Claude Opus 5.5 Is 60% Cheaper to Cache — and That Is the Price Cut

The short version: for anyone paying per token, this is the first Anthropic price cut that changes the arithmetic of running agents. Cache reads — the line item that dominates a coding loop, because the same context is re-read every turn — fell from $0.50 to $0.20 per million tokens, which is Sonnet 5's cache-read price. On input and output the cut is 20% ($5/$25 → $4/$20). If you pay $20 a month instead of per token, none of this is your price: subscribers got higher five-hour limits and a saveable reset that expires October 22. And if you run an agent through the API, budget a migration — four request shapes now return 400s.

The numbers, straight off the pricing doc

Anthropic's own table, read on 2026-09-25, per million tokens:

The same page adds three footnotes the launch post does not lead with: batch is $2/$10, the Fast mode research preview is $8/$40 — exactly double, as the name suggests — and inference that has to stay inside the US carries a 1.1× multiplier.

Anthropic's own framing of the drop is a table of two columns, Opus 5.5 against Opus 5, which is a fair description of the launch and a slightly incomplete description of your bill. Here is what the two columns leave out.

Why the cache read decides your bill

In a chat window, the expensive tokens are the ones the model writes. In an agent loop, they are the ones it re-reads. Every turn ships the whole conversation plus the tool output back through the model, and if you cached that prefix, those tokens bill at the cache-read rate rather than the input rate. Opus 5.5's cache reads bill at 0.05× its base input rate ($0.20 against $4). On Opus 5 and Sonnet 5 that ratio is 0.1×, and on Fable 5.1 it is 0.025×. So the headline "40% less to run than Opus 5" is really a statement about how cache-heavy your workload is.

Three token mixes, priced at today's rates:

Math: cached tokens at the cache-read rate, fresh tokens at input, written tokens at output. Anthropic's docs also list a worked example of the same shape, so the method is theirs, only the mixes are ours.

Read the table as a warning about discounts you were promised on Twitter. If you use Opus 5.5 as a chatbot, you got 20%. If you use it as a coder agent with a big cached repo context, you got closer to 46%. Nobody is lying; the number just moved with your token mix.

Three things the cut does not cover

Token efficiency. The most-quoted worry in the 1,116-comment thread, before anyone had run a benchmark: "need to see its token efficiency to better assess. Been hearing rumors it generates much more output tokens per task" (nickandbro, HN). Output is the priciest line at $20, so a model that thinks longer and writes more can eat a 20% discount without trying. That claim is a rumor, and the reply that matters — the real output-token count on your own workload — is available to you in your usage dashboard and to nobody writing launch posts.

Fast mode is double. The "more than 30% faster output" figure that made the round-up posts (joshstrange, HN quoted it approvingly) comes with a price list: $8/$40, a research preview, still double the standard rate.

Data residency and batch are separate levers. Keeping inference in the US adds 10%; batching halves it. Neither appears in the "40% less" sentence, and for a production pipeline they move the number more than the headline did.

If you pay $20 a month, none of this is your price

Worth saying plainly, because half the reaction to this launch treated it as a subscription announcement. Anthropic's plans page today lists Free at $0, Pro at $17/month billed yearly ($200 upfront) or $20 monthly, and Max from $100/month. The API prices above apply to API customers.

What subscribers actually got is limits. Five-hour usage limits went up across Pro, Max, Team and Enterprise, and the reset handling changed: "with previous releases I've had my usage instantly reset, but now in the Claude app I've got a 'Reset for free' button that expires Oct 22, which seems to effectively be a whole new usage window I can activate whenever's convenient" (cogythea, HN). Another reader dug the same line out of the FAQ: "Reset for free: Get extra wiggle room to explore Opus 5.5. Expires Oct 22" (ayhanfuat, HN).

That is a better deal for a heavy subscriber than a token discount, and it is also the kind of thing that quietly expires. Put October 22 in your calendar if you have been saving the reset.

The migration bill: four 400s and two silent changes

If an agent wrapper of yours pins model names, this release is not a drop-in. The New Stack's write-up on the agent-facing changes lists four request shapes that now fail with a 400 rather than degrading:

• Thinking can no longer be turned off — extended thinking is always on.
• Forced tool choice is gone; you cannot pin a specific tool call.
• Thinking blocks are now bound to both the model and the conversation they came from, so replaying or transplanting them breaks.
• The computer-use toolset changed shape.

Two more changes will not throw errors and are therefore worse: narration output now arrives as empty thinking blocks, and the default thinking effort dropped from high to medium. If your prompts relied on optional thinking or on explicit narration, you will notice it as a quality regression weeks later, not as a stack trace. None of this is unusual for a model release, and none of it appears in a price table — which is the point. The cost of switching models is measured in an afternoon of debugging, not in dollars per million tokens.

What the thread is arguing about

The benchmark contradiction. "They write that at the top, but then on benchmarks, it beats literally every other model, including Fable and Astra?" (pookieinc, HN). The launch post says Opus 5.5 performs at Fable 5.1's level on most work while the benchmark table shows it leading Fable on agentic coding, computer use and knowledge work, and goes on to admit that "benchmark margins have become a less reliable guide to real-world differences." One reader read that as an admission big enough to worry about (km144, HN): if the margins no longer predict the experience, what are the tables for?

The market read. "Opus 5 is the model with highest spend on openrouter and it seems plausible that Opus 5 is/was the highest spend model in the world, and certainly Anthropic's biggest moneymaker. If you are forced to reduce price despite raising capabilities, that certainly tells something about the market, and potentially about Anthropic future profitability too" (GodelNumbering, HN). A company cutting the price of its best-selling product while adding capability is either buying share or defending it. Both explanations are good news for your invoice.

The fatigue. Two sides of it: "Anthropic models are ridiculously expensive. I've stopped using any of their models months ago" (thibran, HN), and "Astra is just so good. And the ChatGPT subscription lets me use my own harness... I don't know how much better a model would have to be to get me to move off OpenAI at this point" (abtinf, HN). Anthropic's own answer to the second one is buried in the same post: Sonnet 5.5 and Haiku 5.5 in the coming weeks, which is the announcement some readers found more interesting than the flagship (sznio, HN on wanting "a decently capable fast model").

And one line worth keeping, given that Anthropic skipped 5.1 through 5.4: "we really are plateauing" (alpineman, HN). Version numbers are marketing; the price list is not.

Who should switch, and who should not

GPT-6 Sol vs Luna, priced → The 34-provider free tier →

FAQ

How much does Claude Opus 5.5 cost? $4 per million input tokens and $20 per million output tokens, per Anthropic's pricing documentation read on 2026-09-25. Cache writes are $5 (5-minute) or $8 (1-hour); cache reads are $0.20. Batch is $2/$10, the Fast mode research preview is $8/$40, and US-only inference adds a 1.1× multiplier.

Is Opus 5.5 cheaper than Opus 5? On every line: 20% on input and output, 60% on cache reads. Your actual saving depends on your token mix — roughly 21% on an output-heavy answer, about 46% on a long agent loop with a large cached context.

Does the Opus 5.5 price cut apply to Claude Pro subscribers? No, those are API prices. What changed for subscribers is usage limits: higher five-hour limits on Pro, Max, Team and Enterprise, plus a saveable "Reset for free" in the Claude app that expires October 22.

Do I have to change code to move to Opus 5.5? Only if you have a custom agent or tool harness. Thinking can no longer be disabled, forced tool choice is unsupported, thinking blocks are tied to the model and conversation, and the computer-use toolset changed — all four now return 400 errors.

Method: prices read from Anthropic's own pricing documentation and plans page on 2026-09-25; model behaviour changes read from The New Stack's agent-migration write-up (linked above); user quotes taken verbatim from the 1,116-comment Hacker News launch thread (1,791 points, 2026-09-22) with comment links. We take no payment from vendors. Prices move — re-check before you commit a budget.