Claude Opus 5.5 Pricing: The Full Rate Card, the Cache Line, and What a Task Really Costs

Claude Opus 5.5 lists at $4.00 per million input tokens and $20.00 per million output, a flat 20% cut from the $5.00/$25.00 that Claude Opus 5 has carried, with cache reads down 60% to $0.20 per million. Those are Ant​hropic's published prices, read from Ant​hropic's own pricing documentation on September 24, 2026, two days after the model shipped. The sticker is the easy part. Ant​hropic also says the model costs "about 40% less to run" on typical workloads, and that number is not on the price sheet — it is a vendor characterisation built from the per-token cut plus an assumption that the model spends fewer tokens and fewer turns finishing the same job. This page separates the two: every rate you can verify on the sheet today, and the arithmetic behind the one you cannot. Claude Fable 5.1, the $10/$50 tier above it, appears throughout as the comparison Ant​hropic itself makes.

The full rate card, every line as Anthropic publishes it

Everything in this section was read from Ant​hropic's pricing documentation at platform.claude.com and claude.com on September 24, 2026. Nothing here is carried over from another write-up.

• Input — $4.00 per million tokens (Ant​hropic's published rate).

• Output — $20.00 per million tokens. Thinking tokens bill as output tokens, and Claude Opus 5.5's adaptive thinking is always on and cannot be disabled.

• Cache read — $0.20 per million tokens. Ant​hropic documents this as 0.05x the base input price, its own footnote carve-out for this model.

• 5-minute cache write — $5.00 per million tokens, the standard 1.25x base input multiplier Ant​hropic applies to every model on the sheet.

• 1-hour cache write — $8.00 per million tokens, the standard 2x base input multiplier.

• Batch API — 50% off both directions: $2.00 input and $10.00 output per million. Ant​hropic publishes the discounted figures directly rather than leaving you to halve them.

• Fast mode — $8.00 input and $40.00 output per million, exactly 2x standard. Ant​hropic's documentation titles the page "Fast mode (research preview)".

• US-only inference — a 1.1x multiplier on input, output, cache writes and cache reads when inference_geo: "us" is set. Global routing, the default, uses standard pricing.

• Context and output ceiling — 1M tokens of context and 128K of synchronous output at standard pricing, with no long-context surcharge. Ant​hropic states that a 900k-token request bills at the same per-token rate as a 9k-token one. On the Batch API the output ceiling rises to 300K behind the output-300k-2026-03-24 beta header.

Ant​hropic also states that the caching multipliers "stack with other pricing modifiers such as the Batch API discount and data residency" — so a batched, cached, US-pinned request multiplies all three, and the arithmetic compounds rather than replacing.

Here is what the two models readers actually weigh cost on the same sheet:

• Claude Opus 5 — $5.00 input, $25.00 output, $0.50 cache read, $6.25 for a 5-minute cache write, $10.00 for a 1-hour write, $2.50/$12.50 on Batch (Ant​hropic's published rates, read 2026-09-24).

• Claude Fable 5.1 — $10.00 input, $50.00 output, $0.25 cache read, $12.50 for a 5-minute cache write, $20.00 for a 1-hour write, $5.00/$25.00 on Batch (Ant​hropic's published rates, read 2026-09-24).

The cache line is where an agent bill is actually decided

Launch coverage treats the cache read as a footnote. It is the largest single lever on the sheet for any workload that re-sends a large prefix, which is most coding and agent workloads.

Ant​hropic's own footnote is unusually explicit about the ratios: cache hits and refreshes cost 0.1x the base input price on most Claude models, 0.025x on Claude Fable 5.1, and 0.05x on Claude Opus 5.5. So Claude Opus 5.5's cache ratio is twice as bad as the standard one, and twice as bad as Claude Fable 5.1's.

In dollars it wins anyway, and this is the part worth internalising: a better multiplier does not win on dollars when the base rate is two and a half times higher.

• Claude Fable 5.1 — 2.5% of a $10.00 base is $0.25 per million cached tokens.

• Claude Opus 5.5 — 5% of a $4.00 base is $0.20 per million cached tokens.

• Claude Opus 5 — 10% of a $5.00 base is $0.50 per million cached tokens.

The same inversion holds on writes. Claude Fable 5.1 charges $12.50 to write a million tokens into the 5-minute cache and $20.00 for the 1-hour window; Claude Opus 5.5 charges $5.00 and $8.00. Cache reads are the one line where Claude Opus 5.5 beats both siblings on absolute dollars despite having the middle ratio.

Two rules decide whether the write is worth paying for at all. Ant​hropic's documentation says a 5-minute write costs 1.25x base input and therefore pays off after a single cache read, while a 1-hour write costs 2x base input and needs two reads to pay off. And the cache is tied to the prefix that produced it: Ant​hropic notes that switching between fast and standard speed invalidates the cache, because requests at different speeds do not share cached prefixes. Changing the model, the effort level, or the tool definitions has the same effect. A cheaper model that forces a prefix rewrite on every switch can cost more than the expensive one it replaced.

What a long agent loop actually costs

Take a 40-turn coding agent that re-sends a 120,000-token prefix — system prompt, tool schemas, file tree — on every turn, adds 1,000 fresh input tokens per turn from tool results, and produces 800 output tokens per turn. Substitution is the point here: change any of those four numbers to match your own traffic and the arithmetic below still holds.

On Claude Opus 5.5:

• Cache write, once — 120,000 tokens at $5.00 per million = $0.60

• Cache reads, 39 turns — 4,680,000 tokens at $0.20 per million = $0.94

• Fresh input, 40 turns — 40,000 tokens at $4.00 per million = $0.16

• Output, 40 turns — 32,000 tokens at $20.00 per million = $0.64

• Session total — $2.34, of which cache reads are about 40%.

On Claude Opus 5, identical token counts: $0.75 of cache write, $2.34 of cache reads, $0.20 of fresh input and $0.80 of output — $4.09.

On Claude Fable 5.1, identical token counts: $1.50 of cache write, $1.17 of cache reads, $0.40 of fresh input and $1.60 of output — $4.67.

Same tokens, same turns, no efficiency claim required: Claude Opus 5.5 comes out 43% below Claude Opus 5 and 50% below Claude Fable 5.1. And $1.40 of the $1.75 gap against Claude Opus 5 — eighty per cent of the saving — comes from the cache read line alone, not from the headline input rate.

Now run the opposite shape: a single-shot call with 20,000 input tokens and 8,000 output, no caching. Claude Opus 5.5 costs $0.08 of input plus $0.16 of output — $0.24. Claude Opus 5 costs $0.10 plus $0.20 — $0.30. That is exactly the 20% sticker cut, and not a point more.

So the rate card alone delivers somewhere between 20% and 43% depending on how much of your bill is cache reads. Any claim above that range is coming from token counts, not from prices.

Where Anthropic's "about 40% cheaper to run" comes from

Ant​hropic's launch material says Claude Opus 5.5 "requires less compute to serve than Opus 5" and that its tests show it "will cost 40% less than Opus 5 on typical workloads" at default settings. The decomposition Ant​hropic gives is that the model "costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs."

Label that for what it is: a vendor characterisation of an average across workloads Ant​hropic selected, not an audited result and not a rate published on the price sheet. The verifiable half is the first half — 20% off input and output, 60% off cache reads, both readable on the sheet today. The second half depends on the model spending fewer tokens and fewer turns on your workload, which is only checkable by running it.

The evidence Ant​hropic offers for the token half is a set of launch-time customer statements. All of the following are vendor-published, describing workloads the vendor and its customers chose:

• Box — the model "used a third of the tokens Opus 5 did," with answers "40% less verbose without losing accuracy."

• Kiro — it "solved more than Opus 5 while making about 40% fewer calls and using half the tokens."

• Factory — it "matched Opus 5 on high effort, while using 20 to 25% fewer output tokens."

• GitHub — it "used among the fewest tokens and steps we measured," and in VS Code "solved more terminal tasks than Opus 5 in less than half the steps."

• Merger analysis, one worked example — two fictional HR software companies, each model building an Excel model and an executive presentation. Ant​hropic reports Claude Opus 5.5 finished in 63 minutes against 93 for Claude Opus 5 and "cost 50% less to produce." Vendor-reported, n=1.

The independent counterweight matters here, because it points the other way. Artificial Analysis measured Claude Opus 5.5 at max effort at roughly 119,000 output tokens per Intelligence Index task, against roughly 73,000 for Claude Opus 5 at max effort — the token count goes up, not down. Its summary of that configuration is that Claude Opus 5.5 is "level with Opus 5 on cost per task despite 1.6x the output tokens," at a published cost of $5.98 per Intelligence Index task. That is an independent measurement at max effort, and it is not a contradiction of Ant​hropic's claim — it is the other end of the effort scale.

Which brings up the caveat that governs every number in this section. Ant​hropic states that unless otherwise noted, all its Claude Opus 5.5 results use adaptive thinking at max effort. Claude Opus 5.5 defaults to medium effort; Claude Opus 5 defaulted to high, and the same named level is not the same thinking budget across the two models. So a vendor table run at max effort and an independent table run at max effort with default fallback enabled are not the same configuration, and neither is the default setting a reader would actually deploy. A 40% average measured across a selected set of workloads, at settings you may not use, is a starting hypothesis — not a budget line.

Fast mode is a research preview, not a tier

Fast mode on Claude Opus 5.5 is $8.00 per million input and $40.00 per million output — exactly double standard pricing — for up to 2.5x higher output tokens per second. It does not improve time to first token; the gain is in output speed, and it is most visible with streaming.

Say the rest plainly, because it is easy to budget this as a tier when it is not one. Ant​hropic labels fast mode a research preview. Access is requested through an account manager or the waitlist. It requires the fast-mode-2026-02-01 beta header and speed: "fast" on the request. It is available on the Claude API only — not on Amazon Bedrock, Goo​gle Cloud, Microsoft Foundry, or Claude Platform on AWS. It cannot be combined with the Batch API or with a Priority Tier commitment. It does stack with prompt caching and data-residency multipliers, so a US-pinned fast-mode cached request pays 1.1x on top of 2x. And switching a workload between fast and standard speed invalidates the prompt cache, which on a long agent loop can cost more than the speed saves.

Claude Opus 5 and Claude Opus 4.8 also support fast mode, at $10.00/$50.00. Claude Opus 4.7 rejects the parameter outright.

The unit that decides the bill is cost per finished task

Every rate above is a price per token. No one buys tokens. A per-token price becomes a cost only when multiplied by the tokens a task consumes and the number of attempts it takes to finish.

Run the arithmetic on the cache-light example from earlier. At $0.24 per pass, a model that finishes a task on the first attempt costs $0.24. The same task on a competitor at $0.30 per pass costs $0.30 — so the cheaper model looks 20% ahead. Now let the cheaper model need 1.2 passes per finished task, which is a generous success rate for agentic work: 1.2 × $0.24 is $0.288, against $0.30 for the single-pass alternative. The saving is gone. At three passes it is $0.72, and the model that looked 20% cheaper per token is 140% more expensive per finished job.

That is why Ant​hropic's own efficiency claims — fewer tokens, fewer calls, fewer steps — are framed in the right unit. It is also why none of them should be taken as a budget line until you have reproduced them. The check is mechanical: price per pass, times passes per finished task, measured on your own traffic. Both numbers are yours to produce, and neither vendor publishes the second one for your workload.

If you want a single reference point from outside the vendor, Artificial Analysis puts Claude Opus 5.5 at $5.98 per Intelligence Index task in its max-effort configuration — an independent figure at a setting most production traffic will not use, useful as a ceiling rather than a forecast.

Lines the headline rate does not include

• Residency. US-only inference is 1.1x on every token category, including cache reads and writes. It stacks with fast mode and with the Batch discount.

• Tool definitions. Ant​hropic publishes a per-model tool-use system prompt overhead: declaring tools adds 286 tokens to every Claude Opus 5.5 request. Small at $4.00 per million, but it is charged per request, and an agent loop makes a lot of requests.

• Cache invalidation. A prefix rewritten after a model, effort, tool-definition or speed change is billed at the write rate — $5.00 or $8.00 per million on Claude Opus 5.5 — not the $0.20 read rate.

• The model below it. Claude Sonnet 5 sits at $2.00 input and $10.00 output per million, which Ant​hropic's pricing footnote confirms is now the standard price rather than the introductory rate it was announced as. Claude Opus 5.5 is exactly 2x Sonnet 5 on both directions.

Calling Claude Opus 5.5 at Anthropic's list price

Claude Opus 5.5 is on OrcaRouter at anthropic/claude-opus-5.5 for $4.00 input and $20.00 output per million — Ant​hropic's own list price, passed through at 0% markup. That pass-through is worth stating precisely on a pricing page: because the provider rate is forwarded rather than cached in a table, a vendor rate change is live here the same day it is published, not after a sync window. If Ant​hropic moves this sheet, the number you call at moves with it.

The same key reaches the tiers this page compares it against — anthropic/claude-opus-5 at $5.00/$25.00 and anthropic/claude-fable-5.1 at $10.00/$50.00 — alongside the rest of the Claude lineup and more than 200 other models, so testing whether the cheaper per-token rate survives contact with your own pass counts does not require a second contract or a code change. Automatic failover is there for the case where you want to run an unfamiliar model on a production path without betting the path on it.

The two-line answer to what this page is about: the rate card is $4.00 in, $20.00 out and $0.20 per million cached read, and the 20% cut is real and verifiable today. The 40% is Ant​hropic's average across workloads it selected. Run one of your own agent loops through it before you budget against the second number.