prize official X post (July 24): https://x.com/arcprize/status/2080716561539907928 | SWE-bench Pro leaderboard: https://llm-stats.com/benchmarks/swe-bench-pro | Frontier-Bench: https://codersera.com/blog/claude-opus-5-benchmarks-explained-2026/