Anthropic Points Opus 5 at Its Own Flagship

The update

On July 24, 2026, Anthropic launched Claude Opus 5, the new flagship of its Opus tier, available on the Claude API, Claude.ai, Claude Code, and Claude Cowork. Pricing holds at $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8 — with an optional Fast mode running about 2.5x faster at 2x the base price. Anthropic positions it as delivering close to the intelligence of its top model, Claude Fable 5, at roughly half the cost.

By the numbers

Under the hood

The benchmark claims are aggressive. On Frontier-Bench v0.1, Anthropic says Opus 5 tops all models and roughly doubles Opus 4.8's score at lower cost. On CursorBench 3.2 it lands within 0.5% of Fable 5's peak at half the cost per task; on ARC-AGI 3 it claims about 3x the next-best model; on Zapier AutomationBench it passes roughly 1.5x more often than the next-best at the same cost — and at its lowest effort setting still beats every competitor. On OSWorld 2.0 it says it beats Fable 5 at a third of the cost. Life-sciences gains over Opus 4.8: 10.2 points on organic chemistry, 7.7 on protein prediction. The capability story leans on long-running agentic workflows that verify their own work and iterate until a task succeeds, plus visual output generation. One demo stands out: given a drawing of a machine part it couldn't view directly, the model wrote its own computer-vision pipeline to extract the geometry and rebuild it as a 3D model. Partners echoed the pitch — Devin's CEO said it 'approaches Fable-level performance at half the cost,' Cursor's co-founder called it 'near Fable 5 intelligence at Opus speed and cost,' and Zapier's CEO reported a 100% pass rate on a churn-prevention workflow that earlier models failed.

The signal

The interesting move here isn't a higher ceiling — it's Anthropic undercutting its own frontier model with a cheaper tier. When your mid-flagship lands within 0.5% of the top model on CursorBench at half the cost per task, the top model's pricing starts to look like a tax on marginal quality. For anyone running agents at volume, where output tokens pile up fast, that spread is the whole decision, and it reframes the competition from 'who's smartest' to 'who's smart enough per dollar.'

The backstory

Opus 5 succeeds Opus 4.8 and slots beneath Fable 5, Anthropic's top model — so the tier structure now has a cheaper 'good enough' option deliberately measured against the flagship above it. Keeping the price identical to Opus 4.8 while roughly doubling the headline benchmark score is the through-line: same sticker, more model.

The timeline

1. Opus 4.8
   The prior Opus flagship, priced at $5/$25 per M tokens.
2. Fable 5
   Anthropic's current top model, holding the frontier.
3. Jul 24, 2026 — Opus 5
   Lands within 0.5% of Fable 5 on CursorBench at half the cost per task.

Who it's for

Developers

→Near-Fable coding at half the cost per task

Enterprises

→Cheaper agentic workflows that self-verify

Researchers

→Gains in organic chemistry, protein prediction

The catch

Every number here comes from Anthropic's own announcement and partner quotes, on benchmarks like Frontier-Bench v0.1 and CursorBench 3.2 that are still young — treat them as vendor claims until independent runs land. On safety, Anthropic reports a 2.3 behavioral-audit score on misaligned behavior (its lowest recent) and says classifiers intervene about 85% less than with Fable 5, but it also notes the model stays substantially behind its 'Mythos 5' cyber threshold with no cyber training applied, and that biology safeguards match Opus 4.8 with acknowledged limits on long-running autonomous research. The self-verifying agentic loops are the most impressive and least externally tested part.

What to watch

Watch for independent benchmark reproductions, especially on the agentic and OSWorld claims where self-verification is doing heavy lifting. The open question is what this pricing pressure does to Fable 5's positioning — if Opus 5 really lands within a fraction of the flagship at half the cost, the flagship needs a reason to exist beyond the last few points of capability.