Claude Opus 5 Is Nearly as Capable as Anthropic’s Best Model at Half the Price

The most consequential part of a new AI model is not always a higher benchmark score. Sometimes it is the moment when expensive capability becomes affordable enough to use every day. Anthropic says Claude Opus 5 reaches close to its more powerful Fable 5 model on many coding and knowledge tasks while costing about half as much.

KEY NUMBER
Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8 and roughly half Anthropic’s price for Fable 5.

Why the price matters more than another leaderboard

Powerful AI agents can consume large amounts of text, code and tool calls while they work through a task. That makes the cost per token only part of the bill. A model that solves a task with fewer attempts, less generated text or fewer tool calls can be cheaper in practice even when its published token price does not change.

Anthropic says Opus 5 more than doubled the performance of Opus 4.8 on Frontier-Bench while reducing the cost per completed task. It also says the model came within 0.5 percentage points of Fable 5 at maximum effort on CursorBench, at about half the cost per task. These are strong claims, but they should be read as launch evidence, not a final verdict.

The practical question for a software team is simple: can the model finish a difficult job reliably enough that a person spends less time supervising and repairing its output? If the answer is yes, the economics of AI-assisted programming change. The value comes from completed work, not from generating more code.

What Opus 5 is designed to do

Anthropic describes Opus 5 as an everyday model for long, multi-step work. It is available through Claude products and the company’s application programming interface, where developers can connect it to coding tools, documents and business systems.

The company reports improvements in debugging, root-cause analysis, computer use, professional research and scientific tasks. It also offers an adjustable effort setting, allowing users to trade speed and cost for deeper processing. A separate Fast mode runs about 2.5 times faster but costs twice the base rate.

That combination points to a wider shift in AI. The competition is moving beyond who owns the single smartest model. Labs are now competing over usable intelligence per dollar, latency, reliability and the degree to which a model can continue working without losing track of the goal.

More useful access, with carefully drawn safety limits

Opus 5 arrives during a sensitive debate over models that can find and exploit software vulnerabilities. Anthropic says the model is close to its more powerful Mythos 5 system at finding vulnerabilities, but remains substantially worse at turning them into working exploits.

Its safeguards therefore allow some defensive source-code analysis while blocking higher-risk activities such as penetration testing, binary vulnerability scanning and exploit generation for ordinary users. Verified cybersecurity researchers may obtain less restrictive access through a separate programme.

Anthropic also says its automated behavioural audit gave Opus 5 the lowest rate of misaligned behaviour among its recent models. That is useful information, but it is still an evaluation designed and reported by the model’s creator. Real deployments can expose failure modes that controlled tests do not capture.

What businesses should test before switching

A benchmark winner is not automatically the best model for every organisation. Teams should test Opus 5 on their own codebases, documents and approval processes, measuring completed-task cost, error rate, review time and the consequences of a wrong action.

For sensitive work, access controls and human approval remain essential. A model that can operate software or edit production systems has a larger blast radius than a chatbot that only drafts text. Companies should also examine data-retention terms, regional availability and whether automatic fallback sends a declined request to another model with different behaviour.

What happens next

The most revealing evidence will come from independent testing and production use over the next several weeks. Developers will want to know whether Opus 5 maintains its performance across different coding agents, languages and long-running tasks, and whether its lower cost survives real workflow overhead.

If the claims hold, Opus 5 may matter less as a spectacular new intelligence record and more as an economic milestone. Capability that once sat behind premium pricing is moving into the everyday tier, which could accelerate the adoption of AI agents across software, research and routine office work.