OpenAI and Anthropic both cut their LLM API prices by up to 80% over the past few months. That’s not a gradual decline. That’s a freefall.

Max Woolf dug into the numbers. GPT-4o dropped from $2.50 per million input tokens in May to $0.50 in October. Claude 3.5 Sonnet went from $3 to $0.80 for cached tokens. Even the flagship models got cheaper: GPT-4 Turbo fell from $10 to $2.50, and Claude Opus went from $15 to $3.

The obvious explanation is competition. When you’re racing to keep market share, you cut prices. But there’s more to it than that. These companies are also getting better at inference optimization. Smaller distilled models, better quantization, smarter caching. The cost of serving a token actually dropped.

This matters for anyone building on LLM APIs. If you’ve been holding off on a feature because token costs were too high, check the current pricing. What was expensive in June might be cheap now.

Microsoft Launches Decision-1

Microsoft just released Decision-1, a 9B parameter reasoning model based on Qwen3.5. It’s small, fast, and designed for structured decision-making tasks.

The model uses a refined training approach focused on multi-step reasoning. Think less “write me a blog post” and more “should we approve this loan application based on these criteria.” Microsoft says it outperforms larger models on specific decision tasks, though they haven’t released full benchmarks yet.

Decision-1 is available through Azure AI. Pricing isn’t public, but given the 9B size, it should be cheaper to run than frontier models. If you’re building approval workflows, classification systems, or anything that needs explainable reasoning steps, it’s worth testing.

IBM’s CEO Talks Up Open Models

IBM CEO Arvind Krishna told The Register that open weight models are now “irresistibly good and cheap” for on-premises deployment. That’s a strong endorsement from a company that sells enterprise infrastructure.

Krishna’s point is that businesses can now run capable models on their own hardware without spending millions on API calls or worrying about data leaving their network. For regulated industries, that’s huge. Banks, healthcare providers, and government agencies have been slow to adopt LLMs specifically because of data residency requirements.

The economics are shifting. If you can run Llama 3.3 70B on your own GPUs and get 90% of the performance of a frontier model, the total cost of ownership drops fast. You pay for hardware once instead of paying per token forever.

A Few Quick Hits

Proc-CLI launched on Hacker News today. It’s a CLI tool for managing processes and ports, designed to be agent-friendly. That means structured output, consistent exit codes, and JSON responses. If you’re building automation that needs to kill processes or check what’s running on a port, it’s worth a look. GitHub

PaperPilot is a zero-cloud PDF engine written in Rust with WASM and MCP support. It runs entirely in the browser, which means you can parse and manipulate PDFs without sending them to a server. The MCP integration lets AI agents interact with PDFs directly. GitHub

Seam Cutter is a weird little tool from Simon Willison that lets you crop out a section from the middle of an image by sliding the two halves together. It’s surprisingly useful if you stitch screenshots together and need to remove something from the middle without re-cropping everything.

The through-line this week is infrastructure. Prices are dropping, open models are getting viable for production, and new tools are making it easier to run AI locally or in constrained environments. If you’ve been waiting for the economics to make sense, they’re starting to.