Claude 3.7 Haiku Gets Extended Thinking at $0.80 per Million Tokens
Anthropic has released Claude 3.7 Haiku, adding extended thinking mode to its fastest, cheapest model tier. It's available via API and Claude.ai at $0.80 per million input tokens.
Original sourceAnthropic has shipped Claude 3.7 Haiku, the latest iteration of its lightweight model line, now featuring extended thinking — the step-by-step reasoning mode previously reserved for heavier models like Claude 3.7 Sonnet. The model is available immediately through the Anthropic API and on Claude.ai, priced at $0.80 per million input tokens, which positions it as one of the more competitively priced options in the extended-thinking tier.
Extended thinking allows the model to work through problems explicitly before returning a final answer, a capability that meaningfully changes what a small, fast model can reliably do. Previously, developers who needed reliable multi-step reasoning had to accept higher latency and cost by reaching for larger models. Haiku with extended thinking is a direct attempt to close that gap — fast inference plus structured reasoning without jumping to a heavier weight class.
The pricing structure suggests Anthropic is positioning Haiku as the workhorse for high-volume, reasoning-adjacent tasks: classification with nuance, structured extraction with edge-case handling, and agentic pipelines that need to think before acting but can't afford Sonnet-level spend per call. Output token pricing has not been prominently featured in initial announcements, which is worth watching, since extended thinking modes tend to generate significantly more tokens than standard completions.
This release continues a broader pattern across frontier labs of pushing reasoning capabilities down the cost curve, making what was a premium feature six months ago a default expectation at the commodity tier. For developers building production systems, the practical question is whether Haiku's reasoning quality is sufficient for their specific task — which means benchmarks need to be run against real workloads, not just announced numbers.
Panel Takes
The Builder
Developer Perspective
“The primitive is clean: extended thinking on a fast, cheap model via the same API surface you're already using for Haiku. The DX bet here is that you don't have to change your integration — just toggle the thinking mode and adjust your token budget. What I actually care about is output token pricing under thinking mode, because extended thinking burns tokens fast and that's where the real cost lives; Anthropic isn't leading with that number, which is a yellow flag before I commit it to a production pipeline.”
The Skeptic
Reality Check
“Extended thinking on a small model sounds like a win until you realize the quality of the reasoning is bounded by the model's underlying capability, not the mode you enable. The scenario where this breaks is multi-hop reasoning tasks requiring deep world knowledge — Haiku will think longer and still get it wrong, just with more tokens billed. What kills this in 12 months is Anthropic shipping Sonnet-level reasoning at Haiku-level prices, at which point 'thinking Haiku' becomes a historical footnote.”
The Founder
Business & Market
“The buyer here is any engineering team running high-volume agentic pipelines who currently flinches every time Sonnet usage spikes their bill — that's a real and large budget category. At $0.80 per million input tokens, Anthropic is pricing to win workloads, not to extract margin, which makes sense if their moat is developer stickiness built through the API rather than per-call revenue. The risk is that extended thinking mode meaningfully inflates output tokens, so the effective cost-per-task may not be as low as the input price implies — teams need to model that before committing.”
The Futurist
Big Picture
“The thesis baked into this release is that reasoning will become a commodity feature at every price tier within 18 months, and the labs that get developers building on their cheap-and-capable models now will own the default integration slot when agents run millions of calls per day. The second-order effect is that extended thinking in small models accelerates the shift from 'use a big model carefully' to 'use a small model confidently at scale' — which is a fundamentally different architecture pattern for production systems. This is on-time to the cost-curve trend, not early, but Anthropic's API ergonomics give it a real shot at being the default runtime for that pattern.”