guide · ai

Claude Opus 5

Claude Opus 5 (Anthropic, July 24, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

A retro robot representing Claude Opus 5

The release

Anthropic released Claude Opus 5 on July 24, 2026 — the fifth major Claude release in eight weeks, and the first time a non-frontier Anthropic model has actually outscored Fable 5 on multiple coding and knowledge-work evaluations. It ships at the same $5/$25 per million input/output tokens as Opus 4.8 and is the new default model on Claude Max. ARC Prize ran it independently the same day and reported a 30.16% score on ARC-AGI-3 at High effort, roughly 4× the previous SOTA.

Why it matters for operators

Opus 5 is the rare mid-tier model that’s worth migrating to on day one. It’s not “almost as good as Fable 5” — on several benchmarks it actually beats Fable 5 (GDPval-AA v2, Frontier-Bench v0.1, OSWorld 2.0) at roughly half the cost per task. The diagnostic-coding numbers are the standout: 85% on debugging tasks versus GPT-5.6’s 5%, 67% on build-and-dependency management versus GPT-5.6’s 40%. For ABS-style long-running coding sub-agents — repo refactors, root-cause analysis, full-stack builds — those category wins matter more than headline leaderboard scores. The two non-benchmark features that change the calculus for operators: (1) zero data retention by default (unlike Fable 5 and Mythos 5, which carry 30-day retention), so routing customer PII through Opus 5 doesn’t need a BAA or retention carve-out; and (2) the strongest prompt-injection resistance of any Claude model yet, per Anthropic engineer Boris Cherny and the system card. If that holds under independent red-teaming, it’s a real unlock for letting agents ingest untrusted web, email, and PDF content without a sandbox layer. The effort parameter (low / medium / high / xhigh / max) is now a first-class dial. Default is high. Drop to medium for caption drafts and classification, push to xhigh/max for the gnarly sub-tasks where Fable-class reasoning matters.

Specs that matter

| Spec | Value | |---|---| | API model ID | claude-opus-5 | | Context window | 1,000,000 tokens (1M) | | Max output | 128K tokens (300K via Batch API with beta header) | | Knowledge cutoff | May 2026 (most current of any Claude model) | | Pricing (input/output) | $5 / $25 per MTok — same as Opus 4.8 | | Fast mode | ~2.5× speed at 2× price → $10 / $50 per MTok | | Batch API | $2.50 / $12.50 per MTok | | Cache writes (5m / 1h) | $6.25 / $10 per MTok | | Cache hits/refreshes | $0.50 per MTok | | Data retention | None (zero-data-retention eligible) | | Effort parameter | low / medium / high / xhigh / max (default high) | | Adaptive thinking | On by default | | Modalities | Text, image (vision), tool use / agentic, PDF upload, voice mode | Sources: Anthropic launch announcement and Anthropic platform pricing docs.

Benchmarks

The headline is ARC-AGI-3: 30.16% at High effort — independently verified by ARC Prize on July 24, 2026. Previous SOTA was GPT-5.6 Sol at 7.8% (Max effort), so Opus 5 is roughly 4× higher. ARC Prize ran the model themselves; Anthropic did not control the harness. ARC Prize’s caveat: Opus 5 was tested only at High effort (not Max) due to pre-launch time pressure. ARC Prize notes “the gap remains large even accounting for this,” but it’s not a clean effort-matched comparison. ARC-AGI-1 is saturated at 97.5%; ARC-AGI-2 is 88.3% (High) / 90.4% (Max). A separate independent check raised concerns: Guanghan Ning’s Witness benchmark, a different puzzle eval in the same spirit, tied Opus 5 with Fable 5 and Kimi K3 (≈43 score), suggesting some ARC-AGI-3 gains may partially reflect training on genre-specific puzzle data rather than general fluid reasoning. Other real numbers from the brief: - SWE-bench Verified: 96.0% (vs. Opus 4.8 at 88.6%, Fable 5 at 95.0%) - Frontier-Bench v0.1: 43.3% — more than 2× Opus 4.8’s 18.7%, beats Fable 5’s 33.7% at half the cost per task - Senior SWE-bench (Snorkel): 28.2% — #2 behind Fable 5 (29.1%) but at one-third the token cost; 85% on debugging, 67% on build-and-dependency management - OSWorld 2.0 (computer use): 70.6% — beats Fable 5 and GPT-5.6 Sol (62.6%) at one-third Fable 5’s cost per task - GDPval-AA v2 Elo: 1,861 — beats Fable 5 (1,747) and all GPT-5.6 variants Where it loses: DeepSWE v1.1 (long-horizon coding) trails GPT-5.6 Sol 68.8% → 72.7%.

Strengths

Best price-to-intelligence ratio Anthropic has ever shipped. Near-Fable-5 capability at half the cost; on Frontier-Bench, GDPval-AA, and OSWorld 2.0 it actually beats Fable 5. - Diagnostic and root-cause coding work is its standout. 85% debugging / 67% build tasks (vs. GPT-5.6’s 5% / 40%) — meaningful for agents doing refactors, root-cause analysis, and long-running repo work. - Zero data retention by default. Unlike Fable 5 and Mythos 5 (30-day retention), Opus 5 is zero-data-retention eligible. Operationally significant for anything touching customer PII. - Strongest prompt-injection resistance of any Claude yet. If it holds under independent red-teaming, removes one of the last serious blockers to giving agents real system access. - Adaptive thinking + mid-conversation tool changes (beta). Designed for long-running agent workflows; the tool-change feature avoids invalidating the prompt cache when adding capabilities mid-session.

Weaknesses

Below Fable 5 on the absolute hardest tasks. Within 0.5% on CursorBench at max effort, but Fable 5 still leads where the gap matters. DeepSWE v1.1: Opus 5 trails GPT-5.6 Sol. - 35% of failures are faulty inference / verification (Snorkel trajectory analysis) — the bottleneck is reasoning, not tooling, and a leaderboard doesn’t show it. - ARC-AGI-3 gains may not transfer cleanly to novel puzzles. Witness benchmark tie with Fable 5 and Kimi K3 raises concerns about training-on-genre contamination. - Not effort-matched on the headline number. Tested at High; competitors tested at Max. Still a real gap, but not a clean comparison. - Below Mythos 5 on cybersecurity exploitation. Near Mythos on vulnerability finding, but substantially behind on turning vulns into material threats — keep Mythos for red-team work. - Context window (1M) is large but not class-leading. Matches Opus 4.6/4.7/4.8; not a step-change over the Opus 4.x line.

The operator call

Migrate Opus 4.8 → Opus 5 as your default production brain. It’s the same price, substantially better on the things ABS agents actually do, and zero-data-retention by default. Swap claude-opus-4-8 → claude-opus-5 in your Hermes config;