On July 24, 2026, Anthropic shipped Claude Opus 5 — and the story a founder should read isn't the benchmark chart. It's the price tag that didn't move. Opus 5 costs $5 per 1M input and $25 per 1M output, exactly what Opus 4.8 cost, and roughly half of Fable 5's input price — while matching or beating the larger, pricier Fable 5 on Anthropic's internal benchmarks (BankInfoSecurity; MarkTechPost).
That's the whole insight up front: the frontier tax collapsed again. The best Claude is now the everyday-priced Claude. If you run agents for a living, that changes your routing math, not just your changelog.
What actually shipped#
Opus 5 is the fourth Claude 5 model in under two months — after Mythos 5, Fable 5, and Sonnet 5. The specs, verified across launch coverage:
- Price: $5 / $25 per 1M tokens (standard). A fast mode at $10 / $50 runs about 2.5x faster for latency-sensitive work.
- Benchmarks: leads SWE-bench Verified at ~96% (ahead of Mythos 5 and Fable 5 near 95%), and 30.2% on ARC-AGI-3 — about 3x the next-closest model (Codersera).
- Context & control: a 1M-token window as default and maximum, extended thinking on by default, and a per-request low / medium / high effort toggle to trade reasoning compute for cost.
- Availability: the API, plus the new default on Claude Max and the strongest model on Claude Pro.
The number that matters is the one that stayed the same. Anthropic put frontier capability at last generation's Opus price.
Why "same price" is the headline#
For two years the mental model was a ladder: cheap tier for volume, mid tier for most work, and the flagship reserved for the hard 10% — because the flagship carried a 2–5x premium. That premium is what made "route cheap first, escalate on failure" the obvious default.
Opus 5 knocks a rung out of that ladder. When the best model sits at the everyday price, the arithmetic of down-routing inverts: the dollars you save by sending a task to a cheaper tier shrink, while the quality you give up grows, because the gap between "default" and "best" just closed to zero. We argued before launch that the Opus 5 decision would be about agent cost, not benchmarks — the pricing confirms it, and it's the same collapsing-frontier-tax pattern we saw when the cheap tiers caught up, running in the other direction.
What a founder does this week#
- If your default was Opus 4.8, switch — it's free. Same price, better model, drop-in. Opus 4.8 is superseded at an identical rate; keeping it as your default leaves capability on the table for no saving.
- If your default was a cheap tier picked to save money, re-run the unit economics. Measure cost-per-accepted-answer, not cost-per-token. When the flagship is this cheap, a workflow that retries, escalates, and post-processes cheap-tier output can quietly cost more than one Opus 5 call that lands it the first time. Our Opus-vs-cheap-tier routing eval is the harness to prove which way your traffic actually breaks.
- Keep a cheap tier for genuinely low-stakes, high-volume calls — classification, routing, extraction. That's still their job. The change is that your default for anything reasoning-shaped should probably be the flagship now.
- Reserve Fable 5 for where it earns the premium. It's still the larger model and may edge specific long-horizon jobs; let your own evals, not the marketing, decide. If you're weighing the top tiers head-to-head, we broke down Fable 5 vs Opus 4.8 vs the frontier ceiling, and the Kimi K3 vs Opus vs GPT-5.6 cost comparison covers the cross-vendor version of this exact decision.
The tell#
Ask what you're optimizing. If it's cost-per-token, cheap tiers still win and always will. If it's cost-per-good-outcome — which is what actually shows up in your margin — the leader may have changed on July 24. The frontier no longer charges rent. Price your agents like it.



