The company's cheapest flagship in years beats its own most powerful commercial model on most benchmarks. The pitch is aimed squarely at CFOs.

For three years, frontier AI labs competed on a single question: what can your best model do on its best day? On Friday, Anthropic released a model that quietly changes the question to something far less glamorous and far more consequential for the people signing enterprise contracts: what does it cost to run this all day, every day?

Claude Opus 5 launched July 24, 2026, priced at $5 per million input tokens and $25 per million output tokens. That is identical to its predecessor, Opus 4.8, and exactly half the price of Claude Fable 5, the company's most powerful publicly available model. The twist is that, according to Anthropic's published benchmark results, the lower-priced model matches or outperforms Claude Fable 5 across most of the company's key evaluations while costing half as much to run.

The numbers that reframe the launch

Anthropic's headline result comes from Frontier-Bench v0.1, a 74-task agentic coding evaluation that measures whether a model can carry a real software engineering job end to end through a terminal.

Figures as published by the vendor at launch. Effort settings vary by row.

Based on Anthropic's published benchmark data, three themes stand out.

The generational jump is real on agentic work. Frontier-Bench more than doubled, from 18.7% to 43.3%, rising to 44.4% at the highest effort setting. ARC-AGI-3, a test built specifically to resist memorisation, went from 1.5% to 30.2%, roughly four times the next-best model on the board.

The flagship-versus-workhorse gap has collapsed. On the coding evaluations where Fable 5 still competes, the margins are inside statistical noise: on one industry coding benchmark Opus 5 lands within 0.5% of Fable 5's peak at half the cost per task. On another it trails by a tenth of a percentage point. SWE-bench Pro is the one place Opus 5 finishes third, at 79.2% against 80.0% and 80.3%, a gap of less than a point.

Cost per task is the metric being sold. At medium effort, Opus 5 still clears 24% on the end-to-end business automation benchmark at roughly $0.89 per task. Even at its lowest effort setting, the company says, it passes more of those tasks than any rival model at any effort level.

Why the price line matters more than the benchmark line

Enterprise AI budgets in 2026 are being eaten alive by agents. A chatbot answers a question and stops. An agent reads a codebase, runs tests, opens a browser, checks its own work, and loops, burning tokens the entire time. Token consumption per useful outcome, not per query, has become the defining line item. That is the gap Opus 5 is engineered for. Anthropic's argument is that the most economically valuable AI work sits in a middle band of difficulty where near-frontier intelligence delivered efficiently beats absolute frontier intelligence delivered expensively. To support that argument, Anthropic highlighted results from several early-access enterprise customers that tested Opus 5 in production-like workloads:

• A financial-modelling team reported 9 percentage points higher accuracy with a third fewer turns and tool calls and 60% less time across effort levels.
• A trading desk measured the strongest Opus result on its internal benchmark using roughly one-seventh the reasoning tokens and under half the latency of Opus 4.8.
• A legal AI firm matched prior quality while generating 26% fewer tokens on average against Opus 4.8 at maximum reasoning.
• An enterprise content platform logged an 8% overall gain, rising to 11% on data analysis and 17% on due diligence.
• A workflow automation company saw the model take a raw account-health workbook through a full churn-prevention sequence and score 100% on a task no previous model had passed.

Those are procurement numbers, not demo numbers. Fewer turns and fewer tokens for equal or better output is the shape of an argument you can put in front of a finance committee.

The safety layer becomes a product feature

The launch also formalises something enterprises have been complaining about all year: safety classifiers that block legitimate work.

On the Frontier-Bench run, Opus 5's classifiers flagged 5% of API calls across 4% of trials. On Fable 5, the same classifiers fired on 42% of calls across 26% of trials. Anthropic expects the cyber classifiers to intervene roughly 85% less often than on its frontier model. Opus 5 is permitted to find vulnerabilities in source code, while binary scanning, penetration testing and exploit generation stay blocked.

More importantly for architects, flagged requests no longer just fail. A new automatic fallback option, in beta, routes blocked calls to another model rather than returning an error, alongside mid-conversation tool changes that don't invalidate the prompt cache. Opus 5 also ships with a 1M-token context window and no data retention requirements for general access, a line that matters in regulated industries.

On the security side, indirect prompt injection success within 15 attempts fell from 5.5% on the previous model to 2.0%. In browser environments, attack success dropped from 31.5% to 3.7% with no safeguards applied, and to 0% across all 129 environments with automatic protections on.

The company's internal behavioural audit scored the model 2.3 on misaligned behaviour, its lowest recorded figure.

The competitive backdrop

Anthropic enters this launch already leading where it is aiming. Anthropic enters this launch with growing momentum in the enterprise AI market. The company has increasingly focused on business customers, software development teams, and enterprise AI deployments rather than competing primarily for consumer chatbot users. Claude has become a popular choice for coding, knowledge work, and long-running AI agents, making Opus 5 a natural extension of that strategy.

Opus 5 is also the new default on the premium consumer tier and the strongest model available a tier below it. Most paying users now get a model that outscores what was, until Friday, the company's flagship offering.

What to watch

Two caveats deserve weight. First, nearly every number above is vendor-published, run on the vendor's harness; the independent evaluations that matter will land over the coming weeks. Second, benchmark efficiency and production efficiency are different animals: a 74-task suite is not a 40,000-request-per-day support pipeline.

The open question is behavioural rather than technical: whether enterprises are comfortable in a world where a safety classifier, not the buyer, sometimes decides which model answers a given request. Automatic fallbacks make that trade-off configurable. They don't make it disappear.