Claude Haiku 5.5: Anthropic Cuts AI Pricing by 75%

Anthropic just made its cheapest model even cheaper, and for once "up to 75% off" isn't just a marketing line buried under footnotes. Claude Haiku 5.5, released on October 7, is now Anthropic's smallest and most affordable model, and the price tag alone is going to change how a lot of small teams build with AI.

What actually changed

Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens (the price rises to $0.50 and $2.50 per million beyond that length). Compare that to its predecessor, Haiku 4.5, which cost $1 per million input tokens and $5 per million output tokens, and you can see why Anthropic is calling this roughly a 75% average cost cut. A token, for anyone new to this, is just a chunk of text the model reads or writes — a few characters to a short word — and API pricing is almost always quoted per million of them.

The model is live on Anthropic's own API as well as Amazon Web Services, Google Cloud, and Microsoft Azure, which matters because most companies don't call a single AI vendor directly anymore — they go through whichever cloud they already run their infrastructure on. Haiku 5.5 also ships with a 1-million-token context window (roughly how much text and conversation history the model can hold in mind at once), a 128K-token maximum output, and is the first Haiku-class model with an adjustable "effort" setting, letting developers dial reasoning up or down depending on whether a task needs a quick answer or a more careful one.

• Short prompts (under 100K tokens): $0.10 input / $0.50 output per million tokens
• Long prompts (over 100K tokens): $0.50 input / $2.50 output per million tokens
• Context window: 1 million tokens; max output: 128K tokens
• Available on Anthropic's API, AWS, Google Cloud, and Microsoft Azure

"The cheapest, fastest, and most capable small model we've ever released." — Anthropic, Claude Haiku 5.5 launch announcement

The fine print nobody's headline mentioned

Before anyone gets too excited about the sticker price, there's a catch worth flagging. Developer and AI commentator Simon Willison pointed out that Haiku 5.5 uses a new tokenizer — the system that breaks text into those billable chunks — and it needs roughly 1.25 times as many tokens to represent the same prompt compared to Haiku 4.5. So the effective cost saving, once you actually run real workloads through it, ends up smaller than the headline 75% figure suggests. It's still a meaningful cut, just not quite as dramatic as the price list alone implies.

This is also arriving at a telling moment for Anthropic. Several reports tied the release to the company positioning itself ahead of a planned IPO, a reminder that aggressive price cuts from frontier AI labs aren't purely generosity — they're also about locking in developers and enterprise customers before going public.

Why this matters more in India than in San Francisco

For a Bangalore or Gurugram startup running customer support chatbots, content moderation, or data classification at scale, the price of a small model isn't a rounding error — it's often the difference between workable unit economics and a product that bleeds cash per user. Indian SaaS and fintech companies typically operate on thinner per-user revenue than their US counterparts, so a model built specifically for "high-volume, cost-sensitive work" lands differently here. A support bot handling thousands of queries a day in regional languages, or a fintech app screening transactions for fraud flags, runs exactly the kind of repetitive, bulk task a cheaper small model is meant for.

It also plugs into a broader shift: as Anthropic's earlier Haiku 4.5 release already showed, the real competition among AI labs right now isn't only about who has the smartest flagship model — it's about who can make a good-enough model cheap enough that running it a million times a day stops being a budgeting headache. That race benefits Indian developers directly, since most local AI products are built on API calls to these foundation models rather than training anything from scratch.

What to actually watch

• Whether rival labs — OpenAI, Google, Mistral — respond with their own small-model price cuts in the coming weeks
• Real-world cost benchmarks from developers once the tokenizer difference is accounted for, not just Anthropic's own numbers
• Whether Indian AI-tooling startups publicly switch workloads to Haiku 5.5 for cost reasons

Price wars between AI labs rarely stay confined to press releases for long. If Haiku 5.5 forces competitors to match it, the actual winners won't be the labs at all — it'll be every small team building on top of them, India included, watching their monthly API bill quietly shrink.