Claude Haiku 5.5: Anthropic's fastest small model, at $0.10 / $0.50 per million tokens
Anthropic released Claude Haiku 5.5 on 2026-10-07, available now on the Claude Platform, AWS, Google Cloud and Azure. Pricing is $0.10 input / $0.50 output per million tokens for prompts up to 100,000 tokens and $0.50 / $2.50 above that. It is the first Haiku with an adjustable effort setting, and Anthropic says it is the fastest model it has released and averages about 75% cheaper to run than Haiku 4.5, though an updated tokenizer uses slightly more tokens per task. Vendor-reported gains over Haiku 4.5 are large (Terminal-Bench 4.0 39.2% vs 0.0%, OSWorld 2.1 offline subset 72.4% vs 15.7%, Humanity's Last Exam without tools 45.9% vs 10.2%); these figures are not independently reproduced, and the context window is not stated. Anthropic positions it for high-volume, narrowly scoped work and says Sonnet 5.5 and Opus 5.5 remain better for complex agentic coding.
Sources & depth
NVIDIA and Microsoft launch RTX Spark PCs and OS-level containers for Windows agents
At a joint event on 2026-10-07, NVIDIA and Microsoft announced RTX Spark, a Windows PC platform pairing a Blackwell RTX GPU (up to 6,144 cores) with a Grace CPU of up to 20 cores and up to 128GB of unified memory, rated at one petaflop of FP4. Laptop preorders opened 2026-10-07 with availability on 2026-10-16; compact desktops follow in November, from Acer, ASUS, Dell, HP, Lenovo, MSI, Gigabyte and Microsoft's own Surface Laptop Ultra. No prices were given. Microsoft also made Microsoft Execution Containers (MXC) generally available, OS-level isolation so agents can run persistently in the background under governance, and NVIDIA previewed DGX Station for Windows (GB300, 748GB coherent memory, up to 20 petaFLOPS FP4) with no date or price.
Sources & depth
Report: Meta and Microsoft cut internal use of Anthropic's Claude as they shift to their own coding tools
The Information reported on 2026-10-05, as relayed by a secondary write-up, that both companies are reducing employee use of Claude. Microsoft is said to have cut per-employee monthly AI spending limits in its cloud and AI division from about $100,000 to about $10,000 (ceilings, not actual spend) and steered staff toward GitHub Copilot and OpenAI tooling. Meta's Claude Code users reportedly fell from about 60,000 to about 30,000 as employees moved to in-house tools, though it still reportedly spent over $105 million on Claude Code over 28 days. Customer spending on Anthropic models through Microsoft's platforms is reported to keep rising. This is single-source and unconfirmed by either company.
Sources & depth
Unsloth adds training of "decision models" from any text or vision LLM
Unsloth's v0.1.904-beta release lets users turn a text or vision LLM into a Jev-style decision model that scores each option with a confidence, with train, test, GGUF export and llama.cpp serving built in. Unsloth claims decision accuracy rises from 30% to 80%, but the release does not describe the benchmark or conditions, so treat the figure as a vendor claim. The release also adds native ComfyUI diffusion model loading and reports up to 4.1x faster Qwen3.5-35B-A3B QLoRA training on A100 and RTX PRO 6000. LiquidAI separately published open d1 decision models for the edge.
Sources & depth
OpenAI says GPT-6 is rolling out globally in ChatGPT with "Intelligent UI"
OpenAI's announcement says GPT-6 is rolling out globally in ChatGPT with Intelligent UI, which it describes as faster responses with visuals and interactive experiences users can explore directly. The page was unfetchable, so tier availability and further details are link only.
Sources & depth