Anthropic's new flagship lets you pick how hard it thinks on every request. Left on high by default, it turns routine work into premium invoices.
Anthropic launched Claude Opus 5 on July 24 and made it the default on Claude Max the same week. Per Anthropic's announcement, it is the company's most capable model, carries a 1-million-token context window, and runs in the API as claude-opus-5. The leaderboard talk will sort itself out. The part that touches your invoice is the dial.
Opus 5 ships with a per-request effort setting: low, medium, or high. You decide how much reasoning the model spends on each call. Low runs faster and cheaper for routine work. High lets it grind longer on hard problems. A separate speed choice, standard or fast, sits on top, and fast doubles the rate.
| Mode | Input / output per million tokens | Speed |
| Standard | $5 / $25 | Baseline |
| Fast | $10 / $50 | About 2.5x faster |
Per launch coverage, standard pricing matches the old Opus 4.8 and runs at half the input price of Fable 5. The raw rate is friendlier than a new flagship usually is. The cost problem is behavioral. Teams wire the model in, leave effort on high and mode on fast because it feels safer, and turn a stack of routine summaries into a premium bill.
Most professional-services work does not need maximum reasoning at double the rate. Intake triage, first-draft summaries, routine classification, all low stakes and reversible. Save high effort for the calls that are novel, expensive, and hard to undo, the contract that actually gets signed, the analysis a client acts on.
That mapping does double duty. It controls cost and it controls risk.
| Task type | Effort setting | Human review |
| Routine, reversible, low stakes | Low, standard mode | Spot check |
| Novel, high stakes, hard to reverse | High | Every output, before it ships |
One honest caveat. Anthropic positions Opus 5 at the front on coding and knowledge work. It is not the top model in every category, so test it on your actual files before you standardize on it.
AI drafts, you verify. Opus 5 adds a second lever underneath that: you also decide how hard it drafts. Set the dial on purpose, task by task, and the flagship earns its keep instead of surprising you.
AI, Properly is the Ayvede field report.
Advise. Vet. Deliver.
Ask a question in writing