Claude Opus 5 Delivers Near-Fable Intelligence at Half the Price

Anthropic announced the official claude opus 5 release on July 24, and the pitch is straightforward: get close to the ceiling-tier intelligence of Claude Fable 5 without paying ceiling-tier prices. Claude Opus 5 is now the default model on Claude Max and the strongest model available on Claude Pro, with immediate availability across the API, Amazon Bedrock, Microsoft Foundry, and GitHub Copilot.

The release is not a subtle increment. Anthropic claims the new claude 5 opus model more than doubles its predecessor's score on Frontier-Bench, triples the next-best model on ARC-AGI 3, and matches Fable 5 within half a percent on CursorBench 3.2 — all at the same $5/$25 per million token price point that Opus 4.8 charged.

Model vitals

Claude 5 Opus: The benchmarks that matter for LLM coding benchmark results

Anthropic published results across nine evaluations, with Opus 5 claiming the top spot on most of them. The standout numbers:

Frontier-Bench v0.1 measures real-world software engineering tasks — Opus 5 surpasses every other model and more than doubles Opus 4.8's performance at lower cost per task. On CursorBench 3.2, the model performs within 0.5% of Fable 5 at peak effort, but at half the cost. On the AA Coding Agent Index, it outperforms all other models at any given price point.

The reasoning results are equally aggressive. On ARC-AGI 3, which tests novel problem-solving, Opus 5 scores three times higher than the next-best model. On Zapier AutomationBench, its pass rate is roughly 1.5x the runner-up at the same cost — and even at its lowest effort setting, Opus 5 passes more business-automation tasks than any other model at any setting. On OSWorld 2.0, a computer-use benchmark, it surpasses Fable 5's best score at just over a third of the cost.

The one area where Opus 5 explicitly does not lead: cybersecurity. Anthropic notes it remains behind the specialized Mythos 5 model on offensive security tasks.

The effort dial

The most interesting new feature is not about raw intelligence — it is about giving developers a cost knob. Opus 5 ships with an effort setting (low, medium, high, max) that controls how many thinking tokens the model burns before answering. At low effort, responses are fast and cheap. At max effort, the model reasons extensively and approaches Fable-tier quality.

This is a practical shift. Instead of choosing between a cheap model and an expensive one, developers can now choose a single model and tune its behavior per request. A quick triage call uses low effort; a complex debugging session uses max. The model also ships with thinking enabled by default, which is a first for the Opus tier — previous versions required explicit opt-in.

Anthropic also introduced mid-conversation tool changes (in beta), allowing users to add or remove tools between turns while preserving the prompt cache. For agentic workflows that evolve as a task progresses, this removes a meaningful friction point.

What early testers are saying

The early-access anecdotes from Anthropic's blog are worth noting for what they reveal about the model's agentic behavior:

On one Frontier-Bench task, Opus 5 was given a drawing of a machine part and asked to rebuild it as a 3D FreeCAD model — but intentionally given no way to view the drawing directly. The model responded by writing its own computer vision pipeline to pull geometry from raw pixels, then reconstructed the full part. No competing model solved it.

A trading firm engineer used Opus 5 to build a market data feed for a new exchange in a single session. When no live feed was available for validation, the model built its own test harness to verify its code parsed the exchange's data format correctly.

On scientific tasks, Anthropic reports improvements across all life sciences evaluations compared to Opus 4.8, with the largest gains in organic chemistry (10.2 percentage points higher on spectroscopy-to-structure tasks) and protein function prediction (7.7 points higher).

Pricing and availability

The pricing is identical to Opus 4.8, which means Opus 5 is a pure performance upgrade at the same cost. The fast mode option doubles the price but delivers roughly 2.5x the speed for latency-sensitive applications.

Opus 5 is available immediately on the Anthropic API, Claude.ai (Max and Pro), Claude Code, Amazon Bedrock, Microsoft Foundry, and GitHub Copilot (Pro+, Max, Business, and Enterprise plans).

Safety claims

Anthropic says Opus 5 has the lowest rates of deceptive and misaligned behavior on their automated behavioral audits. It includes elevated safeguards for high-harm cybersecurity content and adheres strictly to the Anthropic Constitution. These are self-reported claims from Anthropic's own evaluations — independent safety assessments have not yet been published.

What this means

Opus 5 compresses the gap between Anthropic's mid-tier and frontier-tier models to an uncomfortably narrow margin. If the benchmarks hold up under independent evaluation, the practical question becomes whether Fable 5 can justify its 2x price premium for any workload that is not specifically cybersecurity-related.

The effort dial is arguably the more consequential feature. It reframes model selection from "which model?" to "how hard should this model think?" — a shift that every provider will likely follow. Combined with mid-conversation tool changes and thinking-by-default, Opus 5 is not just a faster Opus 4.8. It is a meaningfully different kind of product.

Note: all benchmark figures in this article come from Anthropic's launch materials. Independent reproduction has not yet been published.