Claude Haiku 4.5: fast, cheap, near-frontier coding

Claude Haiku 4.5 is Anthropic's fast, low-cost model that closes most of the gap to Sonnet on coding. It scores 73.3% on SWE-bench Verified — within a few points of Sonnet 4.6 and ahead of every other model in its price tier. It's the first Haiku generation to ship with extended thinking, computer use, and context-awareness features that were previously Opus/Sonnet exclusives. Priced at $1 per million input tokens and $5 per million output, it's 1/15th the cost of Opus 4.7 and 1/3 the cost of Sonnet 4.6. The case for: high-throughput coding agents, fast IDE completions, bulk refactor tasks, vision tasks that don't need Opus-grade nuance. The case against: hairy architecture decisions and long-context refactors where Sonnet's deeper reasoning still wins. For councils, Haiku 4.5 is the workhorse — pair it with one Opus or GPT-5.5 for a balanced quality/cost mix.

Specs

Best at

• High-throughput coding — 73.3% SWE-bench at 1/3 Sonnet's price.
• Fast IDE-style tasks — autocomplete, lint fixes, small refactors, tests.
• Agent fan-out — cheap enough to run many parallel sub-agents.
• Vision-light tasks — chart reading, screenshot QA, document extraction.
• Council workhorse — pair with Opus or GPT-5.5 to balance quality and cost.

Where it loses

• Architecture decisions — Opus 4.7 and Sonnet 4.6 still win on hairy design questions.
• Long-context refactors — Sonnet's deeper reasoning holds coherence better past 100K tokens.
• Writing voice — Opus has more editorial taste; Haiku is competent but plain.

When to use Haiku 4.5 vs Sonnet 4.6

Default to Haiku 4.5 for any task you'd run more than a few times a day. The price gap (3x cheaper input, 3x cheaper output) compounds fast on agent loops and bulk tasks. Promote to Sonnet 4.6 when the task has a hard reasoning step, more than ~100K tokens of context, or customer-facing writing where voice matters. Use Opus 4.7 only when both of those apply and the stakes justify the 15x premium.

Frequently asked questions