Claude Opus 5 Reprices the Frontier

Anthropic shipped Claude Opus 5 this week, and the headline number isn't a benchmark — it's the invoice. The model lands at $5 per million input tokens and $25 per million output, identical to Opus 4.8, while Anthropic claims it gets within 0.5% of its top-tier Fable 5 on CursorBench 3.2 at half the cost and actually beats Fable 5 on OSWorld 2.0 at a third of the spend. Independent coverage from VentureBeat and others confirms the positioning: near-frontier intelligence, workhorse pricing.

That's the story worth paying attention to, because it changes the default answer to "which model do I build on?" For the past year that answer required a spreadsheet: Fable 5 for the hardest problems, Opus for production, Sonnet for volume. Opus 5 collapses most of that decision. Unless you're running the kind of multi-hour autonomous research tasks Fable 5 was built for — and can stomach its $10/$50 pricing and 30-day data retention requirement, which Opus 5 doesn't carry — Opus 5 is now the sensible center of gravity. Fable 5 keeps the crown; Opus 5 takes the traffic.

A drop-in that isn't quite a drop-in

Anthropic markets this as a drop-in upgrade, and mostly it is: same feature set as Opus 4.8, same 1M-token context window, 128K output, the full low-through-max effort ladder. But two API changes will bite anyone who swaps the model string and walks away.

First, thinking is now on by default. On Opus 4.8, omitting the thinking parameter meant no thinking. On Opus 5, the same request runs adaptive thinking — and max_tokens is a hard cap on thinking plus response text. A route that sized max_tokens tightly around its expected answer can now truncate mid-response with no error, just a stop_reason: "max_tokens" you weren't checking for. Audit every call site that never set thinking.

Second, you can no longer disable thinking at high effort. thinking: {type: "disabled"} combined with effort: "xhigh" or "max" returns a 400, validated per request. Teams that ran thinking-off at max effort for latency-sensitive routes need to pick a lane — and honestly, the right lane is usually thinking-on at medium effort, which on this model performs surprisingly close to prior models' top settings at a fraction of the tokens. Run the effort sweep before you assume your old defaults transfer.

One more operational gotcha: Opus 5 gets its own rate-limit bucket, separate from the combined Opus 4.x pool. Shifting traffic over neither frees headroom on the old bucket nor inherits it. Check your tier's limits before flipping the switch on production volume.

The interesting bits are in the betas

Two new beta features say more about where Anthropic is heading than any benchmark.

Mid-conversation tool changes (mid-conversation-tool-changes-2026-07-01) let you add or remove tools between turns without invalidating the prompt cache. Until now, tools rendered at position zero of the prompt, so any change re-billed your entire cached prefix — a real tax on long-running agents that need mode switches or dynamically granted capabilities. You declare tools up front with defer_loading: true, then surface them via tool_addition blocks in a system message. Combined with the prompt-cache minimum dropping from 1024 to 512 tokens, the pattern is clear: Anthropic is optimizing the economics of long-lived, cached agent conversations, where the conversation is the durable object and everything else mutates around it.

Automatic fallbacks (fallbacks: "default", behind the server-side-fallback-2026-07-01 header) address a genuine production problem. Opus 5 ships with elevated cybersecurity safeguards, and its classifiers can decline a request — you get an HTTP 200 with stop_reason: "refusal", which will break any code that reads content[0] unconditionally. The fallback parameter reruns declined requests on another model server-side, routed by refusal category (cyber refusals go to Opus 4.8, which carries lighter classifiers). If you do anything security-adjacent — pentesting tools, malware analysis, even vuln-scanner output triage — opt in from day one. False positives on benign security work are a known failure mode of this classifier generation, and a fallback turns a hard failure into a silent recovery.

The migration cost is in your prompts, not your code

The subtler tax is behavioral. Per Anthropic's own migration guidance, Opus 5 verifies its own work unprompted — which means the "double-check your answer" and verify-with-a-subagent scaffolding everyone bolted on over the past two years now causes over-verification and wasted tokens. The fix is deletion, not rewriting. Same story with delegation: where Opus 4.8 under-used subagents and needed encouragement, Opus 5 reaches for them freely, multiplying cost. If you added "delegate more" prompting for 4.8, rip it out and add a spawn cap instead. And it writes longer — both responses and files on disk — so conciseness instructions earn their place.

This is the part teams consistently underestimate: a model swap is an afternoon; unwinding prompt scaffolding tuned to the previous model's weaknesses is the real week of work. Budget for it.

Who wins, who loses

Winners are obvious: agent platforms (Cursor, Devin, Zapier, and Lovable all shipped glowing launch quotes) and anyone whose unit economics were gated on frontier-class reasoning. Fast mode at $10/$50 — Claude API only, not Bedrock or Vertex — gives interactive products a 2.5x speed lever that undercuts Fable 5's base price.

The squeezed party is Anthropic's own top tier. When your $5 model lands within half a percent of your $10 model on the benchmarks customers actually cite, Fable 5's addressable market shrinks to genuinely frontier workloads. That's likely intentional — reportedly IPO-bound, Anthropic needs a high-volume workhorse more than it needs flagship exclusivity. But it's also the competitive dynamic every frontier lab now faces: your best model's biggest rival is your own second-best model, six months later, at half the price. Opus 5 is the cleanest example yet of that curve — and this time, it's the one to build on.