Quick answer
- What it is
- The fast/cheap tier: subagents, triage, mechanical edits
- Price
- $1 in / $5 out per MTok (cheapest Claude tier)
- Context
- 200k tokens (smallest in the family)
- Not for
- Autonomous feature work or complex multi-step coding
Where Haiku 4.5 Fits, Honestly
Claude Haiku 4.5 was released in October 2025, which makes it the oldest model in the current Claude lineup, everything above it, Sonnet 5, Opus 4.8, and Fable 5, shipped in mid-2026. It carries a 200k tokens window, the smallest in the family, and the API identifier is claude-haiku-4-5-20251001.
None of that is a reason to ignore it. At $1 in / $5 out per MTok it is the cheapest Claude by far, and there is a class of work, high-volume, bounded, mechanical, where paying more buys you nothing. The point of this page is to draw that boundary precisely: what Haiku is genuinely good at, and what it should never be asked to lead. For the full lineup context, see the best Claude model for coding.
Specs and Pricing
| Spec | Claude Haiku 4.5 |
|---|---|
| Vendor | Anthropic |
| API identifier | claude-haiku-4-5-20251001 |
| Released | October 2025 |
| Price | $1 in / $5 out per MTok |
| Context window | 200k tokens |
| Tier | Fast (cheapest and oldest in the current lineup) |
| Claude Code | /model haiku |
Data verified July 18, 2026 against Anthropic's published pricing and docs. Anthropic has not published current agentic-coding benchmark scores for Haiku 4.5 that we can verify, and given its role that is the honest signal: this tier is not competing on SWE-bench, it is competing on price and latency.
The Cost Math vs Sonnet, Opus, and Fable
Every multiple below is computed from the verified list prices, not vibes:
| Model | Price | Input vs Haiku | Output vs Haiku |
|---|---|---|---|
| Claude Haiku 4.5 | $1 in / $5 out per MTok | 1x | 1x |
| Claude Sonnet 5 | $3 in / $15 out per MTok | 3x | 3x |
| Claude Opus 4.8 | $5 in / $25 out per MTok | 5x | 5x |
| Claude Fable 5 | $10 in / $50 out per MTok | 10x | 10x |
Data verified July 18, 2026. Two readings of this table. For a single task, the gap rarely matters, a story's tokens cost cents either way. For anything that runs wide or often, a fan-out of twenty subagents, a pipeline that triages every incoming issue, a summarizer that runs on every PR, the 5x-vs-Opus multiple is the difference between a rounding error and a line item. Note that Sonnet 5's intro pricing ($2 in / $10 out per MTok through August 31, 2026) temporarily narrows its gap to 2x on input, worth knowing while it lasts.
Ship your first agent today
Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.
Download AIDEN freeFree to start · macOS 12+ · No credit card required
What Haiku Is Genuinely Good At
Subagent fan-outs
Triage & classification
Summarization
Mechanical edits
Cost-capped pipelines
Latency-sensitive steps
What It's Not For
Haiku 4.5 is not built for complex autonomous coding, that is a design point, not a flaw. Multi-file feature work, debugging where the cause hides far from the symptom, and long unattended runs all demand reasoning depth this tier does not have, and the 200k tokens window rules out whole-repo context sessions that the 1M-window tiers handle natively.
The Pattern That Works: Opus Leads, Haiku Fans Out
The strongest Haiku deployment in 2026 is not Haiku alone, it is heterogeneous fleets. A lead session on Opus 4.8 owns the plan and the review, and fans bounded subtasks out to Haiku subagents: codebase searches, per-file summaries, mechanical change execution. Each subagent call costs 1/5th of the same call on Opus, so the fan-out can go wide while the judgment stays with the flagship.
The same logic applies one level up. On an AIDEN board, model routing is per story card: the mechanical cards, rename sweeps, doc syncs, lint burn-downs, run on Haiku, while Sonnet and Opus take the cards that need a real engineer's model. Cheap tier, cheap work; it is the whole trick.