Claude Haiku 5.5 Launches: Up to 90% Cheaper, 1M Context and the First Haiku with Effort — What It Is For, What Changes from Haiku 4.5, and Whether You Must Move Now

Nine days after Claude Sonnet 5.5 launched, Anthropic released Claude Haiku 5.5 on Wednesday 7 October 2026, completing all three members of the 5.5 family within two weeks. The news this time is not the scores but the price tag. Where Haiku 4.5 charged $1 / $5 per million tokens, Haiku 5.5 starts at $0.10 / $0.50, a 90% cut for ordinary requests, with a 1-million-token context window and, for the first time on a Haiku, an effort setting. Thai organizations are asking two questions. First, which workloads should move down from Sonnet to this model? Second, do systems still running on Haiku 4.5 really have to migrate before 15 October? This article answers both, and corrects a mistake about the retirement date that we made ourselves.

1. Where Haiku 5.5 Sits in the Claude Lineup, and Why Haiku 4.5 Is Not Shutting Down on 15 October

Anthropic's "current lineup" page now lists four models. Haiku 5.5 takes Haiku 4.5's place at the bottom as what Anthropic calls its "cheapest, fastest, and most capable small model", aimed at high-volume work where response time matters. What is different from the Haiku 4.5 generation is that the API (Application Programming Interface) specification now matches the larger models on every line: context, output, thinking and effort. Only the price and the scores differ.

Source: Claude Haiku 5.5 overview and Model deprecations pages, platform.claude.com (11 Oct 2026)

The retirement date needs to be stated clearly, because in our Sonnet 5.5 article we wrote that Haiku 4.5 "is scheduled to retire on 15 October 2026". That was a misreading of the documentation. Anthropic's table lists Haiku 4.5 as Active, and 15 October sits in the "not retired before" column: a commitment not to shut it down before that date, not a shutdown date. Anthropic's policy is to give at least 60 days' notice before retiring any publicly released model. The most recent example is Sonnet 4.5, notified on 30 September 2026 for retirement on 30 November 2026. At the time of writing, no such notice exists for Haiku 4.5. Systems on Haiku 4.5 will not fail next week. They should start migration testing now, though, because once the notice arrives there will be two months, and this migration involves more than a one-line model ID change (section 5).

Another difference from Haiku 4.5 that hits cost directly is the tokenizer. Haiku 5.5 uses the newer tokenizer shared by every model from Opus 4.7 onward, so the same text counts roughly 30% more tokens, just as happened when Sonnet 5 launched. The 90% price cut therefore needs that adjustment before it goes into a budget. Large images also cost more: Haiku 5.5 uses the high-resolution image tier, so a 2,000 × 1,500 pixel image costs about 2.5 times as many visual tokens as on Haiku 4.5.

2. Pricing: 90% Cheaper Is Real, but Two Conditions Come First

Haiku 5.5 is Anthropic's first model priced in two tiers by prompt length. Requests whose prompt is 100,000 tokens or less are 90% cheaper than Haiku 4.5 on every line. Requests whose prompt exceeds 100,000 tokens are 50% cheaper. Anthropic says about 90% of requests through its platform fall into the first group, and that the blended saving is "around 75%".

Source: Pricing page, platform.claude.com (11 Oct 2026) · The >100K group is 50% below Haiku 4.5 on every line · US-only inference via inference_geo carries a 1.1× multiplier on both tiers

The first condition is how prompt length is counted. Anthropic counts every input token of the request, including tokens read from cache and tokens written to cache. A request that crosses 100,000 pays the higher tier for the whole request, even when most of the prompt is a cache hit, and each request is priced on its own, so earlier requests under the line keep their price. Workloads whose prompts sit near the line should be designed to stay under 100,000. On the new tokenizer, 1 million tokens is roughly 555,000 English words, so the line sits near 55,000 words. Thai text uses more tokens per word, so count it with count_tokens and the model set to Haiku 5.5, not with numbers measured on Haiku 4.5.

The second condition is the 30% tokenizer increase from section 1. Take the same body of text that cost 1 million input tokens on Haiku 4.5: on Haiku 5.5 it costs about $0.13 instead of $1.00, a saving of about 87% rather than 90%. Still an excellent number, but the one a budget should use.

Two things Haiku 5.5 does not have: fast mode (still limited to Opus) and Priority Tier. An organization with a Priority Tier commitment on Haiku 4.5 must plan capacity separately, because the commitment cannot be carried over.

3. Where It Is Strongest: Eight Benchmarks, Plus Independent Numbers by Effort Level

Anthropic reports eight benchmarks against Haiku 4.5, Sonnet 5.5 and OpenAI's GPT-6 Luna, its direct small-model competitor. The picture: Haiku 5.5 beats GPT-6 Luna on every row where both have a score, closes most of the gap between Haiku 4.5 and Sonnet 5.5 on office work, and still trails Sonnet 5.5 clearly on terminal-based coding.

Source: Anthropic, Introducing Claude Haiku 5.5 (7 Oct 2026) · The main table in the announcement does not state the effort level per row; separate "accuracy versus cost" charts show scores moving across all five effort levels

The two rows that matter most to an organization are the first two. GDPval-AA and AA-Briefcase measure the work office staff actually do: writing reports, building spreadsheets, reading documents. Haiku 5.5 moves from 735 to 1,620, closing about 80% of the gap between Haiku 4.5 and Sonnet 5.5 at one twentieth of Sonnet 5.5's input price. The rows to be careful with are Terminal-Bench and FrontierCode: multi-step coding agents still belong to Sonnet 5.5. The OSWorld jump from 15.7% to 72.4% is why Anthropic is, for the first time, promoting browser use and computer use on a small model.

Customer figures in the announcement point the same way. Asana measured over 30% lower latency on task completions and up to 2.5× faster inference per agent turn. HubSpot recorded its best internal suite score yet at 92.8%. Box scored 11 points above Haiku 4.5 at about half the latency, and AlphaSense measured 0.84 against Haiku 4.5's 0.76. These are each company's own test suites: useful for direction, not for comparing against one another.

The independent measurement from Artificial Analysis, published on 8 October, adds the effort dimension that Anthropic's table leaves out.

Source: Artificial Analysis, Haiku 5.5 medium versus xhigh comparison, and The Decoder (8 Oct 2026) · Figures may change as Artificial Analysis adds measurement runs

Two observations. First, output speed is similar at every effort level, but time to first answer differs by more than 4×, because at higher effort the model thinks longer before answering. Work where a user is waiting at a screen belongs at medium or below. Second, running the index at xhigh costs almost 3× what medium costs, for 7 extra points. That is exactly the range where Anthropic itself recommends comparing against Sonnet 5.5 before deciding.

4. Where You Can Use It: claude.ai, Claude Code, the API and the Three Clouds

The announcement says Haiku 5.5 is "available now on all platforms", naming Amazon Web Services, Google Cloud and Microsoft Azure, but it says nothing about claude.ai plans, unlike the Opus 5.5 and Sonnet 5.5 announcements. The web and Claude Code information below therefore comes from third-party reporting.

Source: overview and migration guide pages, platform.claude.com · claude.ai and Claude Code details from Apidog (Oct 2026), which is not an official source

For organizations that procure Claude Team or Enterprise through Grand Linux, this news does not change the per-seat price. Web users gain a faster small model as an option. Where the bill actually moves is the API side: organizations that built automation on Haiku 4.5, or that had to use Sonnet because Haiku 4.5 could not handle the job. The price structure for each channel is on the Claude pricing page.

5. For Development Teams: Six Things That Break on Haiku 4.5 Code, and Two Things Haiku Allows That Its Siblings Do Not

Moving from Haiku 4.5 is more work than moving from Sonnet 5 to Sonnet 5.5, because Haiku 4.5 still sits on the old parameter set: budget_tokens, temperature and prefill, all of which Haiku 5.5 rejects. Anthropic groups the changes three ways: breaking (the request fails), changed (the request succeeds but behaves differently) and new.

Source: What's new in Claude Haiku 5.5 and the migration guide, platform.claude.com (11 Oct 2026)

Two places where Haiku 5.5 is more lenient than Opus 5.5 and Sonnet 5.5, which may surprise teams that just migrated those. First, thinking can be turned off with {"type":"disabled"} at low, medium and high effort (xhigh and max return a 400); no between_tools workaround as on Sonnet 5.5. Second, forced tool use with tool_choice set to any or a named tool still works, except that the response then starts with the tool call and carries no thinking block. To let the model think before calling, use auto and say so in the prompt. On cross-model thinking blocks, Haiku 5.5 reads blocks from Sonnet 5, Opus 4.8 and Haiku 4.5, so a conversation switched mid-way from Haiku 4.5 keeps its reasoning, but it cannot read blocks from Opus 5, Opus 5.5, Sonnet 5.5 or the Fable family. In the other direction, Opus 5.5 and Sonnet 5.5 can read Haiku 5.5's blocks on the Claude API and Google Cloud, which matters for pipelines where Haiku reads first and a larger model decides.

The minimum migration for code that had thinking enabled on Haiku 4.5 looks like this:

# Before: Haiku 4.5 with budget_tokens and a tuned temperature
response = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=4096,
    temperature=0.2,
    thinking={"type": "enabled", "budget_tokens": 2048},
    messages=[...],
)

# After: Haiku 5.5 with adaptive + effort, temperature removed, room for thinking
response = client.messages.create(
    model="claude-haiku-5-5",
    max_tokens=16000,                     # thinking tokens count toward max_tokens
    thinking={"type": "adaptive"},        # or omit thinking; adaptive is the default
    output_config={"effort": "low"},      # start high-volume short tasks at low, then measure
    messages=[...],
)
answer = next(b.text for b in response.content if b.type == "text")  # select by type, not content[0]

One more word on max_tokens. Thinking tokens count toward max_tokens even when the thinking text is not returned. A limit of 256 or 512 that was fine for classification on Haiku 4.5 can now end the request with stop_reason: "max_tokens" after the thinking block and before any answer text. Raise max_tokens or lower effort. Teams already on Claude Code can run /claude-api migrate this project to claude-haiku-5-5 to work through the whole project; Anthropic includes that instruction on its own migration guide page.

6. The Five Effort Levels: Which Setting for Which Job

Effort is the main control on Haiku 5.5, replacing the old budget_tokens. The API default is medium (Sonnet 5.5 defaults to high). Anthropic gives per-level guidance on its prompting page and says most existing Haiku 4.5 prompts work without changes.

Source: Prompting Claude Haiku 5.5, platform.claude.com (11 Oct 2026)

Three techniques from the same page that matter for organizational work. First, if the model has a web or knowledge-base search tool, always put today's date in the system prompt; Anthropic's testing found that the date alone grounded answers in recent results. Second, a customer-service chatbot should carry a sentence stating that the system prompt's rules hold for the whole conversation, even when a user argues, asks for just a small part, or claims someone approved an exception, because a small model is easier to talk around; where this matters most, use high. Third, changing the top-level effort between requests invalidates the prompt cache for the whole conversation; use the per-message effort change (beta), which keeps the cache.

7. Safeguards: Refusals That Haiku 4.5 Never Had

Haiku 5.5 is the first Haiku with the same safety classifiers as the larger models. A declined request comes back as HTTP 200 with stop_reason: "refusal" and a stop_details.category naming one of four categories: cyber (could enable malware or exploit development; finding vulnerabilities in source code is still allowed), frontier_llm (could assist development of competing AI models), bio (dangerous biological methods; everyday health questions are unaffected) and general_harms (other usage-policy areas). Anthropic says Haiku 5.5's cyber safeguards are more restrictive than Haiku 4.5's but somewhat less restrictive than other recent models, while its bio safeguards match Sonnet 5 / 5.5 and Opus 5.

8. Haiku 5.5 or Sonnet 5.5: A Decision Rule

The price differs by 20× on both input and output for requests under 100,000 tokens, but the score gap differs by task. The rule we use is to look at the task type first and the prompt length second.

The conclusion Anthropic writes itself, and which matches our experience: if you would run Haiku 5.5 at xhigh or max, run the same evals on Sonnet 5.5 at low or medium too. A mid-tier model thinking less often scores close, in less time, and sometimes for less money. Haiku 4.5, meanwhile, should no longer be the answer for any new work; Haiku 5.5 wins every row that has a score, at a lower price.

9. The ERP View: Where a Small Model Fits in an Integration

In the Claude integrations Grand Linux builds against back-office systems, the step that generates the most requests is not answering hard questions. It is the step before: reading what a user sent and deciding whether it is a how-to question, an error report or a data request; pulling a document number or an agency code out of free text; summarizing a long report for the main model to read. That work is well-bounded, short and repeated all day. Until now we ran it on Sonnet because Haiku 4.5 could not handle multi-step tasks. Haiku 5.5's OSWorld and Terminal-Bench numbers say a small model can now take on multi-step work. We will test that against real jobs before moving anything, not on the strength of a table.

Saeree ERP has no AI feature built in, and that is deliberate. The ERP (Enterprise Resource Planning) system is the source of truth, running on the customer's own servers. What Grand Linux does is connect Claude to ERP data through MCP (Model Context Protocol), with permissions by user role and an audit log of every call. In that architecture the model can be chosen per tool: a screening tool on Haiku 5.5, a tool that interprets business rules on Sonnet 5.5, without users needing to know which one is behind the answer. Read about the idea in What is MCP and the integration formats at Claude Solutions.

Data needs to be stated separately. ERP data stays on the agency's server or the cloud the agency chose. Claude runs on Anthropic's cloud or the cloud provider selected. Haiku 5.5's 1-million-token context makes "send the whole report and ask" cheap, which does not make it right. An MCP integration still has to decide which data may leave to be asked about and which stays in-house, and a request that ships an entire report crosses the 100,000-token line and pays the higher tier at once. Sending only what the question needs is cheaper on both counts, money and data.

Conclusion

• Claude Haiku 5.5 launched on 7 October 2026 at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens, 90% below Haiku 4.5; longer prompts pay $0.50 / $2.50, 50% below. Anthropic's blended estimate is about 75% cheaper.
• The specification matches the larger models: 1M context, 128K output, adaptive thinking, five effort levels (default medium), knowledge through June 2026. The new tokenizer counts about 30% more tokens, which budgets must subtract first.
• Large gains on office work (GDPval-AA 735 → 1,620) and computer use (OSWorld 15.7% → 72.4%); beats GPT-6 Luna on every row; still well behind Sonnet 5.5 on agentic coding.
• Haiku 4.5 remains Active with no retirement notice. 15 October is the earliest possible date, and Anthropic must give 60 days' notice. Start migration testing now anyway, because there are six breaking changes: budget_tokens, temperature, prefill, computer use, structured outputs on Bedrock, and conversation-history edits.
• Haiku 5.5 is more lenient than its siblings in two ways, thinking can be disabled at high or below and forced tool_choice works, but its safeguards can decline requests with no server-side fallback, so clients need their own escape route.
• Decision rule: classification, screening, extraction, summarization and subagents start on Haiku 5.5 at medium; long office work and coding agents go to Sonnet 5.5; before pushing Haiku to xhigh, compare it with Sonnet 5.5 at low.

A small model ten times cheaper does not replace the mid-tier model. It replaces the work we were paying mid-tier prices for without needing to.

- Saeree ERP team

References

• Anthropic — Introducing Claude Haiku 5.5 (7 Oct 2026)
• Claude Platform Docs — Claude Haiku 5.5 overview
• Claude Platform Docs — What's new in Claude Haiku 5.5
• Claude Platform Docs — Claude Haiku 5.5 migration guide
• Claude Platform Docs — Prompting Claude Haiku 5.5
• Claude Platform Docs — Model deprecations
• Claude Platform Docs — Pricing
• Artificial Analysis — Claude Haiku 5.5 (Medium) vs Claude Haiku 5.5 (Xhigh)
• The Decoder — Claude Haiku 5.5 arrives with massive price cuts (8 Oct 2026)
• heise online — Claude Haiku 5.5: Anthropic cuts API prices by up to 90 percent
• Apidog — How to Use Claude Haiku 5.5 in Claude Code (and as Your Subagent Model)