Claude Opus 5 Is Here. It's Half the Price of Fable 5.

What Happened

Anthropic released Claude Opus 5 on July 24, 2026. It costs $5 per million input tokens and $25 per million output tokens, which is exactly what the model it replaces, Opus 4.8, already cost.

That flat price is the entire story. Anthropic’s most capable widely released model, Claude Fable 5, runs $10 per million input tokens and $50 per million output tokens. Opus 5 is half that, and on Anthropic’s own benchmark charts it lands close enough to Fable 5 that the gap is measured in fractions of a percentage point.

For anyone paying a subscription rather than an Application Programming Interface (API) bill, the change already happened without asking. Opus 5 is “the new default model on Claude Max, and the strongest model on Claude Pro.”

Key Details

• Price: $5 per million input tokens, $25 per million output tokens, identical to Opus 4.8.
• Context window: 1 million tokens, with a 128,000-token maximum output.
• Model ID: claude-opus-5 on the Claude API.
• Fast mode: available at twice Opus 5’s base price.

The Dial That Sets the Price

The headline feature is not a capability. It is a knob.

Opus 5 ships with a feature “enabling users to toggle how much effort … the model expends completing a task or answering a prompt,” with low, medium, and high settings. Anthropic describes it as a control “customers can use to optimize for intelligence or conserve tokens for faster and cheaper results.”

This matters more than it sounds. Tokens are the unit of billing, so a setting that changes how many tokens a model burns is a setting that changes what a task costs. The benchmark results are quoted at specific effort levels for the same reason: on CursorBench 3.2, Opus 5 lands within half a percentage point of Fable 5’s best score at roughly half the cost per task, but only at maximum effort. Run it lower and both the score and the bill come down.

On Frontier-Bench v0.1, Anthropic reports that Opus 5 more than doubles Opus 4.8’s score while costing less per task. Treat that as a vendor figure. Anthropic published its benchmark results as chart images with no underlying numbers, so the scores now in circulation have been read off a graph rather than pulled from a table. Independent replication has not happened yet.

Where Anthropic Says It Falls Short

The more interesting admission sits in the safety section, and it is easy to misread.

Anthropic states that “Opus 5 does not advance the frontier in risky, dual-use capabilities” and that it “remains behind Mythos 5 in both biology research and offensive cybersecurity.” Read that carefully, because the comparison is not the one you would expect. Mythos 5 is not the flagship. It is a separate model offered for defensive cybersecurity workflows under Anthropic’s invitation-only Project Glasswing, with no self-serve signup.

That distinction is the whole point. Anthropic benchmarked its mass-market model’s hacking ability against its specialist security model, not against its flagship. On OSS-Fuzz, Anthropic’s own vulnerability benchmark, the published charts show Opus 5 matching Mythos 5 closely on finding vulnerabilities while scoring far lower on building working exploits. It can find the bug. It is markedly worse at weaponizing it. Mythos Preview, the earlier model in that line, is the one this site covered finding 27 years of hidden bugs in three weeks.

Not everyone reads that restraint as a virtue. Hugging Face chief executive Clem Delangue argued that “closed model APIs have guardrails that flag and refuse a lot of legitimate security work, because analyzing an attack looks a lot like preparing one.”

And Opus 5 is not the top of the stack. Anthropic “still recommends Fable 5 for more advanced projects, including those which the model may handle autonomously for days on end.” Fable 5, the model a US export control directive forced Anthropic to disable for every customer in June before it came back on July 1 and shifted to metered usage credits on July 12, keeps its job.

The Read

Holding price flat across a generation is a price cut that does not look like one. Anthropic did not lower the number on Opus 4.8; it raised what the number buys. For buyers who budget in dollars per task rather than dollars per million tokens, that is the only comparison that matters, and it landed while the effort dial gives them a second lever to push costs down further.

The catch is that every performance claim above traces back to Anthropic’s own evaluations, several of them read off unlabeled charts. Anyone can verify the pricing. Nobody outside the company has yet verified the intelligence.

The Bottom Line

Anthropic left the sticker price alone and moved the capability instead. If you pay for Claude Max or Pro, you are already using it. If you pay per token, the effort setting is the lever that now decides your bill. And if the work involves anything adjacent to offensive security, Anthropic has told you plainly that this is not the model for it.

Sources (5)

• anthropic.com Introducing Claude Opus 5
• platform.claude.com Anthropic Docs: Models Overview
• fortune.com Anthropic debuts Claude Opus 5
• officechai.com Claude Opus 5 Benchmarks
• trendytechtribe.com Fable 5 Back in Your Plan