What changed
ClaudeDevs said on 7 October 2026 that Haiku 5.5 is available in the Claude Platform and in Claude Code. @claudeai called it the cheapest, fastest, and most capable small model Anthropic has released. Anthropic's page uses that same description, and says Haiku 5.5 is its fastest model at each model's standard speed, while Opus models in Fast Mode still run more quickly. Anthropic's page says Haiku 5.5 is available on Amazon Web Services, Google Cloud, and Microsoft Azure, and that Claude Platform developers start with claude-haiku-5-5. A companion @claudeai post names those same three clouds.
For prompts under 100K tokens, ClaudeDevs listed $0.10 / $0.50 per million tokens, with cache reads at $0.01. Above 100K tokens, ClaudeDevs listed $0.50 / $2.50, with cache reads at $0.05. Anthropic's table labels those pairs as input and output for prompts up to 100,000 tokens and over 100,000 tokens, and adds cache writes at $0.125 and $0.625. A cache write is the charge for storing prompt tokens for reuse. A cache read is the lower charge for using those stored tokens on a later call.
Anthropic's around-75% average running-cost cut versus Haiku 4.5 combines a 90% list-price cut on the shorter band, a 50% cut on the longer band, a mix in which 90% of Haiku 4.5 requests were in the shorter band, and a tokenizer that uses slightly more tokens per task.
Sonnet 5.5 cache reads fell from $0.20 to $0.10 per million tokens. Anthropic estimates about 20% lower cost on most agentic tasks from that cut. @claudeai described it as about 20% cheaper on most long-running work.
| Rate | Haiku 5.5 up to 100k | Haiku 5.5 over 100k | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 | $2.00 |
| Output | $0.50 | $2.50 | $5.00 | $10.00 |
| Cache reads | $0.01 | $0.05 | $0.10 | $0.10 |
| Cache writes | $0.125 | $0.625 | $1.25 | $2.50 |
Who it is for
Anthropic positions Haiku 5.5 for summaries, compactions, database queries, and classification, and as a subagent with Opus 5.5 or Sonnet 5.5. On the page, compaction sits with summarization among the narrower jobs, beside the longer agentic coding Anthropic still assigns to Sonnet 5.5 and Opus 5.5, such as Terminal-Bench 4.0. Its table lists Haiku 5.5 at 39.2% on that benchmark. Classification appears on Anthropic's page. ClaudeDevs named summaries, compactions, database queries, and the subagent pairing.
The help center limits the monthly credit to active Max and Team plans, including discounted nonprofit and scientist Team plans. Free, Pro, and Enterprise are ineligible. In the credit table, Max 5x is the $100 level and Max 20x is the $200 level. Team Standard is $20 per seat and Team Premium is $100 per seat. Seats pool into one balance capped at $500, calculated from the seats on the plan when credits are granted. Three Standard seats and two Premium seats equal $260. ClaudeDevs stated Max 5x $100, Max 20x $200, and Team up to $500, pooled.
The credit covers any available Claude model on the Messages API, the Message Batches API, the Console Playground, Claude Managed Agents, and the Claude Agent SDK. Interactive Claude Code in the terminal, IDE, desktop, or web, extra usage in Claude, Claude Code, or Claude Cowork, and Claude on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry sit outside it. ClaudeDevs said the credits work on any model, including Haiku 5.5, in your code or third-party harnesses. The help center says an Agent SDK credit announced in June is a different offer and is not available.
How to try it
On the Claude Platform, Anthropic says developers start with claude-haiku-5-5. ClaudeDevs says the model is in Claude Code as well.
Subscribe on the web, iOS, or Android. After seven days on an eligible plan, claim in a browser on claude.ai: Settings > Billing as the Max subscriber, or Organization settings > Billing as a Team Owner or Primary Owner. Link one Console organization. The Console role is Owner, Admin, or Billing, and the flow can create an organization. No Claude Platform card is required. A plan can link only one organization, and each organization can receive credits from only one plan. Changing the link later means contacting support.
Credits refresh each billing cycle and arrive monthly on annual plans. Unused credits expire. Monthly credits are spent before purchased credits. API keys in the linked organization share the balance. Purchased credits or auto-reload keep requests going after the grant runs out. With neither, API requests stop until the next grant. Organizations invoiced through Anthropic sales are billed as usual past the credit. Usage is never charged to the Claude plan.
Upgrading from Max 5x to Max 20x brings prorated credits immediately, then $200 each cycle. Moving from Max to Team ends the Max link, and a Team Owner or Primary Owner claims after the Team plan has been active for seven days. Canceling, downgrading to an ineligible plan, or a refund stops new credits. Credits already granted stay usable until they expire. The help center says the offer is still rolling out over a few days. Anthropic's launch page said this week.
What is still open
Haiku 5.5 is the first Haiku-class model Anthropic describes with an adjustable effort setting, so a user can trade cost against intelligence. The page text does not number each effort level. The launch page does not print Bedrock, Vertex AI, or Foundry model ids, and points to a migration guide for details. Anthropic says the Claude Python and TypeScript SDKs are being updated with computer use and browser use in beta.
Cyber safeguards are described as more restrictive than Haiku 4.5 and somewhat less restrictive than other recent models. They allow a wider range of defensive tasks than the safeguards for Sonnet 5.5, and they still block penetration testing and other techniques more likely to be used by attackers. Biology safeguards are described as the same as for Sonnet 5, Sonnet 5.5, and Opus 5: research biology questions are allowed, and requests judged likely to cause harm stay restricted. Organizations working on wider-ranging biology and cyber activities can apply to the Life Sciences Verification Program and the Cyber Verification Program. Early comments from Asana, HubSpot, AlphaSense, Box, Rogo, and Cognition are reports Anthropic relayed, not an independent test.
Source
Also from
Sources and editorial approach
This brief uses the ClaudeDevs post from 7 October 2026, two @claudeai posts from the same hour, Anthropic's Claude Haiku 5.5 page, and the Claude Help Center article on monthly API credits for Max and Team plans, checked 8 October 2026. List prices, the model id claude-haiku-5-5, availability on Amazon Web Services, Google Cloud, and Microsoft Azure, and the credit rules are what those pages say. Claude Code availability is the ClaudeDevs post. Anthropic's availability line names those three clouds and the Claude Platform id. The Claude Code changelog checked the same day does not mention Haiku 5.5, and its notes stop mid-entry, so the changelog does not confirm Claude Code. The roughly 75% and 20% savings, the speed comparison, benchmark scores, and customer comments are Anthropic's figures. The system card, the cloud migration guide, and the Platform documentation for the computer-use beta were not covered.