Claude Haiku 5.5 is Anthropic's newest small model, built for high-volume, time-sensitive work such as classification, extraction, routing, and subagent tasks. It arrives as the latest release in Anthropic's foundation model lineup for developers and teams running Claude at production scale, and the headline shift is cost. It costs nearly 75% less than Haiku 4.5 , which matters most to operators whose workloads already run millions of calls a week on the smaller tier.
The model is positioned alongside Sonnet 5.5 and Opus 5.5 as the fast, cheap member of a tiered stack, often used as a subagent that handles narrow lookups while a larger model orchestrates the broader task. It targets high-volume work like summaries, compaction, classification and subagent tasks, and brings a 1M context window, adjustable effort, and $0.10 per million input token pricing. The adjustable effort setting is a first for the Haiku line, letting teams dial between cost and reasoning depth within the same model.
Anthropic also reports notable gains on agentic and computer-use benchmarks over the prior Haiku, with early access partners including Asana, HubSpot, AlphaSense, Box, Rogo, and Cognition citing lower latency and better accuracy on their internal evals. For founders and operators choosing a default workhorse model for retrieval pipelines, browser agents, or customer support flows, this launch resets the price-to-capability floor in the Claude family and reopens the build-versus-buy math for anyone currently routing bulk traffic to a competing small model.
Comments (8)
Sign in to join the discussion.
P
Priya V.3d ago
75% cheaper is nice but what's the retention story once devs realize Haiku 4.5 was already 'good enough' for their eval suite? Curious what the switching behavior looks like in week two.
K
Kenji Morado3d ago
Cool, but is anyone benchmarking latency variance under sustained load or are we all just vibing off the median numbers in the chart?
T
Tobi O.3d ago
The b-roll of tokens flying through the pipeline was shot like a Nike ad. Whoever edited this deserves a raise and a nap.
M
mikhailrune3d ago
Every small model launch gets called the cheapest and fastest. At this rate Haiku 6 ships for negative dollars and arrives before I prompt it.
L
Lailah F.3d ago
Three of my portfolio companies were literally waiting on this exact price point to flip their background workers off GPT-mini. Timing could not be better.
O
Oumar B.3d ago
Mind sharing which portfolio co flipped first? Asking because I'm trying to convince one of mine that migration cost is lower than they think.
D
danya3d ago
Love the drop but one more thing: give us structured output caching at the Haiku tier and you'll own the agent runner market by Q1.
R
Raven S.3d ago
This is probably the last small model release before every launch tweet is written by the model it's announcing. Enjoy the human-crafted copy while it lasts.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Anthropic
The cheapest, fastest, most capable small Claude
The cheapest, fastest, most capable small Claude
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Claude Haiku 5.5 is Anthropic's newest small model, built for high-volume, time-sensitive work such as classification, extraction, routing, and subagent tasks. It arrives as the latest release in Anthropic's foundation model lineup for developers and teams running Claude at production scale, and the headline shift is cost. It costs nearly 75% less than Haiku 4.5 , which matters most to operators whose workloads already run millions of calls a week on the smaller tier. The model is positioned alongside Sonnet 5.5 and Opus 5.5 as the fast, cheap member of a tiered stack, often used as a subagent that handles narrow lookups while a larger model orchestrates the broader task. It targets high-volume work like summaries, compaction, classification and subagent tasks, and brings a 1M context window, adjustable effort, and $0.10 per million input token pricing. The adjustable effort setting is a first for the Haiku line, letting teams dial between cost and reasoning depth within the same model. Anthropic also reports notable gains on agentic and computer-use benchmarks over the prior Haiku, with early access partners including Asana, HubSpot, AlphaSense, Box, Rogo, and Cognition citing lower latency and better accuracy on their internal evals. For founders and operators choosing a default workhorse model for retrieval pipelines, browser agents, or customer support flows, this launch resets the price-to-capability floor in the Claude family and reopens the build-versus-buy math for anyone currently routing bulk traffic to a competing small model.
Comments (8)
75% cheaper is nice but what's the retention story once devs realize Haiku 4.5 was already 'good enough' for their eval suite? Curious what the switching behavior looks like in week two.
Cool, but is anyone benchmarking latency variance under sustained load or are we all just vibing off the median numbers in the chart?
The b-roll of tokens flying through the pipeline was shot like a Nike ad. Whoever edited this deserves a raise and a nap.
Every small model launch gets called the cheapest and fastest. At this rate Haiku 6 ships for negative dollars and arrives before I prompt it.
Three of my portfolio companies were literally waiting on this exact price point to flip their background workers off GPT-mini. Timing could not be better.
Mind sharing which portfolio co flipped first? Asking because I'm trying to convince one of mine that migration cost is lower than they think.
Love the drop but one more thing: give us structured output caching at the Haiku tier and you'll own the agent runner market by Q1.
This is probably the last small model release before every launch tweet is written by the model it's announcing. Enjoy the human-crafted copy while it lasts.