Claude Haiku 5.5: every published spec, where it's free, and what happens to your prompt
The Morse teamPublished 7 min read
Claude Haiku 5.5 is Anthropic's small, fast model, released on October 7, 2026, and you can use it free in the Claude apps or pay from $0.10 per million input tokens and $0.50 per million output tokens on the API. As of November 5, 2026, it reads up to 1M tokens, writes up to 128K tokens in one reply, has a reliable knowledge cutoff of June 2026, and its per-token prices rise fivefold for any request whose prompt is longer than 100,000 tokens.
We read Anthropic's announcement, the Haiku 5.5 model page, the models overview, the price list, the 144-page system card, the Claude plans page, two support articles, the privacy center's retention and training pages, the API data retention docs, and Anthropic's consumer and commercial terms, all on November 5, 2026. Every benchmark and safety figure here is Anthropic's; we haven't rerun them. Anthropic deserves credit for publishing the system prompt Haiku 5.5 runs under in claude.ai and the Claude apps, with each revision dated (October 7 and October 26, 2026 so far).
The Claude Haiku 5.5 spec sheet
Every number Anthropic publishes for the model, as of November 5, 2026, each linked to the page that states it.
| Spec | Claude Haiku 5.5 | Where Anthropic says it |
|---|---|---|
| Released | October 7, 2026 | Model page |
| Model ID | claude-haiku-5-5 everywhere except Amazon Bedrock, which uses anthropic.claude-haiku-5-5 | Models overview |
| Context window | 1M tokens, "roughly 555k words or 2.5M Unicode characters" | Models overview |
| Longest reply | 128K tokens; 300K on the Message Batches API with a beta header | Model page |
| Reads and writes | Text and images in, text out | Model page |
| Knowledge cutoff | June 2026, for both reliable knowledge and training data | Model page |
| Thinking | Adaptive, on by default, effort medium by default | Model page |
| Effort levels | Low, medium, high, xhigh, and max, a first for a Haiku model | Announcement |
| Latency | "Fastest" in the current lineup | Models overview |
| Input price, per 1M tokens | $0.10; $0.50 for prompts over 100,000 tokens | Price list |
| Output price, per 1M tokens | $0.50; $2.50 for prompts over 100,000 tokens | Price list |
| Cache reads | $0.01; $0.05 over 100,000 tokens | Price list |
| Cache writes | $0.125 for 5 minutes, $0.20 for 1 hour; $0.625 and $1 over 100,000 tokens | Price list |
| Batch API | $0.05 in, $0.25 out; $0.25 and $1.25 over 100,000 tokens | Price list |
| What counts toward 100,000 | Every input token, cache reads and cache writes included; each request is priced on its own | Price list |
| US-only inference | 1.1 times every price, including the higher tier | Price list |
| Tool use overhead | 286 extra system prompt tokens (406 when tool_choice is any or tool) | Price list |
| Tokenizer | About 30 percent more tokens than Haiku 4.5 for the same text | What's new |
| Context in the Claude apps | 1M tokens in chat and Claude Code, 500K in Cowork, on paid plans | Support |
| Prompt injection | Attackers succeed 7.1% of the time within 15 tries, against 83.2% for Haiku 4.5 | System card |
| Retirement | Not sooner than October 7, 2027 | Model page |
Anthropic's model page lists six places to call it: the Claude API, Amazon Bedrock, Claude Platform on AWS, Microsoft Foundry, and Claude Models on Google Cloud in both its Google-managed and Anthropic-managed versions. Anthropic describes the model as built "for high-volume, latency-sensitive tasks such as classification, extraction, and routing," and its Haiku page adds chat, voice agents, subagents, and browser automation.
Is Claude Haiku 5.5 free?
Yes, in the Claude apps, within limits Anthropic doesn't put a number on. Its Haiku page says users on every Claude plan, Free included, can pick Haiku 5.5 on the web, iOS, and Android. As of November 5, 2026, the plans page says the Free plan includes Sonnet and Haiku but not Opus or Fable, and that its usage is limited every 5 hours and every week, with no message count.
On the API, Anthropic bills it per token. Its October 7 announcement says Max and Team subscribers now get a monthly API credit usable on any model, $100 on Max and up to $500 pooled across a Team, so for them Haiku 5.5 calls come out of money they've already paid. On Morse, Claude Haiku 5.5 is on the Free plan: 15 messages a day shared with GPT-6 Luna, GLM 5.3 Flash, and more, with no card and no Anthropic account.
What is Claude Haiku 5.5's context window?
It's 1M tokens on the API and on paid Claude plans, and Anthropic doesn't publish a figure for the Free plan. Anthropic's support article on context windows covers paid plans only, from Pro up, and lists Haiku 5.5 at 1M tokens in chat and Claude Code but 500K in Cowork. A second article, on usage and length limits, says "on paid plans, the newest models support up to a 1M token context window," and neither page we read states the Free plan's window.
Haiku 4.5 has 200K tokens of context and 64K of output (Anthropic). Thinking tokens count toward the reply limit, so a small max_tokens setting can end a reply after the model thinks and before it writes any text.
How much does Claude Haiku 5.5 cost?
From $0.10 per million input tokens and $0.50 per million output tokens, until the prompt passes 100,000 tokens. Above that, every token in the request costs five times as much. By our arithmetic from the price list, a 99,000-token prompt with a 1,000-token reply costs about $0.0104, and a 101,000-token prompt with the same reply costs about $0.053. Our GPT-6 Luna vs Claude Haiku 5.5 comparison works out costs at larger prompt sizes against OpenAI's model.
Haiku 4.5 costs $1 in and $5 out, and the announcement puts Haiku 5.5 at "around 75% less to run" on average: 90% cheaper for requests up to 100,000 tokens, 50% cheaper above, after allowing for a tokenizer that uses "slightly more tokens per task."
Is Claude Haiku 5.5 safe?
By Anthropic's own tests, it resists prompt injection far better than any earlier Haiku and turns down fewer harmless questions than Haiku 4.5. The figures below come from its system card, dated October 7, 2026.
Which ASL level is Claude Haiku 5.5?
The system card doesn't give it an ASL label. It measures the model against thresholds in Anthropic's Responsible Scaling Policy instead, and says Haiku 5.5 "does not cross our CB-2 or Autonomy-2 RSP thresholds" but is treated as meeting CB-1 and Autonomy-1, with the matching safeguards applied. Anthropic calls it "broadly less capable than Claude Opus 5 across domains." Our GPT-6 Luna comparison covers how its safety classifiers decline requests.
How well does it resist prompt injection?
Much better than Haiku 4.5. On Gray Swan's indirect prompt injection benchmark, an attacker with 15 tries succeeded 7.1% of the time, against 83.2% for Haiku 4.5. Haiku 5.5's rate was 0.2% in coding, 4.0% in tool use, and 24.4% in computer use through a screen, where most of its remaining weakness sits. Sonnet 5.5 was at 3.4% and Opus 5.5 at 1.0%. Against Gray Swan's adaptive Shade attacker in coding tasks, 0.08% of attempts worked, and none did with Anthropic's prompt injection probes switched on. In 110 browser scenarios run through Claude Cowork, no attack succeeded.
Does it refuse a lot?
Less than Haiku 4.5, and less on the API than in claude.ai. On benign but sensitive questions, Haiku 5.5 refused 0.17% on the API without a system prompt and 0.82% with the claude.ai system prompt, against 0.44% and 3.05% for Haiku 4.5. In the card's broader automated behavioral audit, though, Haiku 5.5 "over-refused more than any other model we tested." The same card calls it "at least as honest under pressure as every other model we tested," though it used a leaked answer without telling the user more often than Haiku 4.5 did.
Who keeps your Claude Haiku 5.5 prompts?
That depends on where you call the model. As of November 5, 2026, Anthropic's API data retention docs split it three ways:
| Where you use Haiku 5.5 | Who processes the prompt | Zero data retention |
|---|---|---|
| Claude API, Claude Platform on AWS, Microsoft Foundry, and Anthropic-managed Google Cloud | Anthropic | Available per organization on request, because Haiku 5.5 isn't one of the "Covered Models" (the Fable and Mythos models) that require 30-day retention |
| Amazon Bedrock and Google-managed Google Cloud | The cloud provider | Set by that provider's own retention and compliance docs |
| claude.ai and the Claude apps, Free plan included | Anthropic, under the consumer rules | Not offered; Anthropic's ZDR doesn't cover the Free, Pro, or Max plans |
In the apps, Haiku 5.5 follows the same rules as any Claude chat: the training switch, the five-year window for chats used in training, and the two-year copies of flagged chats (Anthropic). Our guide to whether Claude is private covers each of them.
Claude Haiku 5.5 on Morse's Free plan
On Morse, Claude Haiku 5.5 is one of the Free plan's models, with no card and no Anthropic account to open. Claude chats run in Anonymous mode: our proxy strips your account, email, and IP address and sends the message under Morse's own account, so the provider reads it without knowing it's you, though it may keep it for a limited time to check for abuse, under its own policy. Your history stays only in your browser, encrypted. Morse doesn't train AI on your content, sell your data, or show ads. For prompts no company should read, Pro at $19 a month adds E2EE chats with open-weight models, whose guarantee rests on the enclave hardware and attestation working as designed.
Send a message from the Morse home page without an account, or start free for 15 messages a day with Claude Haiku 5.5 and no card.
Common questions
Can you set temperature on Claude Haiku 5.5?
No. Anthropic's notes on what changed say temperature, top_p, and top_k can return an error and should be left out. Prefilling the assistant's reply also returns an error, so a request has to end with a user turn.
When will Claude Haiku 5.5 be retired?
Not sooner than October 7, 2027, on the platforms Anthropic runs itself. Amazon Bedrock and the Google-managed version on Google Cloud set their own dates.
Can Claude Haiku 5.5 make images?
No. It reads text and images but answers in text, and its system card says plainly that "the model outputs text only."
Does prompt caching work on short prompts with Haiku 5.5?
Yes, from 512 tokens. Haiku 4.5 needed at least 4,096 tokens before a prompt could be cached, so prompts between those sizes now get the cache-read price.