Review: Claude Opus 5

Anthropic’s July 24 launch of Claude Opus 5 was not another “peak intelligence” press cycle. The pitch was colder and more useful: a model you leave on all day — near Claude Fable 5 on coding and knowledge work, priced like Opus 4.8 at $5 / $25 per million tokens, default on Claude Max, strongest option on Claude Pro.

We ran it as a default coding and research companion for a week: agent loops, long-context briefs, Fast mode sprints, and the usual migration friction from Opus 4.8-era prompts. The SignalSpawn takeaway is simple. Opus 5 is the model most teams should actually standardize on in 2026’s Anthropic stack — and the one that makes Fable 5 feel like a specialist upgrade instead of the only adult in the room.

The product, not the myth

Model id: `claude-opus-5`. Context: 1M tokens, large output ceilings, thinking on by default, effort ladder from low through max. Fast mode claims ~2.5× speed at 2× the base price. Mid-conversation tool changes and automatic safety fallbacks shipped in beta alongside the model — the kind of platform glue that matters more than a single leaderboard spike.

Anthropic’s own boards put Opus 5 at or near the top of Frontier-Bench-style software engineering and several agent/knowledge suites, often within a point of Fable 5 at half the token cost. It stays behind Mythos 5 on the restricted cyber and biology frontiers by design. Independent writeups mostly agree on the shape: a real generation jump from 4.8, with a few teams reporting argumentative or early-stop behavior when old skills and prompts are dropped in blindly.

Hands-on: the daily driver test

Where Opus 5 earns the score is cost-aware competence. Medium effort already clears work that used to demand max on older Opus builds. Long refactors hold state. Tool use is cleaner when you rebuild skills for this generation instead of cargo-culting 4.8 macros. Fast mode is the underrated feature for interactive IDE loops — you feel the latency drop more than you feel the surcharge if your session is short.

Knowledge work — briefs, comparisons, structured research — is where the “half of Fable” claim feels honest. You can stay on Opus 5 for most of a publishing day and only escalate when the problem is genuinely weird. That is the product Anthropic said they were shipping. For once, the week of use matches the blog.

Where it scrapes

Blind swaps from Opus 4.8 skills can underperform; this generation wants prompts rewritten, not just a model id change. Some teams still report stubbornness and incomplete tool chains under elaborate custom skill stacks. Fable 5 remains the better ceiling for the hardest closed-stack problems, and Mythos remains the restricted peak Anthropic will not put on your Pro plan.

Hallucination rate is disclosed as slightly higher than 4.8 in places — not catastrophic, but enough that citation discipline still matters on newsroom and legal-adjacent workflows.

Who it’s for

Claude Max and Pro users who want the default to be the right default. Engineering orgs standardizing one Anthropic model for agents and IDE. Anyone who was paying Fable prices for work Opus 5 now clears. Pause if you need Mythos-class cyber capability, or if your eval harness still prefers GPT-5.6 Sol on a specific coding board you refuse to leave.

The verdict

Claude Opus 5 is the rare frontier release that optimizes for Tuesday afternoon instead of launch-day screenshots. It is not the smartest model Anthropic can run. It is the smartest model most people should leave selected. That is a higher compliment than another leaderboard crown.

Score: 8.8 / 10 — near-Fable workhorse energy at Opus pricing, with Fast mode for when the cursor is waiting on you.

Read the launch note at anthropic.com — Claude Opus 5.