Claude Opus 5 lands at half the price of Fable 5

The frontier tier just got cheaper without anybody switching vendors. Anthropic launched Claude Opus 5 today at $5 per million input tokens and $25 per million output — the same price as Opus 4.8, and half what Fable 5 costs. If you priced an AI workload against Fable 5 in June, that spreadsheet is now wrong. So is the assumption that you know which model answered your last API call.

What actually happened

Opus 5 is available now on the Claude API as claude-opus-5, plus Claude.ai, Claude Code, and Claude Cowork. It becomes the default model on Claude Max and the strongest model available on Pro. CNBC frames it as near-Fable-5 capability at half the cost; Fable 5, from June, runs $10/$50 per million.

Anthropic's own benchmark claims are the interesting part because they're framed in cost, not just score: within 0.5% of Fable 5 on CursorBench 3.2 at half the price, ahead of Fable 5 on OSWorld 2.0 at a third the price, roughly 2x Opus 4.8 on Frontier-Bench v0.1. Benchmarks are vendor-run — treat them as a reason to test, not a result.

Two operational details matter more than any of that. First, Automatic Fallbacks: requests flagged by safety classifiers now "automatically route to another model" instead of erroring, so the API "always route[s] to the best available model by default rather than being blocked." Anthropic also expects those classifiers to "intervene around 85% less often than they do for Fable 5." Second, retention: consistent with prior Opus models, Opus 5 ships with no data retention requirement for general access — the constraint that applies to Fable and Mythos.

Why cheaper AI models don't lower your bill

We've watched this pattern enough times to name it: token prices fall, token consumption rises to fill the gap, and the invoice stays flat. A cheaper frontier model doesn't cut your spend. It changes which workloads clear the bar — the eval you skipped because it cost $400 a month is now $200, and the agent loop you capped at three iterations can run six.

So do the math on cost per completed task, not per million tokens. A model that's 30% cheaper but needs two extra retries is more expensive.

Then handle the two changes that happen whether you act or not. Your default moved — Max users are on Opus 5 now without opting in. Pin explicit model IDs in production code and treat a version bump as a deploy, not a shrug. And Automatic Fallbacks is genuinely useful and means the model that answers may not be the one you requested. If you're generating client-facing output or anything you'd have to defend later, log the responding model on every call. "The AI wrote it" is not an audit trail.

Sources: Anthropic, CNBC, TechCrunch.