Quick answer

October 6 to 10, 2026, in order: Mistral Large 4 preview, Nano Banana 2.1, and the Personal Agent Protocol (Oct 6); Claude Haiku 5.5, GPT-6 to all ChatGPT tiers with Intelligent UI, Perplexity's open pplx-embed-v2-late models, and Nous Research's $90M raise (Oct 7); Free and Go tiers get GPT-6 Luna, Google's Gemini coworker agents in private preview, Grok Imagine Video 1.5 Lite listings, and a $500M+ round for Manus's parent (Oct 8); Liquid AI's open d1 decision models (Oct 9); Cloudflare acquiring Deno (Oct 10). Themes: the budget tier got cheap, interfaces became generative, and agents got identities.

After the flagship launches of September and the DevDay wave at the start of October, this was a week of filling in the rest of the picture: the small models, the free tiers, the standards, and the plumbing.

Models

  • Oct 6 — Mistral Large 4 (Mistral): ~1T-parameter MoE with vision in API preview at $1.36/$4.18, launch sale at half; open weights promised by end of October
  • Oct 6 — Nano Banana 2.1 (Google): image generation and editing update with 4K, 14 reference images, better text, and reported lower output pricing
  • Oct 7 — Claude Haiku 5.5 (Anthropic): $0.10/$0.50 per million tokens up to 100K-token prompts, effort setting, computer use in beta
  • Oct 7 — pplx-embed-v2-late (Perplexity): MIT-licensed late-interaction embeddings (0.6B and 9B) that search PDFs and slides without OCR
  • Oct 8 — Grok Imagine Video 1.5 Lite (xAI): distilled video model from about $0.02 per second, per gateway listings
  • Oct 9 — d1-3B and d1-omni-600M (Liquid AI): open decision models that answer in one forward pass, 8 ms on an RTX 4090 by the company's figures

Products and agents

  • Oct 7–8 — GPT-6 for every ChatGPT tier (Sol for paid, Luna for Free and Go) with Intelligent UI: replies that include charts, forms, and tools
  • Oct 8 — Gemini agent for enterprise (Google Cloud): coworker agents with their own Workspace account, @agents email, calendar, and attested identity, in private preview
  • Oct 6 — Personal Agent Protocol (Sierra and Meta, with Walmart, Shopify, Stripe): an OAuth-based standard for personal agents dealing with businesses; spec due later in October
  • Oct 7 — OpenAI published 722 AI-generated maths manuscripts from an unreleased internal model to a public repository

Money and infrastructure

  • Oct 7 — Nous Research raised $90M at a reported $1.5B valuation for Hermes for Businesses
  • Oct 8 — Butterfly Effect, parent of the Manus agent, reportedly closed a round of more than $500M led by Boyu Capital and IDG
  • Oct 8 — Atomic Machines left stealth with $250M for a "Matter Compiler" that fabricates micro-machines from code, targeting data-centre cooling
  • Oct 10 — Cloudflare acquired Deno; the Deno team moves to Workers and the standalone runtime and Deno Deploy wind down over a year
  • Oct 10 — Anthropic updated its usage policy, effective November 12, adding a section on not undermining democratic processes

What it adds up to

Three threads. The budget tier is now very cheap: Haiku 5.5, DeepSeek V4.1 Flash, and Mistral's launch sale put capable models at or under $0.70 per million output tokens, and the surcharges above 100K to 272K tokens are where the real price differences hide. Interfaces are becoming generative: ChatGPT replies now draw their own charts and forms for a billion-plus users. And agents are getting identities: a Google agent with an inbox, a Sierra and Meta protocol for signing in to shops, and a Hermes that enterprises can audit. The week's quietest item, Perplexity's open embeddings, may matter most to builders: searching slides and PDFs without OCR removes a whole pipeline.

Benchmark, speed, and adoption figures in this roundup are vendor-reported unless an independent source is named. The Mistral and Haiku prices were checked against vendor pages and launch coverage on October 12.

Related reading

Early October 2026 roundup Claude Haiku 5.5 released GPT-6 and Intelligent UI for all users

Bottom line

Re-quote your budget-tier workloads against Haiku 5.5 and V4.1 Flash, try an interactive ChatGPT reply before deciding whether you like generative UI, and if you run a business, read the Personal Agent Protocol when the spec lands. Next up: Mistral's weights and Kling 4.0's full release, both promised for this month.