Anthropic cuts small-model prices and adds credits to plans

Pelicans and a bunch of notes on the pricing of the new Claude Haiku 5.5

Haiku 5.5 Available in Notion Ready when you are

Claude Haiku 5.5 is now available in Devin. On FrontierCode 1.1, Haiku 5.5 scores 58.4%. That puts it ahead of Sonnet 5 at about an eighth of the cost per task. It is also an excellent sidekick model in Fusion. Try it today with Opus 5.5 as the Lead in Devin Desktop and CLI.

Wait, do these new API credits replace being able to use your subscription in Claude Agent SDK apps? Second image is from wayback machine before today.

Anthropic's $200 Claude plan now mints $200 in API credits every month. The subscription pays for itself on paper, and AWS proved 15 years ago why a company would do this on purpose. Credits are denominated at list price. A dollar of credit costs Anthropic a fraction of a dollar to serve, and plenty of subscribers will never spend theirs at all. The ones who do are the exact people worth subsidizing, because developer habits compound. AWS Activate ran the same play starting in 2013. Hand startups free credits, let them build their entire stack on your infrastructure, and collect for a decade. AWS is a $100B+ a year business now, and a meaningful slice of it started as free compute nobody intended to pay for. The clause that actually matters is "third-party harnesses." These credits work…

Cached reads is what agents do the most, now cheaper!

Anthropic just cut the price of its cheapest model 10x and the new one beats OpenAI's small model at operating a computer. 72.4% vs 48.9% on OSWorld. Run the actual numbers on the 10x. Haiku 4.5 charged $1 per million input tokens and $5 on output. Haiku 5.5 charges $0.10 and $0.50 for prompts under 100K tokens, which covers 90% of the traffic the old model served. Cache reads fell from a dime to a penny. So an always-on agent chewing through a million tokens an hour costs about $2.40 a day in input. A tireless digital worker, priced under your morning coffee. This is the part of the AI race that actually hits invoices. Frontier models win benchmarks and headlines. Small models win deployments, because most production AI is grunt work. Summaries, classification, routing. The question…

Anthropic's $200 Claude plan now mints $200 in API credits every month. The subscription pays for itself on paper, and AWS proved 15 years ago why a company would do this on purpose. Credits are denominated at list price. A dollar of credit costs Anthropic a fraction of a dollar to serve, and plenty of subscribers will never spend theirs at all. The ones who do are the exact people worth subsidizing, because developer habits compound. AWS Activate ran the same play starting in 2013. Hand startups free credits, let them build their entire stack on your infrastructure, and collect for a decade. AWS is a $100B+ a year business now, and a meaningful slice of it started as free compute nobody intended to pay for. The clause that actually matters is "third-party harnesses." These credits work…

Shoutout to @cjav_dev for updating the docs to add clarity here. Anthropic are cooking! TLDR; your subscription still works with the SDK & `claude -p `- go enjoy the 5.5 family

I ran 20 of my podcast transcripts through Haiku 5.5, same prompt as 4.5. My bill fell 80%. Anthropic said 75%. Thinking is on by default now, so output grew 3.9x and my old 1,000-token cap cut off 17 of 20 answers mid-sentence. One task type, no quality benchmark.