TL;DR
- Anthropic launched Claude Haiku 5.5 on October 7 at $0.10/$0.50 per million tokens (input/output) for prompts under 100,000 tokens — a 90% price cut from Haiku 4.5's $1/$5. Cache reads: $0.01 per million. This exactly matches OpenAI's GPT-6 Luna pricing tier. - Simultaneously: Anthropic halved Sonnet 5.5's cache-read price from $0.20 to $0.10, reducing the cost of most agentic work by approximately 20%. - Above 100,000 tokens, Haiku 5.5 rates step up to $0.50/$2.50 — still 50% below Haiku 4.5. Approximately 90% of Haiku 4.5 requests fell under the 100K threshold, so most workloads get the lower rate. - Haiku 5.5 ships with computer-use and browser-use support in beta across Anthropic's Python and TypeScript SDKs. It also adds an adjustable effort setting — the first Haiku to have one. - New monthly API credits: $100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team plans. - Available day one on AWS, Google Cloud, and Microsoft Azure. - The same day: Microsoft announced Surface Laptop Ultra ($2,599, ships October 16) and Surface RTX Spark Dev Box ($5,999, November) — both built around Nvidia's RTX Spark chip with up to 128GB unified memory, capable of running 120B+ parameter models on-device. The cloud-first era for AI inference is being challenged from two directions at once.
Two things happened on October 7 that pull the AI pricing conversation in opposite directions and arrive at the same conclusion: running AI inference is getting dramatically cheaper regardless of where you run it.
Anthropic cut the price of its smallest model by 90%. Microsoft and Nvidia announced a laptop that runs 120-billion-parameter models locally. One reduces your cloud bill. The other reduces your dependence on the cloud at all.
The Haiku 5.5 pricing math
The short version: for prompts under 100,000 tokens, Haiku 5.5 is now the cheapest frontier-family model in Anthropic's lineup by an enormous margin.
Full pricing table — Haiku 5.5 vs the rest of the current frontier tier:
Model | Input ($/M) | Cache reads ($/M) | Cache writes ($/M) | Output ($/M) Haiku 5.5 (up to 100K tokens) | $0.10 | $0.01 | $0.125 | $0.50 Haiku 5.5 (over 100K tokens) | $0.50 | $0.05 | $0.625 | $2.50 Haiku 4.5 | $1.00 | $0.10 | $1.25 | $5.00 GPT-6 Luna (OpenAI) | $0.10 | $0.01 | — | $0.50 Sonnet 5.5 (updated) | $2.00 | $0.10 | $2.50 | $10.00 GPT-6.1 Sol (OpenAI) | $2.00 | $0.10 | — | $10.00 Opus 5.5 / Gemini 4 Argon (intro) | $4.00 | ~$0.20 | — | $20.00 GPT-6 Astra | $10.00 | $1.00 | — | $50.00
The $0.01 cache-read rate for Haiku 5.5 under 100K tokens is the number that matters most for agentic loop builders. An agent that reads the same large system prompt or codebase context on every turn pays one cent per million cached tokens — one-tenth of Haiku 4.5's cached rate, one-tenth of Sonnet 5.5's updated cached rate, and one-hundredth of Astra's.
A workload running 10,000 cached context reads per day against a 50,000-token context (well under the 100K threshold) drops from $50/day (Haiku 4.5) to $0.50/day (Haiku 5.5) on cache reads alone. That is the kind of cost reduction that changes which models you use at which tier of your agent stack — not just negotiating a better margin on existing architecture.
The 100,000-token split and what it means
Haiku 5.5 has a two-tier pricing structure: below 100K tokens, the rates above apply. Above 100K tokens, rates quintuple on input and output, and multiply fivefold on cache reads and writes.
Anthropic says approximately 90% of requests to Haiku 4.5 fell under the 100K threshold. For those workloads, the migration to Haiku 5.5 is straightforward: same model tier, 90% lower price, improved capability.
The important engineering implication: if you are building agentic workflows where system prompts, tool definitions, conversation history, or codebase context regularly push above 100K tokens, the Haiku 5.5 pricing structure is not what the headline suggests. A 150,000-token request costs $0.075 input — lower than Haiku 4.5's $0.15, but far above the $0.015 you would pay for a 100K request. Design your context management with the 100K boundary explicitly in mind.
What Sonnet 5.5's cache-read cut adds
The simultaneous Sonnet 5.5 cache-read reduction (from $0.20 to $0.10 per million) is the other half of today's announcement and deserves separate attention.
Cache reads make up a disproportionate share of token consumption in agentic workflows — the same large context (codebase, policy document, conversation history, tool library) gets read on every turn. Anthropic estimates the 50% cache-read price reduction cuts the cost of most agentic Sonnet 5.5 work by approximately 20%.
If you are currently running Sonnet 5.5-based agents at scale, the 20% reduction applies automatically from October 7 without changing any code. No migration required.
What the Haiku 5.5 compute comparison means
Haiku 5.5 is not just a repriced Haiku 4.5. The adjustable effort setting — new to the Haiku line — lets you trade compute cost for answer quality on a per-request basis. Run high-effort for complex reasoning steps, low-effort for classification, routing, and simple extraction.
The computer-use and browser-use beta in the Python and TypeScript SDKs is the more significant capability addition. Haiku 5.5 is now the cheapest Anthropic model that can operate software through its UI and interact with web pages. At $0.10 input and $0.50 output, a computer-use task that consumes 2,000 input tokens and 500 output tokens costs approximately $0.00045 — less than half a cent per action. At Opus 5.5 pricing, the same task costs $0.018.
Computer use at Haiku 5.5 pricing changes the unit economics of browser automation agents substantially. A workflow that previously required Sonnet-level capability for reliability, but where Haiku 5.5 performance meets the task requirements, drops in cost by 95%.
Surface Laptop Ultra and the local inference argument
The same day Anthropic cut cloud inference prices by 90%, Microsoft and Nvidia announced hardware that challenges the premise of cloud inference for many workloads.
Surface Laptop Ultra ($2,599 starting, ships October 16) is built around Nvidia's RTX Spark Superchip — a Grace CPU (18 or 20 cores), Blackwell RTX GPU (up to 6,144 cores), and up to 128GB unified memory in a single package. Microsoft's claim: it can run AI models above 120 billion parameters locally.
The 120B claim requires the $5,899 configuration (128GB, top-spec chip). The $2,599 entry model has 24GB — sufficient for models in the 13B-34B range depending on quantization, not for the full 120B+ claim.
The configuration that makes sense for most developer and enterprise use cases:
Config | Memory | Price | What you can run 18-core, 24GB | 24 GB | $2,599 | Up to ~34B quantized 20-core, 32GB | 32 GB | $3,699 | Up to ~70B quantized 20-core, 48GB | 48 GB | $3,999 | Up to ~90B quantized 20-core, 128GB | 128 GB | $5,899 | 120B+, frontier-scale local
The Surface RTX Spark Dev Box ($5,999, November) is the desktop version — same 128GB/20-core chip, no screen, pre-configured for development (VS Code, GitHub Copilot CLI, WSL, Python, Node.js, .NET SDK installed).
Microsoft's governance architecture for local agents — Microsoft Execution Containers for sandboxing, Microsoft Entra to distinguish agent activity from user activity, Agent 365 controls extended to on-device agents — is the piece that makes this enterprise-relevant rather than just a powerful laptop. An uncontrolled local agent is a security problem; a sandboxed, identity-attributed, policy-governed local agent is an enterprise deployment pattern.
The local inference pitch: for workloads where latency matters (real-time code completion, interactive agents), privacy matters (sensitive data that cannot leave premises), or repetition matters (running the same context thousands of times per day), the economics of a one-time hardware purchase can outperform per-token cloud pricing over a long enough horizon. At Haiku 5.5 pricing, that horizon is longer than it was yesterday. The comparison now requires more careful modeling.
The pricing landscape entering Q4 2026
In the past two weeks, the AI inference pricing table has moved significantly. Updating the full picture:
Budget tier: Haiku 5.5 and GPT-6 Luna are tied at $0.10/$0.50 input/output, $0.01 cached reads. This is the cost floor for production-grade, frontier-family inference. Gemini 3.8 Flash sits nearby.
Mid tier: Sonnet 5.5 and GPT-6.1 Sol at $2.00/$10.00, with $0.10 cached reads. This is where most enterprise agentic workloads should evaluate running.
Frontier tier: Opus 5.5, Gemini 4 Argon (at introductory prices), both at $4.00/$20.00. Gemini 4 Argon's introductory pricing will revert; Opus 5.5 pricing is stable.
Premium: GPT-6 Astra at $10/$50, Astra Ultrafast at approximately $60/$300. Reserved for latency-critical or maximum-capability workloads.
The practical implication for enterprise teams currently running Haiku 4.5 at scale: evaluate migration to Haiku 5.5 now. The 90% price cut for under-100K-token prompts is not a promotional offer — it is Anthropic's competitive response to GPT-6 Luna and the new frontier pricing floor. It is likely to hold.
For teams considering hybrid cloud/local architectures: the Surface Laptop Ultra numbers are now on the table as a real comparison point. Run the math on your workload's token volume, context sizes, and latency requirements against a hardware amortization schedule before assuming cloud-only is cheaper at the new pricing floor.
The race to the bottom in AI inference pricing is not finished. Haiku 5.5 at $0.01 cache reads is not the endpoint — it is the new baseline from which the next round of cuts will be measured.
