Claude Opus 5 Ships at $5/$25 — Near-Fable 5 Intelligence

Claude Opus 5 is the new default frontier workhorse — and the best value in the class. Anthropic released it on July 24, 2026 at $5 per million input tokens and $25 per million output, the same price as Opus 4.8 and exactly half of Fable 5. It sets new state-of-the-art on Frontier-Bench v0.1, matches Fable 5 within 0.5% on CursorBench 3.2 at half the cost, and triggers safety classifiers 85% less often than Fable 5. Our verdict moves from pending to Recommended — Opus 5 is the strongest model on Claude Pro and the new default on Claude Max.

What happened

Anthropic shipped Claude Opus 5 on July 24 — its fourth model release in under two months (Mythos 5, Fable 5, Sonnet 5, now Opus 5). But Opus 5 is the one that reshapes the daily equation for most developers.

Performance. Opus 5 scores 1,861 on GDPval-AA — ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736) — and sets SOTA on Frontier-Bench v0.1. On CursorBench 3.2 at max effort, it's within 0.5% of Fable 5's peak at half the cost. ARC-AGI 3: 30.2%, three times any other model's score. Zapier AutomationBench: the only model to complete a full churn-prevention workflow end-to-end, at 1.5x the next-best model's pass rate.

Pricing stability. Opus 5 costs exactly what Opus 4.8 cost — $5/$25 per 1M tokens. No price increase. Anthropic added low/medium/high effort toggles so teams trade cost for capability per request. Fast mode runs at roughly 2.5x speed for 2x the base price.

Safety. Opus 5's safety classifiers intervene 85% less often than Fable 5's. The model was intentionally not trained on cyber exploitation tasks, and in Anthropic's internal behavioral audit, it scored the lowest misaligned-behavior rate of any recent Claude model (2.3). It's also the least susceptible to being tricked into misuse.

Science. Opus 5 is Anthropic's most capable generally available model for scientific research — 10.2 percentage points higher on organic chemistry tasks and 7.7 points higher on protein function prediction versus Opus 4.8 across every life-sciences evaluation.

Platform. Alongside the model, two beta features shipped: mid-conversation tool changes (swap tools without invalidating the prompt cache) and automatic API fallbacks (flagged requests route to a backup model instead of being blocked). Model ID: claude-opus-5, available on all platforms — Claude API, Claude Code, claude.ai, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Why it matters

This isn't just another model launch. Anthropic priced the frontier for everyday use.

For most of 2026, the frontier tier has been something teams rationed — a model you route to for the hardest 5% of tasks while a cheaper workhorse handles volume. Fable 5 at $10/$50 per 1M tokens and with a mandatory 30-day data retention policy was a capable but expensive tool that many teams couldn't adopt for regulated workloads. Opus 5 collapses that tradeoff: the model that wins the benchmark is also the model that costs half as much, carries no retention requirement, and triggers far fewer safety interruptions.

Early-access teams are already reporting strong results. Harvey's Niko Grupen found Opus 5 matched Opus 4.8 at max reasoning while generating 26% fewer tokens. Zapier's Wade Foster reported it was the first model to complete a full churn-prevention workflow. A trading firm built a complete market data feed in a single session — something previous models couldn't do even with detailed plans. On long-horizon agentic tasks, Opus 5's self-verification behavior — building its own test harnesses, running its own validation — is the difference between an assistant you supervise turn-by-turn and one you can hand a multi-hour task.

The directional signal across every frontier release this year is the same: the competitive axis has moved from headline intelligence to cost per completed task. When Anthropic — the vendor with the strongest model on the market — argues its cheaper model is the right default, the era of rationing the frontier is ending.

What changes for you

If you're on Opus 4.8 today. Migrate to Opus 5. Same price, more than double the Frontier-Bench score, +15 points on OSWorld 2.0, and a 30.2% ARC-AGI-3 result against its predecessor's 1.5%. There is no cost argument for staying.

If you're paying for Fable 5. Test the swap on engineering and operations workloads. Opus 5 wins most evaluation rows and ties the coding benchmarks at half the rate, plus it drops the 30-day retention requirement. Keep Fable 5 for specialist domains — it still leads on DeepSWE v1.1, the Legal Agent Benchmark, and raw recall tasks.

If you're building agents. Opus 5's self-verification behavior — checking its own work, constructing test harnesses, catching edge cases — makes multi-hour autonomous runs economically viable in a way they weren't before. Re-test your ceiling on session length.

If you've been blocked by Fable 5's safety classifiers. Opus 5's 85% reduction in classifier interventions is a material quality-of-life change. Biology requests that were refused on Fable 5 now route to Opus 5 instead of Opus 4.8.

FAQ

Is Opus 5 better than Fable 5? Better in most dimensions at half the price — SOTA on Frontier-Bench, GDPval-AA, ARC-AGI-3, OSWorld 2.0, and AutomationBench. But Fable 5 still edges ahead on DeepSWE v1.1, the Legal Agent Benchmark, and Humanity's Last Exam without tools. For everyday engineering work, the gap has narrowed to the point where paying double no longer makes sense.

Does Opus 5 have the same data retention as Fable 5? No. Unlike Mythos-class models (Fable 5, Mythos 5), Opus 5 carries no data retention requirement for general access — the same posture as Opus 4.8. That opens the door for regulated workloads that couldn't adopt Fable 5.

When should I still use Fable 5? For specialist legal, clinical, and narrow professional domains where it keeps a lead, and for tasks where raw recall matters more than tool-augmented reasoning. Also for teams already approved under Fable 5's Cyber Verification Program who need fewer restrictions than Opus 5's classifiers allow.