Anthropic

Claude Opus 5

Released Jul 24, 2026 · knowledge cutoff 2026-05

Status: Frontier
Location: United States
Type: LLM
Modality: Multimodal
Context window: 1M
Max output: 128K
Speed (medium effort): 54 tok/s · 6.69s to first answer token
Price ($/MTok in / out): $5 / $25
Cost per task (medium effort): $0.62
Open weights: No

Benchmarks8 of 10 reported

GPQA Diamond

93.7%

SWE-bench Verified

96%

SWE-bench Pro

79.2%

Terminal-Bench 2.1

84.6%

Humanity's Last Exam

56.3%

LMArena Elo

1493

GDPval-AA v2

1824

ARC-AGI-2

90.4%

MMLU-Proretired

—

AIME (latest)retired

—

Strengths & weaknesses

Strengths

Verifies its own work unprompted, delegates readily, and will widen a task's scope on its own judgement
Holds the thread across long multi-step work and pins down vague requirements before building
Reads a large unfamiliar codebase carefully before it changes anything
Solves novel-pattern problems that no earlier model could, rather than recalling them

Weaknesses

Verbose and pushy — it narrates constantly, and default output length responds to prompting more than to the effort setting
Early users describe it as neurotic and apologetic, and less willing to follow a prescribed workflow than Opus 4.8
Easy to misconfigure: effort and skill settings change its behavior more than the version bump does

HLE 56.3% no tools / 64.7% with tools. SWE-bench Verified 96.0% is Anthropic's own system-card figure (five-trial average); a third-party aggregator (vals.ai) cites 97.0%. GPQA Diamond varies 93.4-93.7% across trackers; 93.7% used. ARC-AGI-2 90.4% (max effort) and ARC-AGI-3 30.2% (high effort, a separate untracked scale) per ARC Prize (arcprize.org). Max output is 128K on the synchronous API, up to 300K via the Message Batches API. AIME not published — Anthropic appears to be moving to newer evals (IMO 2026, Frontier-Bench v0.1) instead. LMArena Elo 1493 is the "claude-opus-5-high" text-arena listing (arena.ai, 2026-09-13 update, rank 10); the "claude-opus-5-max" listing sits at 1487. SWE-bench Pro 79.2% is Anthropic's system-card figure (the launch post itself gives only relative claims), as relayed by BenchLM, llm-stats and DataNorth; the same table lists Fable 5 at 80.0 and GPT-5.6 Sol at 64.6. Terminal-Bench 2.1 84.64% is Vals AI's independent-harness measurement at high effort (vals.ai/models/anthropic_claude-opus-5, 2026-07-23) — third-party; Artificial Analysis is reported (via morphllm) to measure 89.1% at max effort in the Terminus 2 harness, but that figure could not be read from AA directly.

For developers

API model strings

anthropicclaude-opus-5
aws-bedrockanthropic.claude-opus-5

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

Replaces Claude Opus 4.8