Claude Haiku 5.5 pricing: the 100K token cliff

Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, but only while a single prompt stays at or under 100,000 tokens. Cross that line and the whole request reprices at $0.50 in and $2.50 out, five times the rate. For a chatbot that answers questions about a plumber's hours and service area, you will never get near it. For an agent that drags a growing pile of context behind it, you can hit it without noticing.

Anthropic launched the model on October 7, 2026. We read the launch page and the pricing page this week, then checked what HighLevel charges for the same model. Here is what the cliff costs, who it matters to, and who can ignore it.

What does Haiku 5.5 cost, exactly?

Anthropic lists two tiers. Prompts up to 100,000 tokens pay $0.10 per million input tokens and $0.50 per million output. Longer prompts pay $0.50 and $2.50. Cache reads drop to $0.01 per million under the line and $0.05 over it, and cache writes run $0.125 under and $0.625 over. The pricing page notes the cache figures assume a five-minute TTL.

The previous small model, Haiku 4.5, was $1.00 in and $5.00 out. So the short-prompt tier is a 90 percent cut. Anthropic's own launch page says the model costs "around 75% less to run" on average, then explains in a footnote that the 90 percent figure applies up to 100,000 tokens and the cut is 50 percent above it. Two numbers for one launch is how you end up arguing with a client who read a headline.

One more detail from the launch page. Anthropic says prompts under 100,000 tokens make up about 90 percent of what developers sent to Haiku 4.5. That is a vendor statistic and I cannot check it. It does match what a front-desk bot looks like in practice, which is a few thousand tokens of instructions and a short chat.

How does the 100K cliff actually work?

Not the way most people read it. The tier is not marginal like an income tax bracket. A third-party breakdown from BuildMVPFast says Anthropic's docs count every input token toward the threshold, cache reads and cache writes included, and that once a request passes it the entire request bills at the higher rate. I could not find that exact wording on the two Anthropic pages I opened, so treat the whole-request detail as secondhand. I would still plan around it, because the downside of being wrong runs in your favor.

Run the numbers on one agent turn with 2,000 tokens of output.

• A 90,000-token prompt: $0.009 for input plus $0.001 for output, about one cent.
• A 120,000-token prompt: $0.060 for input plus $0.005 for output, about 6.5 cents.

That is a third more tokens and 6.5 times the bill. The arithmetic is mine, from Anthropic's list prices, and it assumes no caching. With caching the gap narrows in dollars but not in shape.

Where would a small agency's client actually hit it?

Rarely, and here is the short list of places.

• A support agent that keeps the full chat history of a returning customer and never trims it.
• A bot that stuffs an entire knowledge base into every prompt instead of retrieving the three relevant paragraphs.
• Anything that loads call transcripts, PDFs or a month of CRM notes into the context window to "give it memory."
• A long-running agent that appends every tool result and never compacts.

A rough rule of thumb puts a token at about three-quarters of a word, so 100,000 tokens is on the order of 60,000 to 75,000 words. Anthropic's tokenizer changed with this release, and BuildMVPFast reports it produces roughly 30 percent more tokens for the same text, so lean toward the low end. A dental office's price list, hours, insurance rules and 40 FAQ answers might run 5,000 tokens. You would need to be trying to cross the line.

What does HighLevel charge for the same model?

This is the part I did not expect. HighLevel's AI product pricing page, modified October 29, lists Claude Haiku 5.5 under Conversation AI at $0.11 input and $0.55 output per million tokens. That is ten percent over Anthropic's list. Haiku 4.5 sits at $1.10 and $5.50 on the same page, against Anthropic's $1.00 and $5.00, so the markup is consistent.

The page gives one rate per model. It says nothing about a tier above 100,000 tokens, and I do not know whether HighLevel applies one. If you run long-context agents there, ask support before you quote a client. The page itself says third-party model rates may change and tells you to confirm in-app pricing before quoting externally.

None of this makes HighLevel a bad deal. A ten percent pass-through on a fraction of a cent is nothing. It does mean the headline price cut reaches you intact, and the fee layer on top stays where it was.

Does a cheaper model change what you charge?

No. This is the view I would defend on a sales call.

Take a text chatbot doing a thousand conversations a month. Say ten messages each, a 3,000-token system prompt resent every turn, and replies around 120 tokens. That is roughly 38,000 input tokens and 1,200 output tokens per conversation. At Haiku 5.5's short-prompt rates the model bill is about $0.0044 per conversation, so $4.40 for the thousand. On Haiku 4.5 it was about $44. Those are my assumptions, not a client's invoice, and we ran the same shape of math on a different model in our GPT-6 Luna post.

The $40 difference is real and nobody will pay you less because of it. The client is not buying tokens. They are buying a bot that answers at 11pm, books the job and does not make things up about the warranty.

Voice is where the model barely matters. HighLevel's own worked example prices a 10-minute call at $0.60 before language model tokens and phone charges: $0.45 for the voice engine plus $0.15 for OpenAI text-to-speech. Saving a tenth of a cent per turn on the language model does nothing to a number that size. We broke down that stack in what a voice agent costs per minute.

When is Haiku 5.5 the wrong pick?

When the job needs judgment. BuildMVPFast's own tests found it flagging 30 correct sentences as errors in a review task, and it quotes Anthropic saying Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding. Neither is a verdict on a receptionist, but both point the same direction. A cheap model is for narrow, well-fenced work.

We saw a similar pattern in a voice benchmark, covered in our post on newer models and voice receptionists. Good at deciding what a caller wants, weaker at moving a task forward. If your client's bot books appointments, test the booking path on the cheap model before you switch anything.

Also worth saying: this model is not for speech-to-speech work, where HighLevel lists Google and OpenAI live models at $0.10 to $0.20 a minute all in. It sits behind the voice layer, not in place of it.

Where does this leave a Lumina Sales subscriber?

You do not manage any of it. The chat and voice agents on our platform run on our infrastructure, and the plan includes 500 chat messages and 100 voice minutes a month with no API key to hold (pricing). Client builds are priced at wholesale on a 30-minute call before anything is built. If model prices fall again, that is our margin question to work out, and your retail price stays yours to set.

If you build on your own stack instead, the cliff is yours to manage, and it is a fair reason to do so with eyes open.

FAQ

Is Claude Haiku 5.5 cheaper than GPT-6 Luna?

On list price they match: both are $0.10 in and $0.50 out per million tokens for short prompts. I did not check whether Luna has a long-prompt tier of its own, so look before you assume it is flat. For small prompts it is a tie on cost, and the choice comes down to which model behaves better on your client's script.

Does prompt caching keep me under the 100K line?

Not according to the third-party breakdown, which says cache reads and writes count toward the threshold. Cached tokens are cheap, at $0.01 per million under the line, but they still count as prompt length. A 95,000-token cached prefix plus 6,000 new tokens would cross it.

How many chatbot conversations does a $10 budget buy?

On my example math, about 2,270 conversations at Haiku 5.5 short-prompt rates, and about 230 on Haiku 4.5. Change the prompt length and the answer moves, so rerun it with your client's real system prompt.

Should I tell clients the model got cheaper?

Only if they ask. Their price was never tied to a model, and reopening it invites a discount conversation you have no reason to start.

What to do this week

Open one live client chatbot and count the tokens in its longest recent conversation, including the system prompt. If it sits under 20,000, close the tab and ignore the cliff. If it sits over 80,000, you have a context problem to fix before it becomes a billing one. And if you want to see what a delivered build looks like with the model question already handled, open the live demo and talk to it.