Claude Opus 5

Claude Opus 5 is Anthropic’s July 24, 2026 flagship — the successor to Opus 4.8 and, on several independent benchmarks, the new overall intelligence leader. The pitch: near-Fable 5 capability at half the price, with thinking on by default, a May 2026 knowledge cutoff, and agentic scores that beat every other model on tasks ranging from novel problem-solving to computer use to business automation.

The headline number is Frontier-Bench v0.1 (a 74-task successor to Terminal-Bench 2.1): Opus 5 scored 43.3% at max effort, more than doubling Opus 4.8’s 18.7% and clearing GPT-5.6 Sol (37.5%) and Fable 5 (33.7%). It is, by Anthropic’s framing, a “thoughtful and proactive” model — one that plans before acting, flags uncertainty, and executes long-horizon tasks with less human nudging.

Benchmarks

Anthropic published an unusually wide benchmark table at launch. Numbers below are vendor-reported (July 24, 2026); treat as a dated snapshot and prefer independent replications.

Three results define the gap:

• ARC-AGI-3 is the standout: 30.2% is roughly 4× the next-best model (GPT-5.6 Sol at 7.8%). This is a test of novel, out-of-distribution reasoning — the kind that memorized patterns can’t help with.
• OSWorld 2.0: 70.6% surpasses Fable 5’s best result at roughly a third of the cost per task. Computer-use agents just got meaningfully cheaper.
• SWE-bench Pro remains Fable 5 territory: 79.2% vs 80.3%. For the very hardest, longest-horizon coding tasks, Fable is still the peak — but the gap is now roughly 1 point, not 11.

Pricing

Pricing is unchanged from Opus 4.8 — and exactly half of Fable 5 ($10/$50). The cost argument is simple: unless you need Fable’s raw ceiling on the hardest coding or cybersecurity tasks, Opus 5 is a more efficient pick.

Thinking and effort

Thinking is on by default — and unlike Opus 4.8, you cannot disable it above high effort. Attempting to set thinking.enabled = false with effort above high returns a 400 error. This is a breaking change: if your pipeline assumed it could always toggle thinking off, wrap the API call.

The effort dial still runs five levels (low, medium, high, xhigh, max), with high as the default. The interesting data point from Frontier-Bench: xhigh (44.4%) actually outperforms max (43.3%) — suggesting the model occasionally overthinks at maximum effort. Anthropic recommends starting at high and escalating only when the task warrants it.

What’s new vs Opus 4.8

Beyond the benchmark jump, two API-level features ship in beta:

• Mid-conversation tool changes: add or remove tools between turns while preserving the prompt cache, instead of resending a fixed tool list for the entire session. Use the mid-conversation-tool-changes-2026-07-01 beta header.
• Server-side model fallback: a new "default" mode in the fallbacks parameter that applies Anthropic’s recommended fallback models by refusal category, instead of you maintaining a model list. Use the server-side-fallback-2026-07-01 beta header.

Safety

Anthropic reports Opus 5 has the lowest misalignment score (2.30) of any Claude model, with no default data retention. Cyber safeguards relax only for source-code vulnerability finding; exploitation paths stay blocked (same policy as Opus 4.8).

One honest caveat, from Anthropic’s own system card: Opus 5 has slightly higher factual hallucination than Opus 4.8. This is a real trade-off — the model gained creativity and novel-reasoning capability, and the hallucination calibration that made 4.8 famous took a small hit.

Availability

Shipped broadly on July 24, 2026:

How Claude Opus 5 compares

• vs Fable 5: Opus 5 now leads on Frontier-Bench, ARC-AGI-3 and OSWorld 2.0. Fable keeps a ~1-point edge on SWE-bench Pro and retains the ceiling on cybersecurity/biology. The 2× price gap makes Opus 5 the default choice unless you specifically need Fable’s strengths.
• vs GPT-5.6 Sol: Opus 5 takes the AA Intelligence Index (#1 vs #2) and ARC-AGI-3 by a large margin. Sol takes the output price advantage ($30 vs $25 — but Sol’s $5 input is the same). Both are text+image in, text out.
• vs Opus 4.8: a large step up everywhere — Frontier-Bench more than doubles, ARC-AGI-3 jumps from 1.5% to 30.2%, OSWorld from 55.7% to 70.6%. Same price, same API shape. The main reason to stay on 4.8 is the marginally better hallucination calibration.

The Claude lineup (July 2026)

Opus 5 replaces Opus 4.8 outright — same price, better on every published benchmark except hallucination calibration. The question it answers: you no longer need to route to Fable for frontier-level performance. The question it raises: how long before Opus 5’s thinking-on-by-default token usage makes the per-token pricing feel more expensive per task than it looks.

Best for

• Agentic coding — CursorBench, Frontier-Bench, SWE-bench Verified 96%
• Computer-use automation (OSWorld 2.0 70.6%)
• Business task automation (Zapier AutomationBench 26%, ~1.5× next-best)
• Complex reasoning and novel problem-solving (ARC-AGI-3, IMO gold-medal level)

Pros & cons

Frequently asked questions

Sources

• Anthropic — Introducing Claude Opus 5
• Anthropic — Claude Opus 5 System Card (PDF)
• Anthropic — What's New in Opus 5
• Anthropic — Pricing
• TECHSY — Claude Opus 5: What Changed and Where It Loses
• MarkTechPost — Claude Opus 5: Frontier-Class Agentic Coding