Claude's Price Dropped Up to 10 Times per Token, but Margin for Those Building on the API Didn't Grow the Same Way

The cut that shows up in the table, not in the announcement

Anthropic didn't issue a press release about a price cut. The change is in the pricing documentation (docs.anthropic.com/en/docs/about-claude/pricing), in the comparison table across model generations. It's there, not in any press note, that whoever charges by LLM usage needs to read the numbers before closing the next contract with a client.

And the correct reading isn't "Claude got cheaper." It's more specific: it got much cheaper in one segment and remains expensive, on purpose, in another. This difference is the strategic bet behind the table, and it's what changes the math for those building a product on top of the API.

Where the price really plummeted

The most aggressive cut is in Claude Haiku. Generation 4.5 charges $1/MTok for input and $5/MTok for output. Generation 5.5, for prompts up to 100,000 tokens, charges $0.10/MTok for input and $0.50/MTok for output. That's a 10x cut on both sides.

A simple example illustrates the size of this: a product that processes 100 million input tokens per month in a classification or routing task used to pay $100 on generation 4.5. On 5.5, it pays $10. This isn't fine-tuning of price, it's a reshaping of what kind of product becomes viable to run at volume.

Sonnet also dropped, but more modestly: from $3/$15 per MTok (Sonnet 4.6) to $2/$10 (Sonnet 5.5), a drop of one third. Opus dropped from $15/$75 in the original version 4 to $4/$20 in 5.5, a reduction of nearly 75% across generations.

Where the price didn't move (and why that's the important part)

Claude Fable 5.1 and Claude Mythos 5.1, Anthropic's current top-of-the-line models, charge $10/MTok for input and $50/MTok for output. That's exactly the same price as the previous versions, Fable 5 and Mythos 5. No cut there.

In short: Anthropic commoditizes the repetitive-task layer (classification, extraction, routing, where Haiku lives) and defends the price of the cutting-edge reasoning layer (where Fable and Mythos live). This isn't a spreadsheet accident, it's market segmentation: whoever competes on production volume cuts price, whoever competes on unique capability keeps margin.

For those deciding product architecture, the implication is direct: LLM economies of scale aren't uniform. They show up at full strength only for those who can push the workload to the cheaper model without losing enough quality for the use case.

The new tokenizer charges back part of the bill

The documentation itself has a note that changes the reading of the cut: models from the 4.7 series onward (including Sonnet 5.5, Opus 5.5, and Mythos Preview) use a new tokenizer that generates approximately 30% more tokens for the same text, compared to the previous tokenizer used up through Sonnet 4.6.

This means the nominal per-token price cut doesn't translate one-to-one into a per-task cost cut. Doing the math: if Sonnet 4.6 charges $3/MTok for input and Sonnet 5.5 charges $2/MTok but needs 30% more tokens to process the same text, the effective cost for that text comes out to around $2.60, not $2. The real savings drop from 33% to something close to 13%.

This isn't a calculation Anthropic hides, it's documented. But it's a calculation the founder has to do on their own, because no pricing card shows the real cost per task, only the nominal cost per token.

The multipliers nobody calculates in the pitch deck

The base per-token price is just the starting point. The pricing table lists at least four multipliers that change the final cost of a request, and that rarely make it into the financial model of those validating unit economics:

• Data residency: using inference_geo: "us" to guarantee that inference stays in the US adds a 1.1x multiplier across all token categories, including cache. For those serving clients with a contractual requirement for data to stay in the US (finance, healthcare), this is a fixed cost, not optional.
• Fast mode: on Opus 5.5, fast mode charges $8/MTok for input and $40/MTok for output, double the standard price ($4/$20). It's a second product line within the same model, for those who need low latency in a customer-facing feature.
• Cache: a cache hit costs 10% of the standard input price in most models, but drops to 5% on Opus 5.5 and Sonnet 5.5, and to 2.5% on Fable 5.1 and Mythos 5.1. Architecture with a long, reused prompt (large system prompt, conversation history) changes cost depending on which model is behind it.
• Batch: asynchronous processing gives a fixed 50% discount on input and output, across any model in the table. It doesn't work with fast mode, but it's the simplest discount to capture for any workload that doesn't need a real-time response.

Those running on Amazon Bedrock or Microsoft Foundry via marketplace have one more layer: billing uses Claude Consumption Units, converted at $0.01 per CCU based on the standard dollar price. Negotiated discounts show up as "fewer CCUs metered," not as a visible percentage on the invoice, which makes it harder for the finance team to audit the real impact of any enterprise deal.

The counterpoint: today's discount could be a war subsidy

The strongest argument against treating this cut as a permanent margin gain is simple: a commodity price falling 10x in one model generation isn't necessarily an inference cost falling 10x. It can also be market competition, with Anthropic, OpenAI, and Google competing to be the cheap layer that processes the production volume of those building AI products.

If that's the case, today's price isn't a structural floor, it's a tactical position. A founder who builds the product's entire margin model on top of Haiku at $0.10/MTok runs the risk of seeing that number rise again if competitive pressure shifts, or if the specific model is discontinued and replaced by a successor with different pricing, as has already happened between generations in this very table.

The defense against this risk isn't ignoring the cut, it's not baking the current price into a long-term contract with the end client without an adjustment clause. Separating infrastructure cost pricing from the price charged to the client gives room to maneuver for the next change in the table, whether it's a cut or a hike.

What changes for those who decide

Anthropic's pricing table gives those building on the API three concrete decisions, not just one. First, segment tasks by model: high-volume classification, extraction, and routing move to Haiku, where the cut is real and large; complex reasoning and long-running agentic work stay on Opus or Fable, where Anthropic didn't lower the price and won't compete on cost there.

Second, separate what's real-time from what isn't. Any workload that can tolerate delay (report generation, batch processing, data enrichment) goes to the Batch API and already includes a 50% discount, with nothing to negotiate. Third, model cost per task, not per token: the new tokenizer charges back part of the nominal cut, and caching, fast mode, and data residency change the final bill more than the headline price line in the table.

Those who only look at the big number ($2/MTok, $0.10/MTok) are reading half of the story Anthropic tells. The other half is in the footnotes, and that's where the product's real margin lives.