DeskTheory is where founder-CEOs learn to run their companies on AI leverage.

What is Claude Opus 5?

Claude Opus 5 is Anthropic's workhorse flagship, released July 24, 2026. It comes close to the frontier intelligence of Fable 5 at half the price, $5/$25 per million tokens, and it's the first Claude with an effort dial: low, medium, or high, chosen per task.

Anthropic shipped four models in under two months this summer: Mythos 5, Fable 5, and Sonnet 5 in June, then Opus 5 on July 24. If the names blur together, here's the one that matters day to day: Opus 5 is the model most of your Claude work actually runs on now. It's the default for Max subscribers and the strongest model you can select on Pro.

What it is (in plain English)

The Claude family is a ladder. At the top sits Mythos 5, the version Anthropic restricts to approved organizations. Fable 5 is its public face, the frontier model you pay premium rates for. Sonnet 5 is the fast, cheap workhorse at $2/$10. Opus 5 slots in between: near-Fable capability at $5/$25, half of Fable's price and the same base price as its predecessor, Opus 4.8.

The practical specs: a 1M-token context window, up to 128K tokens of output, API name claude-opus-5. It's live in the Claude apps, Claude Code, the API, and the big clouds (Bedrock, Vertex, Microsoft Foundry), plus GitHub Copilot.

What changed beyond the number: Anthropic built this one to need less supervision. It requires less back-and-forth than its predecessors, and it "verifies its own work and recovers from errors without intervention." It's also, by Anthropic's account, the most aligned Opus yet and the "least susceptible to being tricked into misuse," with stronger cybersecurity guardrails than 4.8.

The effort dial is the actual news

Opus 5 ships with a toggle: low, medium, or high effort, set per task. Low effort answers fast and cheap; high effort thinks longer and costs more. You choose how hard the model works based on what the task deserves.

Pricing an AI used to be one number; now it's a dial you turn per job. This is the same shift OpenAI made from the other direction in August, selling speed tiers for Sol. The labs have stopped selling "a model" and started selling a range you position yourself on. Reformatting a document doesn't deserve the same effort as reasoning through your pricing strategy, and now the bill reflects that.

Why you should care as a CEO

If your team is on Claude, this model is already answering you; knowing what it is stops being trivia the moment you budget for AI or pick a model in Claude Code. The operator playbook that falls out of the ladder is simple: run the week on Opus 5, drop routine bulk work to Sonnet 5, and reserve Fable 5 for the genuinely hard runs. Anthropic's own guidance points the same way; it recommends Fable for the most advanced projects, the kind a model handles autonomously for days.

The half-price part deserves a beat. A year ago, frontier-adjacent intelligence was the expensive tier by definition. Opus 5 at $5/$25, Sonnet 5's cancelled price hike, and OpenAI's Sol cut all landed within weeks of each other. The verdicts you formed about what AI "costs" in the spring are due for a rerun.

Where you'll see it

• In Claude Code's model picker, where /model shows you which Claude you're actually talking to.
• In the Claude apps, as the default on Max plans and the top selectable model on Pro.
• Next to its siblings: What is Claude Mythos?, our Fable 5 coverage, and the Opus 4.8 launch it supersedes.
• In the Frontier Watch hub and our ChatGPT vs Claude vs Gemini comparison.

What you should do next

Open Claude Code, run /model to see which Claude you're on, then run the same real task once at low effort and once at high and compare what the difference actually buys you.