Claude Opus 5 Is Anthropic Undercutting Itself, on Purpose

Anthropic shipped Claude Opus 5 on July 24 — its fourth model release in under two months, after Mythos 5, Fable 5, and Sonnet 5 landed in June. The pitch is blunt: near-Fable-5 intelligence at exactly the price you were already paying for Opus 4.8, $5 per million input tokens and $25 per million output. That's half of Fable 5's input price for a model that, on Anthropic's own agentic-coding numbers, lands within 0.5% of its flagship.

Read past the benchmark charts and this is the more interesting story: Anthropic is deliberately cannibalizing its own premium tier. That's not generosity. It's a bet about where the money in this market actually is.

The price curve is the product

Anthropic's model line now reads like a ladder rather than a flagship-plus-discounts lineup: Mythos 5 restricted to approved organizations, Fable 5 as the premium frontier, and Opus — for years the top of the range — repositioned as the value-frontier workhorse. Opus 5 is where that repositioning gets teeth.

The reason is agents. A chat session cares about capability per response; an agent harness cares about capability per million tokens, because a single long-running coding or computer-use task can chew through tens of millions of them across tool calls, retries, and context reloads. At that volume, a model that's 99.5% as good at 50% of the price isn't a compromise — it's the obvious default. Anthropic's own comparison makes the economics explicit: on OSWorld 2.0, the computer-use benchmark, Opus 5 beats Fable 5's best score at one-third the cost per task.

The launch also confirms that inference economics is becoming an API surface rather than a SKU decision. Opus 5 ships with an effort setting that trades capability against cost per request, plus a fast mode at double the price that runs roughly 2.5× faster. You're no longer choosing a model tier once per project; you're tuning spend per call. Expect your agent framework's config surface to grow accordingly.

What the numbers actually say

The vendor-reported results are strong even by launch-day standards: more than double Opus 4.8 on Frontier-Bench's agentic coding suite, within 0.5% of Fable 5 on CursorBench 3.2 at max effort, roughly 1.5× the next-best model on Zapier's AutomationBench, and 30.2% on ARC-AGI-3 — about three times the nearest competitor. Treat those with the usual launch-day skepticism, but the independent signal points the same way: Simon Willison notes it's currently leading the Artificial Analysis leaderboard, ahead of Fable 5 itself.

The result that will get quoted for years is the math one. Per the system card, Opus 5 solved all six IMO 2026 problems — 42/42, gold-medal territory, where the cutoff was 29 — with no tools and no agent harness, graded by both a model judge panel and human experts. Two years ago that required DeepMind's purpose-built AlphaProof running formal proof search. Now it's a checkpoint on a general-purpose API.

The behavioral shift matters as much as the scores. Willison highlights a telling example: asked about a machine-part drawing it couldn't directly view, Opus 5 wrote its own computer-vision pipeline to extract the geometry from raw pixels and reconstructed the part. Anthropic calls this "proactive." If you run agent harnesses, read that word carefully — a model that invents multi-step workarounds instead of reporting failure is exactly what you want in a coding agent and exactly what blows token budgets and permission boundaries in a poorly fenced one. When you swap it in, re-check your max-iteration limits and tool allowlists before trusting old assumptions.

Adoption is a config change — the evaluation isn't

The migration path is about as low-friction as this industry gets: the model ID is claude-opus-5 on the Claude API, and it's live in the Model Garden on Google Cloud's Agent Platform (the platform formerly known as Vertex AI), with traffic staying inside your private GCP network, IAM-scoped access, and Zero Data Retention compatibility for the compliance-constrained. If you're on Fable 5 for agentic coding today, the experiment costs an afternoon: point your eval suite at the new ID, compare pass rates, and look hard at cost per completed task rather than cost per token — effort settings make per-token math misleading.

One under-reported change developers will feel day-to-day: Anthropic says its safety classifiers intervene about 85% less often than Fable 5's on cybersecurity-adjacent work. Anyone who's had a legitimate dependency audit or CTF writeup tripped by over-eager refusals knows how much friction that removes. The line Anthropic is drawing instead is architectural: Opus 5 got better at finding vulnerabilities but was deliberately not trained on exploitation, which stays gated behind Mythos 5 and a verification program for security researchers. Whether that find/exploit split holds up under fine-grained probing is an open question, but it's a more honest control point than refusal-classifier whack-a-mole.

Genuine shift, with an asterisk

Call it what it is: this is a real shift, not launch hype — but the shift is economic, not scientific. Opus 5 doesn't do much that Fable 5 couldn't; it does it at a price where "just use the near-frontier model for everything" becomes a defensible architecture instead of a budget request. The losers are premium-tier pricing everywhere — Fable 5's own positioning now depends on a shrinking set of tasks where that last half-percent matters — and every competitor whose mid-tier now benchmarks below a model costing the same or less.

The asterisk is that most of the eye-catching deltas are Anthropic's own measurements, and independent replication of the agentic benchmarks will take weeks. Don't re-platform on a press release. But do run the eval — at these prices, not benchmarking it against your current stack is the expensive choice.