Anthropic released Claude Haiku 5.5 on October 7 with a price cut that rearranges the math of production AI work: about 90 percent cheaper than Haiku 4.5 on the prompts that matter. Short-context requests up to 100,000 tokens now cost 10 cents per million input tokens and 50 cents per million output tokens, with cache reads at a penny. Above the 100k line the discount drops to 50 percent. Anthropic estimates the average workload lands roughly 75 percent cheaper once a new tokenizer is factored in, and this is the first small Claude to ship with an effort dial, so you can trade cheap for smart per task instead of per model.
That matters because the small, boring work is where the volume lives. Summarization, classification, ticket routing, field extraction. Nobody runs those once; you run them ten thousand times a day, and at that scale a 90 percent cut is the difference between a project that pencils out and one you shelved on cost. The short-context rates match OpenAI's GPT-6 Luna token for token, though Luna's higher tier starts above 272,000 tokens and comes out cheaper on list price for a 150,000-token prompt. For builders, the practical question is no longer which frontier model is smartest. It is which routing scheme puts the cheapest adequate model in front of each call.
Speaking of putting models to work, Google Cloud introduced its Gemini agent for work on Thursday: a single persistent agent that plans work, calls tools, plugs into business systems, and returns finished output in documents, email, and developer environments. It lives inside Gmail, Docs, Sheets, and Calendar, and also works through Microsoft 365 and Slack. Two things stand out. First, the operational framing: coworker agents get their own email address, narrow permissions, and spending caps. Google is selling autonomy with the guardrails attached. Second, the agent picks its own model per task and runs on Gemini and Anthropic's Claude models today. A Google agent that ships with Claude inside is a quiet admission that the model layer has commoditized faster than anyone's marketing will admit.
On the infrastructure side, the White House announced Thursday that 11 companies have committed 2.4 billion dollars in AI resources to the Genesis Mission Consortium, aimed at federal science challenges in energy, health, and space. Nvidia put up 1 billion, AMD 500 million, OpenAI 200 million, and Anthropic and Google 150 million each, with AWS, Crusoe, Micron, and others filling out the rest. Compute credits and AI-for-science tools for more than 15 federal agencies. Read it as patriotism or procurement strategy; either way, the frontier labs now compete partly on who can furnish the state's research apparatus.
The rest of the week filled in around the edges. Mistral released a public preview of Mistral Large 4, a natively multimodal mixture-of-experts model with 1.05 trillion total parameters and 49 billion active, with open weights promised for the end of October. DeepSeek open-sourced its first native vision model under the MIT license. And Tencent is reportedly weighing a 5 billion dollar bond sale to fund its AI buildout, after SoftBank raised 11.1 billion in junk bonds last month for the same purpose. Debt markets are now a standard instrument of the AI arms race. The through-line for anyone building this quarter: tokens keep getting cheaper, agents keep getting more autonomous, and the hard part keeps moving up the stack, from the model to the harness to the process the harness automates.
