Claude Opus 5.5 Costs: The 60% Cache Cut Beats the 20% Price Cut
Opus 5.5 dropped to $4/$20 per million tokens and $0.20 cache reads. The same 2.8M-token task costs $11.20 or $0.99 depending on cache hit rate. Full pricing and benchmarks.
The 20% cut is not the story. The cache-read price is.
On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in its Claude 5.5 family. The headline is unglamorous: near-Fable 5.1 performance at roughly 40% less cost to run than Opus 5. Per-token prices dropped 20% to $4 input / $20 output per million tokens.
The more useful part of the announcement is the cost breakdown Anthropic published alongside it. Anthropic states plainly that cache reads make up the majority of agentic and coding work costs — and cut that line item from $0.50 to $0.20 per million tokens, a 60% reduction. That single number matters more to your invoice than the 20% headline cut.
The reason is mechanical. An agent resends the entire conversation on every turn: the same system prompt, the same file contents, the same instructions, dozens of times per session. A one-hour session against a million-token context can push millions of cumulative input tokens through the API. The per-token rate applies to that cumulative pile, and whether it bills at $4.00 or $0.20 depends entirely on cache hits.
The arithmetic: roughly 11x on the identical task
Anthropic's own figures are concrete enough to check by hand.
- A task using 2.8 million input tokens: $11.20 with no caching, $1.62 at a 90% cache hit rate, and about $0.99 at 96%
- 60,000 output tokens on a typical task: about $1.20
- Compacting a 150,000-token conversation: about $0.25
- Enterprise guidance: roughly $13 per developer per active day, with under $30 per active day covering 90% of users
The same task costs $11.20 or $0.99 depending on cache behavior. That is a ~11x spread, and no model swap closes it. A 20% unit-price cut buys you 20% — not 10x.
Reading Opus 5.5 as "tokens got cheaper" gets you half the picture. The fuller reading: Anthropic put numbers on the fact that how you structure caching determines your bill by an order of magnitude.
The spec changes matter as much as the price
The implementation deltas deserve attention before the pricing.
- Thinking is now always on. Opus 5 let you disable adaptive thinking at
higheffort or below. Opus 5.5 does not. - Default effort dropped from
hightomedium. - Forced tool calls are no longer supported via the API.
- Context window remains 1M tokens; standard max output remains 128K tokens — no change from Opus 5.
- Knowledge cutoff moved from May 2026 to June 2026.
Because thinking cannot be disabled, thinking tokens bill at the standard output rate of $20 per million. Anthropic's own example: setting effort to high adds roughly 20,000 thinking tokens to a task, costing about $0.40. Trivial per task; material across thousands of daily tasks if high becomes a thoughtless default. The shift of the default to medium is Anthropic pricing that risk in.
Full pricing, one place
- Standard: $4.00 input / $20.00 output per million tokens
- Fast mode: $8.00 / $40.00 — up to 2.5x speed in Claude Code and the Claude Platform
- Batch API: $2.00 / $10.00 — 50% off both directions, asynchronous, and it unlocks a 300,000-token output ceiling
- Cache read: $0.20
- Cache write (5-minute): $5.00, down from $6.25
- Cache write (1-hour): $8.00, down from $10.00
Fast mode costs exactly double the standard rate. It is a fair trade for latency-sensitive work — and a quiet reversal of the entire price cut if it drifts into being a default. Access runs through the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, plus Pro, Max, Team, and Enterprise subscription plans. The Free plan does not include Opus 5.5 — it stays on Sonnet and Haiku tiers.
Benchmarks: vendor numbers and third-party numbers diverge
Anthropic's reported results at max/xhigh effort:
- Terminal-Bench 4.0 (agentic coding): 66.4%, up from 52.3% on Opus 5
- AutomationBench (business workflows): 40.0%
- FrontierCode v1.1: 54.4%
- CursorBench 4.0 (IDE coding): 57.8%
- GDPval-AA v2.1 (Elo): 1846
- Humanity's Last Exam (with tools): 67.7%
- Terminal-Bench-Science: 58.7%
- OSWorld 2.0 (desktop use, partial credit): 81.8%
Independent testing from Artificial Analysis reports lower figures on at least two — 59.6% on Terminal-Bench 4.0 and 61.4% on Humanity's Last Exam. This gap between vendor-reported and third-party results is routine, not a red flag, but it is the number to quote when someone asks what the model actually scores.
Anthropic also notes it found benchmark margins to be a less reliable guide at this capability level, and that in its own use the distance between Opus 5.5 and Fable 5.1 is narrower than the scores imply. That is a rare and useful admission.
What to configure first
The order of operations is not ambiguous.
high.One more caveat on the numbers: a ten-turn retry loop with 100,000 tokens of cached context and 10,000 output tokens costs roughly the same as a single well-cached turn. Cache hit rate dominates turn count — which is the whole argument in one sentence.
Caveats on the performance claims
The anecdotes — a 680,000-line code migration in under a day, a 200,000-line codebase audit in under three hours (versus more than 20 hours on Opus 5), and a HAProxy C-to-Rust translation in 9.5 hours — are vendor and early-tester reports, not independently reproduced results. Treat them as directional and re-measure on your own workload. Anthropic also shipped Opus 5.5 with preserved thinking, an anti-distillation safeguard, applying to API accounts created on or after August 31, 2026, which is a compatibility consideration if you manipulate prior context.
The bottom line
Claude Opus 5.5 cuts input and output by 20% and cache reads by 60%. What actually moves your bill is neither: it is your cache hit rate and your effort setting. The same 2.8-million-token task costs $11.20 or $0.99. If you run agents, the first number worth measuring is not on Anthropic's pricing page — it is in your own logs.