Claude Haiku: the full family history

Claude Haiku is the small, fast, cheap tier. Every Haiku release, the platform model IDs, latency and price progression, and which versions are retired.

Haiku is the small, fast, cheap tier of the Claude range, and the current member is Claude Haiku 4.5 - $1 per million input tokens, $5 per million output tokens, a 200,000-token context window and 64,000 maximum output tokens. It is the only Haiku version still callable: Haiku 3.5 was retired on 19 February 2026 and Haiku 3 on 20 April 2026.

The Haiku tier has had only three members in two and a half years, which makes its history short but its practical details unusually important - Haiku behaves differently from the rest of the current lineup in ways that break code copied from a Sonnet or Opus integration. Those differences, and the platform model IDs for every cloud Haiku runs on, are below.

#Every Claude Haiku version, in order

Three releases, two of them retired. Figures Anthropic's current documentation no longer publishes for the older versions are marked rather than reconstructed.

#Claude Haiku 4.5 model IDs on every platform

This is the part most people arrive here for. Haiku 4.5 uses a different identifier on each cloud, and unlike the 4.6-and-later generation it still carries a snapshot date - with a different separator on each platform.

Two things to know about that table. First, a dated ID is a pinned snapshot, not an evergreen pointer - Anthropic does not update weights behind an existing ID, and a new version ships under a new one. Second, the retired IDs are claude-3-5-haiku-20241022 and claude-3-haiku-20240307; both return an error on the Claude API, and Anthropic's stated replacement for each is claude-haiku-4-5-20251001. Anthropic's pricing page still lists Haiku 3.5 rates for AWS Bedrock and Google Cloud, so a partner-cloud deployment may survive the first-party retirement - confirm with the cloud vendor, not with this page. The full cross-tier ID list lives on the Claude models page.

#What changed between Haiku generations

Haiku 3 in March 2024 was positioned purely on speed and cost, arriving a few days after the rest of the Claude 3 family. Haiku 3.5, launched in October 2024 alongside the upgraded Claude 3.5 Sonnet and the first public beta of computer use, was priced at $0.80/$4.

Haiku 4.5, a year later, is the version that changed the tier's character. Anthropic reported a SWE-bench Verified score of 73.3%, averaged over 50 trials with a 128k thinking budget - a figure that sat within two points of the then-current Opus 4.1 flagship at 74.5%, for a fifteenth of the input price. SWE-bench Verified has since been dropped by frontier vendors and should be read as historical, but it is the clearest published evidence for Anthropic's "near-frontier intelligence" framing of the tier. Haiku 4.5 also raised the price slightly, from $0.80/$4 to $1/$5 - the only Claude tier where the current model costs more than its predecessor.

One quiet regression worth noting: Haiku 4.5's reliable knowledge cutoff is February 2025 even though its training data runs to July 2025. Anthropic distinguishes the two, and the earlier of the pair is the one to plan around for anything time-sensitive.

#How Haiku differs from the rest of the current lineup

Haiku 4.5 is a 4.5-generation model surrounded by 5-generation siblings, and the gap shows up in the API rather than in the marketing.

Other differences that bite in practice: the minimum cacheable prompt is 1,024 tokens; the newer code-execution tool versions run on Haiku but behave like code_execution_20250825, without REPL state persistence or programmatic tool calling; the output-300k-2026-03-24 batch header does not apply; and fast mode is not offered. Haiku 4.5 also does not appear in Anthropic's list of models supporting the computer use and browser use toolsets that went generally available on 19 August 2026 - those are documented for Fable 5, Opus 5, Sonnet 5 and Opus 4.8.

What it does keep is generous throughput. At the Start rate-limit tier Haiku 4.5 gets 1,000 requests, 2,000,000 input tokens and 400,000 output tokens per minute - identical to Opus 5 and Sonnet 5, and four times Fable 5's input allowance. For fan-out work, the cheap model is also the least constrained one. Rate-limit mechanics are on the Claude API page.

#Who the Haiku tier is for

Haiku earns its place wherever the unit of work is small and the number of units is large: classification, extraction, routing, moderation triage, log summarisation, real-time interfaces where latency is the product, and sub-agent tasks inside a larger orchestration where a bigger model does the planning. Pairing Haiku with the Batch API takes the effective rate to $0.50/$2.50 per million tokens, and prompt caching drops repeated context to $0.10 per million. Patterns for that kind of pipeline are on the automation page, and the sub-agent framing is covered under MCP and Claude agents.

It is the wrong choice when the task needs judgement across a long document, when you need the 1M-token window, or when adaptive thinking and the current tool surface matter. In those cases move up: the Sonnet family is the usual next step at twice the price, and the Opus family is where the long-horizon agentic work belongs. The tier-by-tier comparison puts cost-per-task numbers against each of those choices, and current subscription rates are on the pricing page.

#Frequently asked questions