The Vibe Father
Model Review
Claude Haiku 4.5 is Anthropic's small, fast, cheap tier — released back in October 2025, which makes it the oldest model we still track on our live leaderboard. That age shows in the scores. But writing Haiku off entirely would be a mistake, because there is one job in a modern agent team it does brilliantly, and it does that job for pocket change. This review is about being precise on where that line sits.
The essentials: $1 per million input tokens, $5 per million output, 96 tokens per second — the fastest model in the Claude family — and a 200k context window, notably smaller than the 1M windows on Sonnet 5 and up. It is built for speed and volume, not depth, and the benchmarks say exactly that.
The numbers
Here is Haiku against its own family. Our Vibe Coding Index weights SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30%.
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) |
|---|---|---|---|---|
 Claude Haiku 4.5 | 66.6 | 35.5 | 41.2 | $1 / $5 |
| Claude Sonnet 5 | 85.2 | not yet published | 82.4 | $3 / $15 |
 GPT-5.3 Codex | 74.8 | 78.4 | 87.3 | $1.75 / $14 |
Let us be blunt about two of those cells. A 35.5 on Terminal-Bench means Haiku fails roughly two out of three realistic terminal-driven agent tasks — that is disqualifying for autonomous building, full stop. And 41.2 on LiveCodeBench puts it far behind everything else on our board on fresh problem-solving. The 66.6 on SWE-bench Verified is the one respectable number: enough to navigate a repository and answer questions about it, not enough to be trusted with unsupervised changes to it. Note also GPT-5.3 Codex in that table — for 75 cents more per million input, it more than doubles Haiku's Terminal-Bench score. The cheap tier has gotten competitive.
In practice, the profile is even more lopsided than the table suggests. Give Haiku a read-and-report job — "map this module and tell me where authentication happens" — and the answer comes back in seconds, accurate and usefully compressed. Give it a build job and you watch the Terminal-Bench number come to life: it runs the wrong command, misreads the error, "fixes" the wrong file, and does all of it with total confidence at 96 tokens per second. Speed in the wrong direction is not a feature. The 200k context is also a real constraint on larger codebases; the bigger Claudes hold five times as much, which matters even for scouting work on sprawling monorepos.
The price math
Our reference heavy month: 50M input tokens, 10M output tokens.
Haiku 4.5: 50 × $1 + 10 × $5 = $50 + $50 = $100 per month. The same firehose costs $300 on Sonnet 5, $500 on Opus 4.8, $1,000 on Fable 5, and $227.50 on GPT-5.3 Codex.
One hundred dollars for 60 million tokens of work is genuinely remarkable — if the work is work Haiku can do. The trap is obvious once you have fallen into it: a "cheap" model that fails a task three times before you escalate to Sonnet did not cost $100, it cost $100 plus the Sonnet run plus your afternoon. Cheap tokens are only cheap when the pass rate is high, a theme we hammer throughout our cheapest coding models guide.
Its best seat on the team
Haiku 4.5 is a scout, and only a scout. In a multi-agent team, that means: reading files and summarizing them, grepping and mapping unfamiliar codebases, triaging which of fifty test failures share a root cause, classifying and routing incoming tasks, and running the cheap first pass over logs before a bigger model spends real tokens on the interesting parts. At 96 tok/s and $1 in, it is close to free reconnaissance, and its 66.6 SWE-bench score is plenty for comprehension-shaped work.
What it must never be is a solo builder. The 35.5 Terminal-Bench score is not a nuance, it is a verdict — autonomous multi-step execution is where Haiku collapses. Route implementation to Sonnet 5 or GPT-5.3 Codex, keep planning with Fable 5 or Opus, and let Haiku do the legwork in front of them. The full division of labor is in best model for each agent role.
Verdict
Haiku 4.5 earns its place on a well-run team — as the cheapest competent reader in the business. Used for reconnaissance, triage, and summarization in front of stronger builders, it pays for itself many times over. See where it sits overall in our best coding model rankings.
Who should pay for it: multi-agent operators who need a high-volume scout tier, and teams processing mountains of read-heavy work — codebase Q&A, log triage, classification.
Who should not: anyone looking for one cheap model to do everything. If you want a single budget engine that can actually build, GPT-5.3 Codex at $227.50 a heavy month is the honest answer, and our cross-lab shootout shows why. Solo developers hoping Haiku will write their app will spend more time cleaning up than they saved. Know what you hired it for, and it is a bargain. Forget, and it is not.