Claude Opus 5 — Anthropic Just Made Frontier AI Affordable. The Benchmarks Back It Up.

In less than two months, Anthropic has released three frontier AI models. Claude Fable 5 came first — the company's most capable model ever, restricted in access due to its power. Then Claude Sonnet 5, the everyday workhorse. And now, on July 24, 2026, Claude Opus 5 — the model that sits between them and, according to the benchmarks Anthropic published at launch, does something genuinely unusual: it beats Fable 5 on several tests while costing exactly half as much per token.

This is not a minor release. Anthropic's own framing calls it "a generational leap" — and for once that kind of language is backed up by numbers that are hard to dismiss. Opus 5 scores 43.3% on Frontier-Bench v0.1, the industry's most rigorous software engineering benchmark — more than doubling its predecessor Opus 4.8's score of 21.1%. It triples the next-best model on ARC-AGI 3. It beats Fable 5 on computer use at a third of the cost. And it arrives at the same API price as the model it replaces — $5 per million input tokens, $25 per million output. No price increase for a generation-defining capability jump. That combination is what makes this week's release the most important AI story of July 2026.

What Opus 5 Actually Is — And Where It Sits

Claude Opus 5 is Anthropic's new top-tier production model, sitting between Claude Sonnet 5 and Claude Fable 5 in the company's current lineup. It is available immediately across every Anthropic platform — Claude.ai, the Anthropic API, Claude Code, Claude Cowork, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry — making this one of Anthropic's broadest simultaneous rollouts to date. The model ID is claude-opus-5.

For Claude Max subscribers, Opus 5 is now the default model — the one that powers every conversation unless you select otherwise. For Claude Pro subscribers, it is the strongest available option. The positioning is deliberate: Anthropic is not hiding its best production model behind a paywall that prices out developers and enterprises. It is making Opus 5 the experience most people encounter first.

The release also signals a broader strategic shift. In less than two months, Anthropic has introduced Claude Fable 5, Claude Sonnet 5, and now Claude Opus 5, signalling a shift away from occasional flagship launches toward a portfolio of models optimised for different budgets and workloads. The days of one model per year are over. Anthropic is now running an AI product line that competes at every price point simultaneously.

The Benchmarks — What the Numbers Actually Mean

Benchmark claims are easy to dismiss as marketing. Opus 5's numbers are harder to wave away than most because they were independently tested and verified by developers running real workloads within 24 hours of launch.

Frontier-Bench v0.1 is the most respected software engineering benchmark in the industry — it tests real coding tasks across Terraform, Kubernetes, Bash, Python, and full-stack web work, with automated linters rather than human judges as the scoring mechanism. Scoring 43.3% on Frontier-Bench, doubling its predecessor, Opus 5 is not just an incremental update — it is a fundamentally more capable system. The previous leader on this benchmark was Fable 5. Opus 5 surpasses it.

ARC-AGI 3 tests abstract reasoning and generalisation — the ability to solve novel problems using patterns rather than memorised answers. Tripling the next-best model on this benchmark is the kind of result that gets attention from AI researchers who are otherwise sceptical of commercial benchmark claims. It suggests Opus 5 is doing something qualitatively different from its predecessor in how it approaches problems it has not seen before.

OSWorld 2.0 tests computer use — the ability to control a real computer interface, navigate applications, complete tasks, and recover from errors the way a human operator would. On OSWorld 2.0, Opus 5 surpasses Fable 5's best result at just over a third of the cost. For anyone building AI agents that interact with software — automation tools, coding assistants, browser agents — this is the benchmark that matters most. Beating Fable 5 on it at a third of the price is a direct commercial argument for switching.

What Is Actually New — The Capabilities That Matter

Self-verification and error correction. According to Anthropic, Opus 5 is better at verifying its own work, revising intermediate results, and continuing complex tasks until they are completed. This is the capability that separates models that are useful for a quick answer from models that can run an autonomous multi-step task reliably. If Opus 5 makes a mistake partway through a complex workflow, it is more likely to catch and fix it before handing the result back — which is what makes it viable as an agent rather than just a chatbot.

Computer use. Opus 5 can control a real browser and desktop environment — opening pages, navigating interfaces, filling in forms, catching UI bugs, and completing multi-step workflows without human intervention. On its benchmark, Opus 5 opened pages in a browser at desktop and phone widths, caught a product hidden below the mobile fold and an off-screen checkout button, and fixed both before handing the work back. For developers building browser agents, this is the capability that was previously only available at Fable 5 pricing.

Scientific research. Life sciences workloads see major improvements over Opus 4.8. Organic chemistry gains 10.2 percentage points on internal benchmarks — tasks like inferring molecular structures from spectroscopy data. Protein sequence prediction improves by 7.7 points. For researchers using AI to accelerate laboratory work, Opus 5 at Opus 4.8 pricing is a direct cost reduction for the same capability level.

Visual outputs. Opus 5 can generate stronger visual content than any previous Claude model — including interactive technical visualisations. Anthropic demonstrated it producing a working wind tunnel simulation showing airflow over aerodynamic objects, generated entirely from a prompt. This puts complex data visualisation within reach of users who previously needed dedicated graphics tools or engineering support.

The Effort Level System — A Feature Worth Understanding

One of the less-discussed additions in Opus 5 is its adjustable effort settings. Rather than always running at maximum processing depth — which costs tokens and takes time — Opus 5 lets you dial the effort level up or down depending on the task. Quick factual lookups run cheap and fast. Complex multi-step reasoning problems run deep and thorough. This is a significant practical improvement for anyone running Opus 5 at scale: you only pay for the processing depth the task actually requires.

The Fast Mode variant doubles the price per token but delivers approximately 2.5× the speed — a trade-off designed for latency-sensitive production environments where response time matters more than per-token cost. For most developers running batch processing, the default effort settings will be sufficient. For real-time user-facing applications, Fast Mode changes what is possible.

Who Should Switch — And Who Should Wait

For developers on Opus 4.8: switch. The migration is a model-ID change — the API shape and Claude Code integration carry over unchanged. You get double the benchmark performance at the same price. There is no reasonable argument for staying on 4.8 unless you have a specific workflow that has been tested and certified on the old model.

For Claude Max subscribers: Opus 5 is already your default. You are already using it. The question is whether to explore the effort settings and Fast Mode for specific use cases.

For enterprises on Fable 5: evaluate the specific benchmarks that match your workload. On software engineering, computer use, and automation tasks, Opus 5 matches or beats Fable 5 at half the cost. On cybersecurity vulnerability research and health/biology workloads, Fable 5 and Mythos 5 still lead. The answer depends on what you are actually doing with the model.