Haiku 5.5 at Luna Pricing: What to Do with a $0.10 Token

At these rates, the risk lies in long requests.

A request with 99,000 input tokens costs about $0.01.

A 101,000-token request at the higher rate will cost about $0.05 if the new price applies to the entire request.

If it applies only to tokens above the threshold, the difference will be smaller. Note

Willison does not clarify this, so the condition should be checked in Anthropic's documentation before budgeting.

This leads to a simple architectural rule: measure and limit context. An agent that adds the entire conversation history to every prompt can cross the threshold unnoticed, making the bill five times higher without a single code change.

That is what RAG is for: only the required 5,000–10,000 tokens enter the model from the knowledge base, while the full corpus remains in storage.

Around the same time, OpenAI released the Decisions API, while

Willison built a plugin for it llm-openai-decisions.

In this mode, the gpt-6-luna model accepts text and images, and you pay only for input: 10 cents per million tokens. The competing Jev works the same way, but costs less: 4.2 cents.

This is suitable for tasks with short, structured responses: selecting a category, confirming compliance, or making a rule-based decision.

The operation’s cost is known in advance because it depends only on the input size, which we control ourselves.

The price list changes monthly, while an integration lasts for years.

That is why model calls should not be hard-coded into business logic.

We build them through LLM & Security Gateway: a single point through which requests pass, where tokens are counted, context-size limits are enforced, and data is filtered.

In this setup, switching from Haiku to Luna or back only requires changing the route; the service code stays the same. The same applies to tools.

If the model accesses systems through MCP, switching providers changes only the model; the toolset stays in place.

This is AI-native integration: a layer is built around the model that does not depend on any specific model, so the model can be replaced.

Cheap tokens matter most where operations are numerous: bulk product-card enrichment, facet assignment, taxonomy maintenance, and tag updates in Pimcore or Akeneo catalogs.

At this volume, the difference between $2,250 and $225 per month matters to the budget.

Decide whether to switch only after measuring performance on your own data.

A tenfold discount makes many tasks viable.

Track the cost of a completed task: token costs multiplied by the number of attempts, plus the cost of an error.

A cheaper model that makes twice as many errors can cost more.

This can be verified by measuring performance on 200-300 labeled examples.

My approach: take two or three real processes, run them through both models, calculate the cost of a correct result, and make sure requests stay below the 100,000-token threshold.

In our estimate, choosing a model after this will take about one day.

When provider prices match, performance on your data is what matters.