Anthropic

Claude Haiku 4.5

Provider

Anthropic

Status

available

Context

200,000 tok

SWE-bench

73.3%

Price

$1 / $5 /MTok

Knowledge

2025-02

Claude Haiku 4.5 is Anthropic’s small, fast, cheap tier — the model you reach for when you need to do something simple a million times, or quickly, or both. Released on 15 October 2025 at $1 input / $5 output per million tokens, it sits beneath the Sonnet 4.6 workhorse and the Opus 4.8 flagship, and it’s one of the default models on the free tier of claude.ai.

The pitch that got attention at launch was coding value: Anthropic reported Haiku 4.5 scoring 73.3% on SWE-bench Verified — roughly level with the older Sonnet 4 flagship — while running at “one-third the cost and more than twice the speed.” That’s a genuinely useful trade. But there’s an honest caveat worth stating early, which Simon Willison flagged on day one: Haiku 4.5 is not the cheapest small model on the market. OpenAI’s GPT-5 mini and Google’s Gemini Flash tier undercut it on raw price. Anthropic is competing on coding quality and tool reliability, not on being the rock-bottom option — and whether that trade makes sense depends entirely on your workload.

Quick specs

TRY CLAUDE HAIKU 4.5 →

What’s new in Claude Haiku 4.5

Haiku 4.5 is a big jump over Haiku 3.5 — there was no “Haiku 4”, so the line skipped a generation. The headline changes (Anthropic):

The first Haiku that can think

Haiku 4.5 is the first small Claude to support extended thinking — the model can reason through a problem before answering, with a configurable thinking-token budget. This is what closes much of the gap to the bigger models on harder coding and maths tasks; on AIME 2025 it jumps from 80.7% without tools to 96.3% with Python tools and thinking enabled. It’s a toggle rather than the adaptive low/medium/high/max effort dial that Opus 4.8 and Sonnet 4.6 expose, but the capability is new to the tier.

Computer use and a much bigger output

Haiku 4.5 adds computer use — controlling a desktop, browser and files — scoring 50.7% on OSWorld-Verified, the highest of any Haiku and ahead of the old Sonnet 4 (42.2%). Maximum output rises from Haiku 3.5’s cramped 8,192 tokens to 64,000, which makes it viable for longer code generation and structured extraction rather than just short replies. It also gains context awareness: the model is trained to track how much of its context window it’s using, which Anthropic says reduces “agentic laziness” in long tool-calling loops.

Built to be a worker, not just a chatbot

The most interesting framing in the launch was architectural. Anthropic explicitly demonstrated Sonnet 4.5 breaking a complex problem into a plan, then “orchestrating a team of multiple Haiku 4.5s to complete subtasks in parallel.” This is the barbell pattern — a smart, expensive model plans and coordinates while a fleet of cheap, fast Haiku workers does the legwork. It’s the use case Haiku 4.5 is really designed for, and it’s why the speed and tool reliability matter more than the headline benchmark scores.

A note on price and safety

The trade-off is a 25% price rise over Haiku 3.5 ($0.80/$4 → $1/$5). On the other side of the ledger, Anthropic released Haiku 4.5 under the lighter ASL-2 safety standard and called it their “safest model yet” at launch by automated misalignment rate — reporting a statistically significant lower rate of misaligned behaviours than both Sonnet 4.5 and Opus 4.1 (Anthropic).

The Haiku 4.5 model family

Haiku 4.5 ships as a single model with two API identifiers — a pinned, build-dated snapshot and a dateless alias.

Within Anthropic’s lineup, Haiku 4.5 is the entry tier: below Sonnet 4.6 and the Opus flagships, and far below the Mythos-class Fable 5. It’s the cheapest and fastest current Claude — the trade is capability ceiling, not reliability.

Benchmark performance

Anthropic published Haiku 4.5’s scores in its launch post, run with extended thinking and large (128K) thinking budgets. All figures in the tables below are Anthropic-reported unless noted; the independent verification section follows. The most useful comparison isn’t against Opus — it’s against the older Sonnet 4 flagship, because that’s the “you get yesterday’s top tier, cheaply” claim.

Coding

The claim holds up: Haiku 4.5’s 73.3% on SWE-bench Verified is fractionally ahead of where the Sonnet 4 flagship sat five months earlier, at a fraction of the price (Anthropic). The launch partner Augment reported it hitting “90% of Sonnet 4.5’s performance” on their agentic coding eval. Where the tier shows is agentic terminal work and the harder end of computer use, where the current Sonnet 4.6 is clearly ahead. For the full field, see our best AI for coding rankings.

Reasoning, knowledge and maths

This is where the tier ceiling is real. On graduate-level science reasoning (GPQA Diamond), Haiku 4.5’s 73% sits well below Sonnet 4.6’s 89.9% and the Opus flagships — this is not the model for hard scientific or novel-pattern reasoning. The maths story is more nuanced: 80.7% on AIME 2025 unaided is respectable for a small model, and tools plus thinking push it to 96.3%. Anthropic did not publish Humanity’s Last Exam, MMLU-Pro, or MATH-500 figures for Haiku 4.5, so treat any you see elsewhere as unverified.

Independent verification

Third-party data broadly confirms Haiku 4.5 is strong for its class but not a giant-killer.

Artificial Analysis scored the non-reasoning variant 31 on its Intelligence Index — above the ~24 average for comparable small models — and measured output speed at roughly 93 tokens/sec with sub-second time-to-first-token, the metrics that actually matter for the latency-sensitive jobs Haiku is built for. vals.ai placed it 3rd on the Vals Index (40.9%) at launch, noting strength on coding and agentic tasks but middling results on GPQA, MMLU-Pro and MMMU.

The honest independent read, captured by vals.ai and broader trackers: newer budget rivals like Gemini 3 Flash now score higher on general reasoning composites (~52% vs Haiku’s ~42% on the Vals Index), so Haiku 4.5’s case rests on coding quality, tool-calling reliability and speed rather than topping the leaderboard. One data note: Artificial Analysis lists Haiku 4.5’s input at $1.25/MTok — its measured, blended figure — against Anthropic’s $1.00 list price.

Pricing breakdown

Haiku 4.5 costs $1 input / $5 output per million tokens (Anthropic pricing docs).

For high-volume Haiku workloads, prompt caching and the Batch API are the two biggest cost levers — caching can cut input cost by up to 90% on repeated context, and batch halves both rates. Cache writes are billed separately (1.25x for a 5-minute TTL, 2x for one hour).

Cost comparison with small-model rivals

This table is the whole argument. On price alone, Haiku 4.5 loses to GPT-5 mini and the Gemini Flash family. It earns its premium where coding accuracy, structured-output discipline and reliable tool-calling matter — the launch partner Gamma, for instance, reported Haiku 4.5 hitting “65% accuracy versus 44% from our premium tier model” on slide-text instruction-following. The right move is to benchmark it against the cheaper rivals on your own tasks rather than assuming either way.

How to access Claude Haiku 4.5

Via API

Haiku 4.5 is generally available with no waitlist as claude-haiku-4-5-20251001 on the Claude API, and on Amazon Bedrock (anthropic.claude-haiku-4-5-20251001-v1:0), Google Cloud Vertex AI, Microsoft Foundry and GitHub Copilot. Output throughput varies by host — Artificial Analysis measures the fastest providers at over 110 tokens/sec on Bedrock — and regional Bedrock/Vertex endpoints carry a ~10% premium over the global ones.

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(message.content)

Via the Claude apps

Haiku 4.5 is one of the models behind the free tier of claude.ai and is the default model for a subset of free users; it also powers Claude in Chrome by default. Access by tier:

Inside Claude Code, Haiku 4.5 is commonly used as the fast sub-agent model — the cheap worker in the orchestration pattern described above. There is no free API tier; free access is app-only.

How Claude Haiku 4.5 compares

Haiku’s job is to be the cheap, fast default. The two comparisons that matter are up to Sonnet (when do you need more) and sideways to the budget rivals (when is something cheaper good enough).

vs Claude Sonnet 4.6

Sonnet 4.6 is the step up, and it’s a real one: it beats Haiku 4.5 on every shared benchmark, often by a wide margin (GPQA 89.9% vs 73.0%, SWE-bench Verified 79.6% vs 73.3%, OSWorld 72.5% vs 50.7%). At 3x the price, the question is whether your task needs that headroom.

Choose Haiku 4.5 for high-volume, latency-sensitive, well-scoped work — classification, routing, extraction, real-time chat, and parallel sub-agents. Choose Sonnet 4.6 when accuracy, harder reasoning, longer context (its 1M window) or higher coding quality justify the cost. A common production design uses both: Sonnet plans and handles the hard cases, Haiku does the bulk at scale.

vs GPT-5 mini and Gemini 3 Flash

This is the comparison that decides whether Haiku is the right small model at all. On raw price, GPT-5 mini and the Gemini Flash tier are cheaper, and Gemini 3 Flash leads Haiku on several general-reasoning composites. Haiku’s counter is coding/agentic quality and tool-calling reliability, plus the consistency of the Claude tool-use stack. There’s no universal winner here — the deciding factor is whether the cheaper models clear your accuracy bar on your tasks. See the Google and OpenAI hubs for their current small-model lineups, and our best AI models ranking for the wider field.

The practical consensus

Haiku 4.5 is the right default when speed and cost dominate and the task is well-scoped — and especially as the worker tier in a Sonnet-orchestrated agent system. It’s the wrong choice for hard reasoning (use Sonnet/Opus) or when pure cost-per-token at massive scale is the only thing that matters and quality is forgiving (benchmark the cheaper rivals first).

Known limitations

Not the cheapest small model. GPT-5 mini and Gemini Flash-Lite undercut Haiku 4.5 on price. If your workload is cost-dominated and quality-tolerant, test them head-to-head before defaulting to Haiku.

Trails newer budget rivals on raw reasoning. On broad reasoning composites, Gemini 3 Flash now scores higher (vals.ai). Haiku’s edge is coding and tool reliability, not leaderboard-topping intelligence.

Hard reasoning ceiling. GPQA Diamond at 73% is well below Sonnet 4.6 and Opus. This is not the model for graduate-level science, novel-pattern problems, or anything where being exactly right beats being fast.

200K context only. Unlike Sonnet 4.6 and the Opus line, there’s no 1M-token option — a real constraint for large-document or whole-codebase ingestion.

Vendor-reported headline figures. The benchmark scores are Anthropic-run, with extended thinking and large thinking budgets; real-world results without those settings will be lower. Independent trackers (Artificial Analysis, vals.ai) broadly agree on its standing but use different methodologies.

Community reception

Reception was positive but clear-eyed. The praise centred on speed and the surprising coding quality for the price; the criticism centred on the price itself relative to rival mini-models.

Anthropic’s launch partners reported concrete wins (Anthropic):

• Augment: “achieves 90% of Sonnet 4.5’s performance” on its agentic coding eval, “matching much larger models.”
• Warp: “a leap forward for agentic coding… makes AI-assisted development feel instantaneous.”
• GitHub Copilot: “comparable quality to Sonnet 4 but at faster speed.”
• Gamma: “65% accuracy versus 44% from our premium tier model” on slide-text instruction-following.

Independent coverage was measured. Simon Willison welcomed it as the cheapest Claude 4.5-family model and highlighted the context-awareness training detail, but was openly disappointed it wasn’t price-competitive with the cheapest OpenAI and Gemini models — reading the release as Anthropic “continuing to focus squarely on the ‘great at code’ part of the market.” That framing has held up well: Haiku 4.5’s enduring role has been as the fast, reliable worker tier, not the budget-price leader.

Version history

Note there was no “Haiku 4” — the line jumped straight from 3.5 to 4.5 to align Haiku’s versioning with Sonnet and Opus. As of June 2026, Haiku 4.5 is still the current Haiku; Haiku 3.5 and Haiku 3 were both retired on 19 February 2026, with Anthropic recommending migration to 4.5.

FAQ

Is Claude Haiku 4.5 better than Haiku 3.5?

Substantially. It’s the first Haiku with extended thinking and computer use, raises max output from 8K to 64K tokens, and is a large jump on every benchmark — its Artificial Analysis Intelligence Index of 31 compares to an estimated 12 for Haiku 3.5. The cost is a 25% price rise to $1/$5.

How much does Claude Haiku 4.5 cost?

$1 per million input tokens and $5 per million output tokens. The Batch API halves both rates, and prompt caching cuts repeated-context input cost by up to 90%. On the apps it’s included on every tier, including Free.

Can I use Claude Haiku 4.5 for free?

Yes, in the apps — it’s one of the free-tier models on claude.ai and the default for some free users. There’s no free API tier; API use is paid.

How does Haiku 4.5 compare to Sonnet 4.6?

Sonnet 4.6 is clearly stronger across the board (e.g. GPQA 89.9% vs 73.0%) and has a 1M context window, at 3x the price. Use Haiku for high-volume, latency-sensitive, well-scoped tasks; step up to Sonnet when a task needs more reasoning, accuracy or context.

Is Haiku 4.5 the cheapest small model?

No. GPT-5 mini ($0.25/$2) and Gemini 3.1 Flash-Lite ($0.25/$1.50) are cheaper. Haiku 4.5 competes on coding quality and tool-calling reliability rather than price, so benchmark it against the cheaper rivals on your own tasks before deciding.

What is the context window and knowledge cutoff?

A 200,000-token context window (no 1M option) with up to 64,000 output tokens, and a reliable knowledge cutoff of February 2025 (Anthropic also documents a broader July 2025 training-data cutoff).

Was Haiku 4.5 affected by the June 2026 export-control suspension?

No. The 12 June 2026 US export-control directive suspended only the Mythos-class Fable 5 and Mythos 5; Haiku 4.5 and the rest of the Claude lineup stayed fully available throughout. Those export controls were lifted on 30 June 2026, and Fable 5 is generally available again as of 1 July 2026 (Anthropic).

Last verified 1 July 2026. Benchmark figures are Anthropic-reported via the Claude Haiku 4.5 launch post unless otherwise noted; independent figures from Artificial Analysis and vals.ai are labelled as such. Pricing and availability current as of the publication date and subject to change.