Claude Opus 5: near-Fable-5 coding at roughly half the price
The API details the daily drop skipped: 1M context, effort up to max, mid-session tool swaps, a 512-token cache floor, and two breaking migration changes.
Copy markdown
Frontier-tier coding, Opus pricing
On CursorBench 3.2 it lands within 0.5% of Anthropic's frontier Fable 5 while charging $5/$25 per Mtok — about half Fable 5's rate. It also doubles Opus 4.8 on Frontier-Bench, triples the next model on ARC-AGI 3, and beats Fable 5 on OSWorld 2.0 at a third of the cost.
1M context standard, thinking on by default
Every request gets the full 1M-token window and 128k max output, with extended thinking now enabled by default (model id `claude-opus-5`). Practical catch: raise your `max_tokens`, because reasoning tokens now eat into your budget.
Two breaking changes before you migrate
Thinking is on by default, and you can no longer disable it at `xhigh` or `max` effort — you'll get a 400. Effort now climbs low → medium → high → xhigh → max, with `high` the default. Swap `claude-opus-4-8` for `claude-opus-5` and check both before shipping.
Swap tools mid-session without dumping your cache
A new beta (header `mid-conversation-tool-changes-2026-07-01`) lets you add or remove tools between turns while keeping the prompt cache warm — no more pinning a fixed tool list for a whole agent session.
Server-side fallbacks and a lower cache floor
A `fallbacks: "default"` beta (header `server-side-fallback-2026-07-01`) auto-routes safety-refused requests to Anthropic-recommended models by category, and the prompt-cache minimum drops to 512 tokens from 1,024, so shorter prompts cache for free.
Fast mode is API-only and costs double
A research-preview fast mode runs ~2.5x quicker for $10/$50 per Mtok, but only on the Claude API — not Bedrock (`anthropic.claude-opus-5`), Vertex, or Foundry, which serve the model at standard speed.