Most of the AI work inside a marketing agency is not clever. It is repetitive. Tagging search terms. Summarizing call notes. Checking that UTMs follow the naming convention. Drafting the first pass of a weekly report so a human can fix it instead of writing it from scratch.
For every one of those tasks the automation question is the same. Does the cost per run, multiplied by the number of runs, beat the time a person would spend? When small models were relatively expensive, many of these tasks failed that test, or passed it only barely. Anthropic's release of Claude Haiku 5.5, which Simon Willison covered on his weblog on 7th October 2026, moves that line. It does not move it as far as the headline price suggests.
This piece walks through what changed, where the hidden costs are, how to compare cost per run instead of cost per token, and the routing rule we would use to decide which model handles which step.
What actually changed
The headline is price. According to Simon Willison's weblog, the previous Haiku, 4.5, was priced at $1 per million input tokens and $5 per million output tokens. The new model exactly matches the price of GPT-6 Luna, at $0.10 input and $0.50 output, up to 100,000 tokens. Beyond 100,000 tokens the price increases 5x to $0.50 and $2.50.
| Model and prompt size | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Haiku 4.5 | $1 | $5 |
| Haiku 5.5, up to 100K tokens | $0.10 | $0.50 |
| Haiku 5.5, beyond 100K tokens | $0.50 | $2.50 |
Source: Simon Willison's weblog, with the same tiers reported in Latent Space's AINews roundup.
A few other changes matter for agency workflows:
- Context window. Latent Space reports a "1M-token context, up from 200k for Haiku 4.5."
- Cheap cache reads. Latent Space's pricing table lists cache reads at $0.01 per 1M tokens under 100K and $0.05 over 100K. If your agents reuse the same long system prompt or brand guidelines on every call, this is where real savings sit.
- Effort settings. Latent Space notes it is "the first Haiku with Anthropic's effort settings and adaptive thinking." The Simon Willison write-up adds that the new Haiku does not let you disable reasoning and defaults to medium effort.
- Positioning. According to Latent Space, Anthropic positions it as a subagent paired with Opus 5.5 or Sonnet 5.5, for high-volume, cost-sensitive work such as summaries, compactions and database queries.
That last point is the important one. Anthropic is not pitching Haiku 5.5 as the model that does everything. It is pitching it as the worker underneath a smarter lead model. That is exactly the shape most agency workflows should have.
The hidden costs inside the headline price
Three things eat into the per-token discount, and all three show up in the sources.
A less generous tokenizer. Simon Willison found that "the same long prompt uses around 1.25x as many tokens with Haiku 5.5 compared to Haiku 4.5, so there's a hidden price increase there." Your existing token estimates from Haiku 4.5 logs will undercount.
Verbosity at high effort. Latent Space, citing Artificial Analysis, reports that at max effort the model "uses about 162k output tokens per Index task, roughly 3x GPT-6 Luna at max (~50k)." Cheap tokens are less cheap when you use a lot more of them.
The 100K cliff. Above 100,000 tokens of prompt, the price steps up by the 5x described above. Long agent loops that accumulate history can cross that line without anyone noticing.
These effects explain why the savings claims vary. Anthropic said, per Latent Space, that the model costs "about 75% less to run than Haiku 4.5 on average." Cursor claimed "10x less than Haiku 4.5" on shorter requests. Latent Space's own reading is that heavy token use and the price step above 100K likely explain the gap between the two. Both claims can be true. Which one applies to you depends on your prompts.
It is also worth knowing that Latent Space flags the independent cost figures as provisional, because Artificial Analysis does not yet model the 5x price step above 100K tokens.
Cost per run, not cost per token
Per-token prices are useful for comparing vendors. They are useless for deciding whether to automate a task. What you need is cost per run:
Cost per run = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)
Then multiply by runs per month and compare against the human time the task currently takes.
Simon Willison's pelican test is a good illustration of why effort level dominates this calculation. Using the same prompt, the low effort run cost 0.0936 cents and took 7 seconds, while the max effort run took 5 minutes 9 seconds to generate and still cost only 3.3826 cents. A year earlier, the same test on Haiku 4.5 cost 0.7583 cents.
Read those three numbers together. At low effort, Haiku 5.5 came in at 0.0936 cents against Haiku 4.5's 0.7583 cents. At max effort, it came in at 3.3826 cents, above what Haiku 4.5 cost. Same model, same prompt, and the effort setting decided whether the new model was cheaper or more expensive than the old one.
So for each task you are considering, measure real token usage from logs at the effort level you will actually run in production. Do not estimate from the length of the answer you see.
Here is how we think about the four task types in the brief:
- Tagging and classification. Short inputs, tiny outputs, easy to validate against a fixed list. The best fit for low effort.
- Summarizing. Fits well if the source material stays under 100K tokens. Chunk long transcripts rather than sending them whole.
- QA checks. Naming conventions, missing parameters, broken links in a draft. Rule-like, checkable, a strong fit.
- Report drafts. A fit for the first pass, not the final. The draft is cheap. The judgment about what the numbers mean is not, and that belongs to a stronger model or a person.
A routing rule for agency workflows
The rule we would use is simple: default to the cheap model, escalate on purpose.
Send a step to Haiku 5.5 when all four of these are true:
- The prompt stays under 100K tokens. Above that, you are paying the higher tier and losing the reason to route there.
- The output is checkable. A schema, a fixed list of labels, a rule, or a quick human glance can tell you whether it is right.
- The facts come from the input, not from the model's memory. Latent Space reports Artificial Analysis's AA-Omniscience results for Haiku 5.5 as 36% accuracy and a 40% hallucination rate, and notes that part of the lower accuracy comes from the model being more willing to say it doesn't know. Give it the data. Do not ask it to remember things.
- Effort starts at low or medium. Raise it only when a measured quality check fails.
Escalate to Sonnet or Opus when the step involves client-facing judgment, strategy, or long-running context that will cross 100K tokens. Also escalate when the task depends on reading charts: Latent Space summarizes independent ParseBench testing as finding it weaker on chart understanding, while good at tables for the price.
One more caveat from the same roundup. Artificial Analysis's AutomationBench score was depressed by a pre-release safety bug that made the model over-refuse, and Anthropic is working on a fix. If you see unexpected refusals in an automation step, that is a known issue rather than your prompt.
If you are still deciding which parts of your marketing automation need a language model at all, our piece on decision models versus text models in marketing automation is a useful companion to this routing rule.
Why this matters for agency economics
Cheaper runs do not just shave the API bill. They change which tasks clear the bar for automation in the first place. Work that was not worth wiring up at the old price may be worth it now, provided the prompts stay short and the effort stays low.
That shift matters most to agencies already rethinking how they staff and price work, which is the thread running through how eight agencies rewrote their business model with AI. The routing rule is the operational side of that change.
The same release also touched the tier above. Latent Space reports that Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per 1M tokens. If your lead model reuses long context, that cut lowers the cost of the escalation path too.
What to do this week
- List your high-volume steps. Every place an agent tags, summarizes, checks or drafts.
- Pull real token counts. From logs, not guesses. Remember the tokenizer difference Simon Willison measured if your baseline is Haiku 4.5.
- Compute cost per run at low and medium effort, and check how many runs cross 100K tokens.
- Apply the four conditions. Route what passes to the cheap model and keep the rest where it is.
- Set a hard budget limit. Anthropic lets you disable auto-reload, so that API requests stop when your balance runs out. For an experiment, that is exactly the safety net you want.
The model got cheaper. The discipline of measuring each step did not get any less necessary.
If you want help building model routing, cost tracking or the analytics underneath your agents, talk to Bitegrico. We implement MCP connections, GA4, BigQuery and GTM pipelines so your AI workflows run on data you can trust and costs you can see.