AIHow AI Works

Claude Haiku 5.5

Legacy model. Claude Haiku 5.5 replaced Haiku 4.5 as the current Haiku on October 7, 2026. Haiku 4.5 is still available on the API and in Claude Code;…

Legacy model. Claude Haiku 5.5 replaced Haiku 4.5 as the current Haiku on October 7, 2026. Haiku 4.5 is still available on the API and in Claude Code; it leaves the claude.ai model selector on October 29, 2026.

At a glance

Field Value
Developer Anthropic
Model type Language Model
Status Legacy (Active on the API)
Successor Claude Haiku 5.5
Context window 200K tokens
Input price $1
Output price $5
Knowledge cutoff February 2025
License Proprietary
Released October 15, 2025

Overview

Claude Haiku 4.5, released October 15, 2025, was Anthropic’s fast, low-cost tier for almost a year. On October 7, 2026 Claude Haiku 5.5 replaced it, and Anthropic now lists Haiku 4.5 as a legacy model that is still available.

(adsbygoogle = window.adsbygoogle || []).push({});

At launch its pitch was simple: coding roughly on par with Claude Sonnet 4 at one-third the cost and more than twice the speed, and better than Sonnet 4 at computer use. That made near-frontier output affordable for real-time chat, high-volume classification and sub-agent fan-out.

It is now two generations behind its siblings. The Opus and Sonnet tiers moved through 4.6 and 5 to Claude Opus 5.5 and Claude Sonnet 5.5, with Claude Fable 5.1 above them. The Haiku line jumped straight from 4.5 to 5.5. For new work, Haiku 5.5 is cheaper per token on most prompts; Haiku 4.5 mainly matters to teams with a tested pipeline that depends on its exact behavior.

(adsbygoogle = window.adsbygoogle || []).push({});

Status & lifecycle

Haiku 4.5 is legacy but not deprecated: the API still serves claude-haiku-4-5-20251001, and no retirement date has been set.

Date Event
October 29, 2026 Removed from the claude.ai model selector (chat, Cowork, Design) on Free, Pro and Max; selected chats switch to Haiku 5.5. Claude Code and the API are unaffected.
October 15, 2026 Anthropic’s “not sooner than” retirement floor passes with no deprecation notice.
October 7, 2026 Claude Haiku 5.5 released; Haiku 4.5 moves to Legacy.
April 20, 2026 Claude 3 Haiku retired, with Haiku 4.5 named as its replacement.
February 19, 2026 Claude 3.5 Haiku retired, with Haiku 4.5 named as its replacement.
October 15, 2025 Claude Haiku 4.5 released.

How long will it last? Anthropic commits to at least 60 days’ notice before retiring a publicly released model. With no notice issued as of October 26, 2026, the earliest possible API retirement is late December 2026. Google Cloud lists its own floor of December 7, 2026.

Sources: Anthropic docs, Claude Help Center, Google Cloud, Layer3 Labs deprecation tracker.

Capabilities

  • Speed: More than twice the speed of Claude Sonnet 4 at launch, per Anthropic. Haiku 5.5 is now faster: Anthropic reports about half the latency in early testing.
  • Coding: 73.3% on SWE-bench Verified, under a heavy configuration (see Benchmarks).
  • Computer use: Beat Claude Sonnet 4 at browser navigation and tool use, which is why it powered latency-sensitive surfaces like Claude in Chrome.
  • Extended thinking: Manual thinking with a configurable token budget; Terminal-Bench rose from 40.21% to 41.75% with a 32K budget.
  • Vision: Accepts text and images, outputs text. Uses the standard image tier, which downscales above 1,568 pixels on the long edge.
  • Alignment: In Anthropic’s automated assessments, lower misaligned-behavior rates than Sonnet 4.5 and Opus 4.1. Deployed at ASL-2.

Technical Specifications

  • Context window: 200K tokens
  • Max output: 64K tokens per request
  • Thinking: Manual extended thinking, thinking: {"type": "enabled", "budget_tokens": N}. Minimum 1,024, and strictly less than max_tokens.
  • Effort parameter: Not supported; output_config.effort errors.
  • Input → output: Text and images → text
  • Reliable knowledge cutoff: February 2025
  • Training data cutoff: July 2025
  • Tokenizer: The pre-Opus 4.7 tokenizer. The same text uses about 30% fewer tokens than on Haiku 5.5.
  • Safety level: ASL-2
  • Architecture: Transformer-based, developed under Anthropic’s Constitutional AI approach

Model IDs by platform

Platform Model ID
Claude API claude-haiku-4-5-20251001 (alias claude-haiku-4-5)
Amazon Bedrock anthropic.claude-haiku-4-5
Amazon Bedrock (InvokeModel) anthropic.claude-haiku-4-5-20251001-v1:0
Google Cloud (Google managed) claude-haiku-4-5@20251001
Google Cloud (Anthropic managed) claude-haiku-4-5-20251001
Microsoft Foundry claude-haiku-4-5
Claude Platform on AWS claude-haiku-4-5

Source: Anthropic docs.

Supported features

Extended thinking, vision, prompt caching (minimum cacheable prefix 4,096 tokens), Message Batches, structured outputs, tool use, the memory tool, bash and text-editor tools, code execution, tool search, computer use (computer_20250124), and Priority Tier.

Not supported: adaptive thinking, the effort parameter, fast mode, task budgets, the browser use toolset, and mid-conversation system messages.

API Behavior

Haiku 4.5 predates the request-surface changes that arrived with Opus 4.7, so it behaves like the older generation.

  • Extended thinking, not adaptive. Thinking is off unless you enable it with a budget. A budget_tokens at or above max_tokens returns a 400.
  • Sampling parameters depend on the API version. On anthropic-version: 2023-06-01, Haiku 4.5 still accepts temperature, top_p and top_k. On API version 2026-10-21 (released October 21, 2026), all three return a 400 for every model, Haiku 4.5 included. New SDK majors send 2026-10-21 by default.
  • Bedrock caveat. Amazon Bedrock supports API version 2026-10-21 only for Claude Opus 4.7 and later. Upgrading your SDK while still calling Haiku 4.5 on Bedrock can produce 400s; pin bedrock-2023-05-31 or stay on the older SDK.
  • Assistant prefill works when thinking is off. Haiku 5.5 rejects prefill in every mode.
  • Summarized thinking by default. Haiku 4.5 returns readable thinking summaries; Haiku 5.5 returns empty thinking blocks unless you ask for "display": "summarized".
  • Separate rate-limit pool. Haiku 4.5 has its own limits. Check your tier before shifting traffic between Haiku versions.

Sources: Haiku 5.5 migration guide, API 2026-10-21 migration guide, Tokely on DEV.

Haiku 4.5 vs Haiku 5.5

Haiku 5.5 is cheaper on prompts under 100K tokens and more capable, but it changes the request surface enough that migration is more than a model-ID swap.

Claude Haiku 4.5 Claude Haiku 5.5
Released October 15, 2025 October 7, 2026
API ID claude-haiku-4-5-20251001 claude-haiku-5-5
Input / output per MTok $1 / $5 $0.10 / $0.50 up to 100K-token prompts; $0.50 / $2.50 above
Cache read per MTok $0.10 $0.01 ($0.05 above 100K)
Context / max output 200K / 64K 1M / 128K
Thinking Manual, off by default Adaptive, on by default
Effort Not supported low to max, default medium
Sampling params, prefill Accepted (on API 2023-06-01) Rejected
Tokenizer Older Newer; ~30% more tokens for the same text
Image tier Standard (1,568 px) High-res (2,576 px)
Knowledge cutoff February 2025 June 2026
Retirement floor Not sooner than October 15, 2026 (passed) Not sooner than October 7, 2027

What it costs in practice. A month of 50M input and 10M output tokens, all prompts under 100K, costs $100 on Haiku 4.5. On Haiku 5.5 the same text becomes about 65M and 13M tokens under the new tokenizer, costing about $13. Anthropic’s own figure is about 75% cheaper on average, which allows for long prompts billed at the higher rate.

Migration checklist (Haiku 4.5 → 5.5)

  1. Change the model ID (claude-haiku-5-5 on every platform, anthropic.claude-haiku-5-5 on Bedrock).
  2. Recount tokens with count_tokens against the new model; raise max_tokens to leave room for thinking.
  3. Replace {"type": "enabled", "budget_tokens": N} with {"type": "adaptive"} plus an effort level. Where you ran without thinking, start at low.
  4. Read content blocks by type, not position: responses can now open with thinking blocks.
  5. Remove temperature, top_p and top_k.
  6. Replace assistant prefill with structured outputs or a system-prompt instruction.
  7. Move computer use from computer_20250124 to the computer_toolset_20260801 toolset (Bedrock: computer_20251124).
  8. Keep conversations append-only when sending thinking blocks back, and handle stop_reason: "refusal".
  9. Plan separately if you rely on Priority Tier, which Haiku 5.5 does not support.

Sources: Haiku 5.5 migration guide, T-Minus AI pricing roundup, Unite.AI.

Use Cases

For new projects, start on Haiku 5.5. Haiku 4.5 still earns its place in a few specific situations.

When Haiku 4.5 is still the right call

  • A validated pipeline you cannot re-test yet. Classification or extraction jobs whose outputs are scored against a fixed eval set; a model swap means re-validating.
  • Code that depends on sampling control or prefill. temperature: 0 for repeatable outputs, or assistant prefill to force a format, both still work on API version 2023-06-01.
  • Hard token limits tuned for the old tokenizer. Fixed max_tokens caps or token-based quotas will cut off or overrun under Haiku 5.5’s ~30% heavier tokenizer until re-tuned. (Haiku 5.5 stays cheaper even on long prompts: a 150K-token prompt costs about $0.15 input on Haiku 4.5 and about $0.10 on Haiku 5.5.)
  • Priority Tier commitments. Haiku 5.5 does not support Priority Tier.
  • Bedrock on older SDKs. Teams pinned to bedrock-2023-05-31 for other pre-4.7 models.

What it was built for (and Haiku 5.5 now does better)

  • Real-time chat and customer-service agents where latency is visible
  • Sub-agent fan-out under an Opus- or Sonnet-tier orchestrator
  • High-volume classification, routing and extraction (model routing)
  • Browser navigation and computer use at scale
  • Fast inline completions and small refactors

Performance / Benchmarks

Anthropic published Haiku 4.5’s comparison table as an image, so only three scores exist as citable text, all in the launch post’s methodology notes. The system card is a safety document with no capability table.

Benchmark Claude Haiku 4.5 Setting
SWE-bench Verified 73.3% 128K thinking budget, averaged over 50 trials, two-tool scaffold, prompt nudging heavy tool use
Terminal-Bench (Terminus 2) 41.75% 32K thinking budget, 5 runs
Terminal-Bench (Terminus 2) 40.21% No thinking, 6 runs

The settings matter. The 73.3% used a 128K thinking budget and a system-prompt addition telling the model to test first and call tools often. A default, no-thinking call will score lower.

Two widely quoted numbers are not Anthropic’s. The “90% of Sonnet 4.5” figure comes from Augment’s own agentic coding eval. The “4-5x faster than Sonnet 4.5” line is a customer quote with no workload attached. Both appeared as testimonials in the launch post.

Against Haiku 5.5. Anthropic says Haiku 5.5 scored 11 points higher than Haiku 4.5 in early testing at about half the latency. Its internal ECI index puts Haiku 5.5 at 167.11. No ECI figure for Haiku 4.5 has been published.

Artificial Analysis lists two entries on Intelligence Index v4.1: Non-reasoning at 24 (flagged as estimated) and Reasoning at 30. Scores are not comparable across index versions.

All vendor figures here are upper bounds. Validate on your own eval set.

Sources: Haiku 4.5 announcement, Haiku 5.5 announcement, Unite.AI.

Limitations

  • Superseded. Haiku 5.5 is cheaper on most workloads and stronger at coding, computer use and knowledge work. Anthropic recommends migrating.
  • No retirement date yet, but one is coming. The October 15, 2026 floor has passed. Watch the deprecations page; you will get at least 60 days’ notice.
  • Leaving the Claude apps. Gone from the claude.ai model selector after October 29, 2026; continued chats answer with Haiku 5.5.
  • Smaller context window. 200K tokens, versus 1M on every current Claude model.
  • Lower max output. 64K tokens per request, versus 128K.
  • Oldest knowledge in the lineup. Reliable cutoff February 2025, versus June 2026 on Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1. Pair it with web search for anything recent.
  • No adaptive thinking or effort control. Reasoning depth is set with a raw token budget.
  • API version friction. Sampling parameters break on API version 2026-10-21, and Bedrock does not support that version for Haiku 4.5 at all.
  • No internet access unless you add tool use or the server-side web search tool.

Pricing & Access

API Pricing

Available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry.

Item Price per MTok
Input $1
Output $5
5-minute cache write $1.25
1-hour cache write $2
Cache read $0.10
Batch API 50% off input and output

Source: Anthropic docs. For comparison, Haiku 5.5 lists at $0.10 / $0.50 for prompts up to 100K tokens.

Consumer Plans

Selectable under “More models” on Free, Pro and Max until October 29, 2026, then removed from chat, Claude Cowork and Claude Design on web, desktop and mobile. It remains available in Claude Code. Source: Claude Help Center.

Ecosystem & Tools

Official SDKs

  • Python SDK and TypeScript SDK, plus Java, Go, Ruby, C# and PHP, and the ant CLI. New SDK majors default to API version 2026-10-21; stay on the previous major if you need sampling parameters on Haiku 4.5.
  • REST API for direct HTTP access.

Developer Tools

  • Claude Console for testing prompts
  • Claude Code, where Haiku 4.5 stays selectable after it leaves the chat apps
  • The bundled Claude API skill in Claude Code can automate the move to Haiku 5.5

Community & Resources

Common questions

Is Claude Haiku 4.5 still the current Haiku?
No. Claude Haiku 5.5 replaced it on October 7, 2026, and Anthropic lists Haiku 4.5 as a legacy model.

Is Claude Haiku 4.5 being retired?
Not yet. It is still Active on the API with no deprecation notice. Its “not sooner than October 15, 2026” floor has passed, and Anthropic gives at least 60 days’ notice, so it will run until at least late December 2026.

Can I still use Haiku 4.5 in the Claude app?
Until October 29, 2026, under “More models”. After that it leaves the selector on Free, Pro and Max, and those chats continue on Haiku 5.5. It stays available in Claude Code and on the API.

What is the model ID for Claude Haiku 4.5?
claude-haiku-4-5, which resolves to the pinned snapshot claude-haiku-4-5-20251001. On Bedrock it is anthropic.claude-haiku-4-5.

How much does Claude Haiku 4.5 cost?
$1 per million input tokens and $5 per million output tokens, with cache reads at $0.10 and a 50% Batch discount. Haiku 5.5 costs $0.10 / $0.50 for prompts up to 100K tokens.

Should I switch from Haiku 4.5 to Haiku 5.5?
For most workloads, yes: it is cheaper even after its heavier tokenizer, has a 1M context window and a June 2026 knowledge cutoff. Budget time for the breaking changes: thinking configuration, sampling parameters, prefill and computer-use tool version.

What is the context window for Claude Haiku 4.5?
200K tokens, with up to 64K output tokens per request.

Does Claude Haiku 4.5 support adaptive thinking or effort?
No. It uses manual extended thinking with budget_tokens (at least 1,024 and below max_tokens). Passing effort errors.

Does temperature still work on Haiku 4.5?
On API version 2023-06-01, yes. On API version 2026-10-21, which new SDK majors send by default, temperature, top_p and top_k return a 400 for every model.

What is the knowledge cutoff for Claude Haiku 4.5?
Reliable cutoff February 2025, training data through July 2025. Every current Claude model has a June 2026 cutoff.

What safety level is Claude Haiku 4.5 deployed under?
ASL-2.

  • Claude Haiku 5.5: Anthropic’s current small model. $0.10 / $0.50 up to 100K-token prompts, 1M context, adaptive thinking with effort.
  • Claude Sonnet 5.5: Anthropic’s current mid-tier, released September 28, 2026 at $2 / $10.
  • Claude Opus 5.5: Anthropic’s current Opus, released September 22, 2026 at $4 / $20.
  • Claude Fable 5.1: Anthropic’s most capable widely released model, $10 / $50.
  • Gemini 3.8 Flash: Google’s fast model, 1M context.