A 2,000 by 1,500 photo costs 2.5 times the visual tokens on Claude Haiku 5.5 as on Haiku 4.5

Send a 2,000 by 1,500 pixel photograph to Claude Haiku 5.5 and it counts as about two and a half times as many visual tokens as the same file did on Claude Haiku 4.5. The figure is Anthropic's own. The Haiku 5.5 migration guide puts it in one sentence: "An image of 2,000 by 1,500 pixels costs about 2.5 times as many visual tokens on Claude Haiku 5.5 as on Claude Haiku 4.5."

The per-token price has nothing to do with it. The cause is the resolution tier the model sits in. Haiku 5.5 "uses the high-resolution image tier, which downscales images above 2,576 pixels on the long edge or 4,784 visual tokens", the guide says, while "Claude Haiku 4.5 and earlier Haiku models use the standard tier, which downscales images above 1,568 pixels on the long edge or 1,568 visual tokens". The older tier shrank a three-megapixel photo before counting it. The newer one does not.

Anthropic publishes the arithmetic, and it varies by size

Anthropic's vision documentation carries a table of worked examples. A visual token is a 28 by 28 pixel patch, so an image costs the number of patches it covers. On the standard tier a 2,000 by 1,500 image is downsized to 1,269 by 952 and costs 1,564 visual tokens; on the high-resolution tier it is not resized at all and costs 3,888. That is the 2.5x, and the division gives about 2.49.

The multiplier is not one number, though, and the same table shows why. A 1,920 by 1,080 screenshot goes from 1,560 visual tokens to 2,691, a factor of about 1.7. A 3,840 by 2,160 frame goes from 1,560 to 4,784, about 3.1. An image inside the standard tier's limits costs the same on both, such as 1,092 by 1,092 at 1,521 tokens. The 2.5x belongs to the size Anthropic chose to illustrate, and a real workload's figure depends on what it sends.

The pictures still cost less, and the counts still bite

Visual tokens are input tokens, and Haiku 5.5's input price is a tenth of its predecessor's. Anthropic's pricing page gives Haiku 5.5 $0.10 per million input tokens on a prompt up to 100,000 tokens against Haiku 4.5's flat $1, so 3,888 visual tokens on the new model work out cheaper than 1,564 on the old one by roughly a factor of four. Above 100,000 tokens Haiku 5.5 charges $0.50, which still leaves it under the model it replaces.

What the higher count changes is everything measured in tokens rather than dollars. The guide's instruction is to "Recount your prompts, and revisit max_tokens limits and cost estimates, because the same text counts as more tokens and large images count as more visual tokens." A per-request token budget tuned on Haiku 4.5 fills faster, a token-counting call returns a bigger number, and anyone metering a customer by tokens is metering more of them for the same picture. The text half of that sentence is the separate change the same guide puts at "approximately 30% more tokens" for the same input text, the change that narrowed the saving Anthropic claimed at launch.

Dating a page Anthropic keeps editing

The high-resolution tier is not a Haiku feature. Anthropic assigns it to "Claude 4.7 and later models", with the standard tier for "All other models", and says it "is automatic on the listed models and requires no beta header or client-side opt-in". Haiku 5.5 arrived on October 7 as the first Haiku-class model on that side of the line.

One caution on the dates. Anthropic edits published release-note entries and documentation after the fact, including text under earlier date headings, so an entry dated October 7 describes what the page said when it was read rather than what shipped that day. The migration guide and the vision documentation here were read at 21:36 UTC on October 10 and the pricing page at 21:33 UTC, and that guide also runs to a checklist of code changes a Haiku 4.5 request needs before it will run.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.