Your senior engineer pastes a Jira link into Slack and asks the channel agent to summarize the sprint blockers, cross-reference the open GitHub PRs, and post a brief. Three tools, one request, maybe 40 seconds. A month ago that roundtrip cost roughly 20× what it costs today.
Anthropic released Claude Haiku 5.5 on October 7, 2026, and the headline is price: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens - a 90% cut from Haiku 4.5. That is not a marginal discount. It is the kind of number that changes which problems are worth automating.
What $0.10 per million tokens actually buys a team
Haiku 5.5 is Anthropic's small, fast model for high-volume, cost-sensitive work - summarization, subagents, and browser use - with a one-million-token context window and a maximum output of 128,000 tokens. The context window alone is a meaningful change from previous Haiku generations: it means a subagent can hold a long thread, a chunked document, and its tool results in one call without truncating.
It is also the first Haiku-class model with an adjustable effort control - adaptive thinking is on by default, with effort defaulting to medium.
In practice that means you can dial it down to low for pure classification tasks and back up to high when a subagent needs to reason through an ambiguous edge case. Most routing steps will run fine at low or medium - and at medium, you are paying a tenth of a cent per thousand input tokens.
Here is how the pricing stacks against the previous generation and Sonnet:
| Model | Input / 1M tokens | Output / 1M tokens | Cache read |
|---|---|---|---|
| Haiku 4.5 | $1.00 | $5.00 | $0.10 |
| Haiku 5.5 (≤100K prompt) | $0.10 | $0.50 | $0.01 |
| Haiku 5.5 (>100K prompt) | $0.50 | $2.50 | $0.05 |
| Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
In an illustrative subagent run, 20 Haiku 5.5 subagents account for roughly 7% of the total bill. The orchestrator - the frontier model doing the planning - stays the expensive part. That is a meaningful rebalancing: it means you can spin up more subagents without the economics collapsing.
The 100K-token cliff nobody budgeted for
The catch is in the docs, not the launch post: code written for Haiku 4.5 can return 400 errors on the new model, and the same text counts as roughly 30% more tokens.
That second point is the one to internalize. If your prompts currently land at 80,000 tokens on Haiku 4.5, they may land at 104,000 on 5.5 - pushing you across the pricing cliff where the Claude Haiku 5.5 price for prompts up to 100,000 tokens is $0.10 per million input tokens and $0.50 per million output tokens, but a prompt over 100,000 tokens pays $0.50 input and $2.50 output. That is a 5× jump, not a rounding error.
The practical rule: measure your actual token counts on 5.5 before you set a budget. If you are feeding a subagent a long conversation history plus a big tool result, count carefully.
What it changes for teams running agents in Slack
The practical shift is in the economics of fan-out. An orchestrator - whether that is a Copilot agent, a custom bot, or a teammate like Beagle - can now spin up multiple cheap subagents to work tasks in parallel without those subagents dominating the bill. The expensive token budget belongs to the planner; the workers are nearly free.
Channel Agents in Teams channels can securely interact with MCP servers, enabling multi-step workflows through natural language prompts - so instead of switching between apps, users can ask the agent to fetch GitHub pull requests, create Jira issues, or sync Asana tasks, all through conversational prompts. The subagent price drop makes that kind of fan-out tractable even at volume.
For a support team fielding 200 Zendesk tickets a day, routing each ticket through a Haiku 5.5 classifier (200 tokens in, 50 tokens out) costs less than $0.03 a day at short-prompt rates. The decision to automate that triage used to involve a cost-benefit argument. Now it does not.
Beagle in action#eng-standup, 9:02amClaude Haiku 5.5 subagent cost: common questions
How much does Claude Haiku 5.5 cost per API call?
Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, then $0.50 and $2.50 above that, with a 1M token context window and 128K max output. A single short subagent call - say 500 input tokens, 100 output - costs roughly $0.00005, or about one-twentieth of a cent.
Is Claude Haiku 5.5 good enough to use as a subagent, or does it make too many errors?
Claude Haiku 5.5 scores 66.8 out of 100 across tracked benchmarks, ranking 27th of 218 models, with its strongest eligible category being Agentic at rank 18. For well-scoped routing, classification, and tool-call tasks with clear schemas, it performs reliably. For open-ended reasoning steps where errors compound, you still want a larger model in the orchestrator role.
What is the 100K-token pricing cliff in Haiku 5.5?
Prompts up to 100,000 tokens use the low tier ($0.10 input). Prompts above that jump to $0.50 input - a 5× increase. Code written for Haiku 4.5 can also return 400 errors on the new model, and the same text counts as about 30% more tokens on 5.5. Measure real token counts on the new model before setting cost assumptions.
Can Haiku 5.5 run in Slack or Teams agents?
Haiku 5.5 is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Any agent framework that calls the Claude API - including custom bots in Slack or Copilot Studio agents in Teams - can route subagent steps to it today.
Does the adjustable effort setting change the price?
The effort toggle changes compute depth, not the token price directly. Running at low effort produces shorter thinking traces and fewer output tokens, which lowers your bill through lower output volume - not through a discounted rate.
It is the first Haiku model with adjustable effort; thinking is adaptive and on by default, with effort as the main lever for trading off depth, latency, and cost.
For pure classification steps, low effort is usually enough.