- Home
- Learn
- Deep Dives
- Claude Haiku 4.5 and 5.5: Speed, Pricing and When to Use It
Claude Haiku 4.5 and 5.5: Speed, Pricing and When to Use It
Claude Haiku 4.5 and the newer Haiku 5.5 compared: price per token, speed, benchmarks vs Sonnet, the jobs it handles and where it fails. Checked 16-11-2026.
Published 16-11-2026. Last verified on 16-11-2026 against Anthropic's Haiku launch posts, pricing page and deprecations page, and the Artificial Analysis model page for Haiku 5.5. Model pages go stale fast, so we re-check this guide monthly.
Quick answer: is Claude Haiku 4.5 still the one to use?
Claude Haiku 4.5 was Anthropic's fast, cheap model from 15-10-2025, at $1 per million input tokens and $5 per million output tokens with a 200K-token context window. Since 07-10-2026 its successor, Claude Haiku 5.5, costs from $0.10 and $0.50 for prompts up to 100,000 tokens, has a 1M-token window and beats Haiku 4.5 on every benchmark Anthropic reports for both. This guide is for teams that run volume work on Claude (support replies, classification, summaries, hooks and subagents) and want to know which jobs can move to Haiku and where it breaks. Our answer: start new work on Haiku 5.5, plan to move Haiku 4.5 workloads soon, and keep Haiku away from jobs where it has to lead multi-step coding.
How we judged Haiku
Our method in three lines: specs and prices come from Anthropic's own pages, read on 16-11-2026; capability numbers come from Anthropic's launch posts (vendor-reported) and from Artificial Analysis (independent, but not our own measurement); and every recommendation names the number it rests on.
Our gate: we recommend Haiku for a job only if the published numbers put it close to Sonnet on a benchmark that matches the job, or if the job has a checkable right answer, such as a label or an extracted field, so you can catch its mistakes cheaply. The briefing for this page asked us to measure Haiku against Sonnet on our own classification and summarisation tasks. We had no API access in this production run, so that comparison is not tested by us; the last section gives the exact test so you can run it.
The gate disqualified one use that other Haiku pages promote: Haiku as the lead model for agentic coding. Anthropic's own table puts Haiku 5.5 at 39.2% on Terminal-Bench 4.0 against 70.6% for Sonnet 5.5. A gap of more than 31 points is too big for a model that has to plan and fix its own mistakes. Simple, well-defined coding subtasks under a larger model passed; leading a coding agent did not.
What is Claude Haiku 4.5?
Claude Haiku 4.5 is the small Claude model Anthropic launched on 15-10-2025, with API ID claude-haiku-4-5-20251001. In its launch post, Anthropic said it offered coding performance similar to Claude Sonnet 4 "at one-third the cost and more than twice the speed", and reported 73.3% on SWE-bench Verified (vendor-reported). It has a 200K-token context window and 64K tokens of maximum output, according to the models overview.
Its price is $1 per million input tokens and $5 per million output tokens, with cache reads at $0.10 per million, on the pricing page on 16-11-2026. That was cheap in 2025; next to Haiku 5.5 it is now ten times more expensive per short prompt.
Is it still available? Yes. Anthropic's deprecations page lists claude-haiku-4-5-20251001 as Active, with a retirement date "not sooner than October 15, 2026". That date has passed without a deprecation notice. We read that as a model on borrowed time: Anthropic promises at least 60 days' notice, so nothing breaks tomorrow, but we would not start anything new on it.
What changed in Claude Haiku 5.5?
Claude Haiku 5.5 (claude-haiku-5-5) launched on 07-10-2026. Anthropic describes it as its cheapest and fastest small model so far, and its most capable one, and says it costs around 75% less to run than Haiku 4.5 on average (vendor-reported). We see four changes that matter for a team that runs volume.
The context window grew from 200K to 1M tokens and maximum output from 64K to 128K. Haiku 5.5 is the first Haiku with an adjustable effort setting, so you can trade cost against intelligence per job; its default is medium. The price became tiered: cheap up to 100,000 input tokens, five times higher above that. And the capability jump is large. From Anthropic's Haiku 5.5 launch post, all vendor-reported:
| Benchmark | Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 |
|---|---|---|---|
| GDPval-AA v2.1 (knowledge work, Elo) | 1620 | 735 | 1840 |
| OSWorld 2.1, offline subset (computer use) | 72.4% | 15.7% | 83.9% |
| Humanity's Last Exam, no tools | 45.9% | 10.2% | 56.9% |
| Terminal-Bench 4.0 (agentic coding) | 39.2% | 0.0% | 70.6% |
| Chartography, no tools (charts) | 46.4% | 6.4% | 61.6% |
Our reading: Haiku 5.5 closes most of the gap to Sonnet on knowledge work and computer use, and very little of it on agentic coding. That pattern is the basis for every recommendation below.
Claude Haiku pricing
We took these prices per million tokens from Anthropic's pricing page, 16-11-2026:
| Input | Output | Cache read | Batch input and output | |
|---|---|---|---|---|
| Haiku 5.5, prompts up to 100K | $0.10 | $0.50 | $0.01 | $0.05 and $0.25 |
| Haiku 5.5, prompts over 100K | $0.50 | $2.50 | $0.05 | $0.25 and $1.25 |
| Haiku 4.5 | $1 | $5 | $0.10 | Half of standard |
What that means per 1,000 jobs, in our arithmetic: a support ticket of 800 input tokens and a 20-token label costs $0.09 per 1,000 tickets on Haiku 5.5 and $0.90 on Haiku 4.5. Through the Batch API, Haiku 5.5 drops to $0.045. Two warnings before you budget. Prompts over 100,000 tokens move into the higher tier, so a job that reads whole documents can cost five times more per token than you planned. And the release notes state that Haiku 5.5 uses a tokenizer that produces more tokens for the same text, so recount your prompts instead of reusing Haiku 4.5 counts.
For the full price list of every model, see our Claude API pricing guide. Skills that watch these numbers in production are in the LLM ops collection.
How fast is Claude Haiku?
Haiku 5.5 is Anthropic's fastest model: the models overview labels its latency "Fastest", against "Fast" for Sonnet 5.5 and "Moderate" for Opus 5.5. How fast in practice depends on the effort setting, and that is where we found the most useful independent data.
Artificial Analysis measured Haiku 5.5 at five effort levels, read on 16-11-2026:
| Effort | Intelligence Index | Minutes per index task | Cost per index task |
|---|---|---|---|
| Low | 19 | 2.2 | $0.0438 |
| Medium (default) | 25 | 4.0 | $0.0918 |
| High | 30 | 6.5 | $0.16 |
| Xhigh | 35 | 9.9 | $0.25 |
| Max | 36 | 14.7 | $0.41 |
For comparison, Artificial Analysis lists Opus 5.5 at max effort at 59 on the same index, 24.9 minutes and $10.04 per task. We have ordered the rows by effort because the trade-off is the point: from low to max, Haiku 5.5 nearly doubles its index score and takes almost seven times as long. Xhigh gets 35 of the 36 points at two-thirds of the max-effort time, which makes it our pick when Haiku needs to think; for labels and routing we would test low first. Artificial Analysis also reports an output speed of 257 tokens per second at max effort and calls the model "very verbose", which matters because thinking tokens are billed as output. The Haiku cost figures carry an asterisk on their page, so check their footnote before you quote them.
You can pull the same data from inside your agent. Through Agensi Tools, Artificial Analysis returns model details from $0.013 per call and model cost per task from $0.13 per call (starting prices read on 16-11-2026), with no Artificial Analysis account or API key, and a failed call costs nothing.
Claude Haiku vs Sonnet: which should you use?
Use Haiku 5.5 when the job is short, repeats at volume and has a checkable answer; use Sonnet 5.5 when the job needs several reasoning steps or writes something a customer reads without review. On price, Haiku 5.5 is twenty times cheaper than Sonnet 5.5 for prompts up to 100,000 tokens ($0.10 and $0.50 against $2 and $10).
On capability, the gap depends on the job. In Anthropic's table, Haiku 5.5 reaches 88% of Sonnet 5.5's GDPval-AA score (1620 against 1840) and 86% of its OSWorld score (72.4% against 83.9%), but only 56% of its Terminal-Bench 4.0 score (39.2% against 70.6%). For knowledge work and browser automation, Haiku is close enough that the price decides. For agentic coding, it is not.
Haiku 4.5 against Sonnet 4.5 is a historic comparison now: Sonnet 4.5 retires on 30-11-2026 according to the deprecations page, so we would not plan around either.
Which jobs and skills should run on Haiku?
Our ranking, from the job where Haiku is the safest bet to the one where we would test hardest first:
- Classification and routing. First because the answer is a label you can check, and a wrong label is cheap to catch. Customer Case Classification and Routing, $2.99, triages support cases by topic, urgency P1 to P4 and owning team with deterministic rules, which pairs well with a cheap model. The business operations collection has more triage skills.
- Templated replies. Second because the format is fixed and a person can approve before sending. Email Reply Generator, $1.99, drafts replies from your intent, tone and language.
- Summaries and review analysis. Third because quality varies with the source, so spot checks matter. Customer Review Insight Analyzer, $1.99, turns reviews into a report with sentiment and recurring themes. To feed it, your agent can collect the reviews through the Trustpilot tool on Agensi Tools, from $0.0011 per call (read 16-11-2026), without a Trustpilot account or API key. More analysis skills are in the research and analysis collection.
- Browser and desktop automation. Fourth: Haiku 5.5 scores 72.4% on OSWorld 2.1 offline (vendor-reported), strong for form filling and data entry, but a wrong click can have real effects, so we would keep a person on the first runs.
- Simple coding subtasks. Last: only as a subagent under a larger model, never as the lead (see the gate above).
What would move a job up this list: a quality check you can automate. Any job with an automated check can safely start on Haiku; any job without one should start on Sonnet. To make that split automatic, Route Each Task to the Right Model, $12, sends easy work to a cheaper model and escalates when a check fails. Pipelines that run these jobs at volume are in the workflow automation collection.
Should hooks and subagents run on Haiku?
Mostly yes, and this is where we see the clearest case for Haiku. Hooks and subagents run often, have narrow jobs and report back to a larger model that can catch mistakes, which is exactly the profile Anthropic describes for Haiku 5.5 when it says the model "pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work". A hook that summarizes a test log or a subagent that searches a codebase is a Haiku job in our view; a subagent that writes the fix is not.
Two skills help set this up. Claude Code Starter Pack, $12, gives you rules, handoff templates and three cheap pinned-model subagents. Subagent Dispatch, $6.99, decides when to delegate at all, because a subagent call costs tokens even on Haiku. Our guides to Claude Code hooks and Claude Code subagents cover how to pin a model per hook or subagent, and the agents and orchestration collection has more orchestration skills.
Where does Haiku fail?
Haiku fails where a task needs many dependent steps. The clearest number is Terminal-Bench 4.0: 39.2% for Haiku 5.5 against 70.6% for Sonnet 5.5, vendor-reported. On Humanity's Last Exam without tools, Haiku 5.5 scores 45.9% against 56.9% for Sonnet 5.5, so hard reasoning questions are also a weaker area.
Three failure cases we would plan for. Long multi-step agent work: errors compound, so a small per-step gap becomes a failed run. Long documents: the 100K price tier and the larger context can tempt you to stuff whole archives into one Haiku call, which costs more and, in our view, invites missed details. And high effort without a reason: Artificial Analysis lists Haiku 5.5 at max effort as "notably slow" per task, 14.7 minutes against 2.2 at low effort, so turning effort up to fix quality can cost you the speed you chose Haiku for.
Moving from Haiku 4.5 to Haiku 5.5
We would plan it as a code change, not a model swap. According to Anthropic's release notes for 07-10-2026, code written for Haiku 4.5 can break on Haiku 5.5: manual extended thinking with budget_tokens, the temperature, top_p and top_k parameters and assistant message prefill can each return a 400 error. Adaptive thinking is on by default, so responses can start with thinking blocks. Anthropic recommends claude-haiku-4-5-20251001 as the replacement for retired Haiku 3 and 3.5 models on its deprecations page, so many older pipelines landed on Haiku 4.5 and now face a second move.
Model Migration Checker, $15, audits code for hardcoded model versions, deprecated parameters and token limits before you switch. Our articles on reducing Claude Code token usage and skills that cut Claude Code costs help once you are on the new model.
Run our test on your own jobs
Vendor and index numbers tell you where to look; your own data tells you whether to switch. We would not move a production job without it. This is the test the briefing asked us to run, written so you can run it in an afternoon:
- Take 100 real items per task. First because everything depends on it: 100 support tickets with a known correct category, and 20 documents with a summary you trust.
- Use one fixed prompt. For classification: "Classify this support ticket into exactly one of these categories: billing, bug, feature request, account access, other. Reply with the category only." Keep it identical for both models.
- Run Haiku 5.5 at low and medium effort and Sonnet 5.5 at its default. Third: record accuracy against your labels, median seconds per item and cost from the usage data.
- Apply a rule you set beforehand. Last: for example, switch to Haiku if its accuracy is within two points of Sonnet. Without a rule written down first, people tend to keep the more expensive model.
Prompt Eval Harness Pro, $9, gives you the case templates, graders and cost and latency reporting for exactly this kind of run. More testing skills are in the testing and debugging collection.
Your next step
We sell skills and pay-per-call tools that your agent uses through one MCP connection, and we name ours in the sections above. No vendor paid to be listed.
Pick the one job your team runs most often, check whether it has a right answer you can verify, and if it does, run the test above on Haiku 5.5. To pull model data and the review data for your first Haiku job from your own agent, connect it once through Agensi MCP.
Written by Samuel Rose, Founder of Agensi. Next scheduled review: 16-12-2026.