Everything that mattered in AI, one page a week.

Z.ai released GLM-4.6 on September 30 with open weights under the MIT license on Hugging Face: the GLM-4.5 mixture-of-experts architecture — 355B total parameters with 32B active per token — with the context window expanded from 128K to 200K tokens and maximum output raised to 128K. Z.ai reports gains over GLM-4.5 across eight public benchmarks covering agents, reasoning and coding, and says GLM-4.6 beats Claude Sonnet 4 in its own 74-task real-world coding suite run inside the Claude Code environment, while trailing Claude Sonnet 4.5 on coding.

Two things separate this from a routine model drop. First, published pricing of $0.60 per million input tokens and $2.20 per million output, against Sonnet 4.5's $3/$15 — roughly a fifth on input and a seventh on output — with self-hosting free under MIT if you have the GPUs. Second, Z.ai published all 74 CC-Bench agent trajectories, prompts and tool calls included, as a dataset, which lets you inspect failure modes instead of trusting a bar chart.

DeepSeek published DeepSeek-V3.2-Exp on September 29: a 671B-parameter mixture-of-experts model with 37B active parameters per token, a 128K context window and MIT-licensed weights, posted to Hugging Face alongside a technical report and TileLang and CUDA kernels. It is V3.1-Terminus plus DeepSeek Sparse Attention, which uses a learned lightning indexer to score the context and compute attention over roughly the top 2,048 tokens per query instead of the whole sequence. DeepSeek reports benchmark parity with the dense-attention predecessor — 85.0 MMLU-Pro on both, Codeforces 2121 versus 2046 — while processing long inputs two to three times faster.

The price change shipped with the model rather than after it. DeepSeek's release post says API prices dropped more than 50% effective immediately, to about $0.28 per million input tokens, $0.028 per million cached input tokens and $0.42 per million output tokens, with the old V3.1-Terminus endpoint kept alive for comparison only until 15 October. The sparse-attention kernels are released as readable TileLang code plus production CUDA, so the efficiency trick is reproducible on your own serving stack rather than locked behind the vendor API.

Anthropic released Claude Sonnet 4.5 on September 29 — the same day DeepSeek's sparse-attention checkpoint dropped — claiming state of the art on SWE-bench Verified at 77.2% and 61.4% on OSWorld computer use, up from 42.2% for Sonnet 4 four months earlier. Pricing was held flat at $3 per million input and $15 per million output tokens with a 200K context window, and the model shipped under ASL-3 protections with CBRN classifiers that Anthropic says have cut false positives tenfold since their introduction. Anthropic reports sustained focus of more than 30 hours on multi-step tasks.

The more consequential half was the harness. Anthropic gave developers the Claude Agent SDK — the same scaffolding behind Claude Code — plus checkpoints with instant rollback in Claude Code, a native VS Code extension, and context editing and a memory tool in the API. Those are the unglamorous primitives of a production agent stack: durable memory, cheap context compaction, and a way to rewind a run without re-paying for it.

Governor Gavin Newsom signed Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, on September 29, a year after vetoing its predecessor SB 1047 over fears it would chill California's AI economy. The law binds large frontier developers — anyone training a foundation model above 10^26 floating-point operations — to publish a frontier AI framework covering catastrophic-risk assessment, mitigation, internal governance and weight-security measures, and to report critical safety incidents to the Office of Emergency Services within 15 days, or 24 hours where there is an imminent threat. It adds whistleblower protections, empowers the attorney general to levy civil penalties of up to $1,000,000 per violation, establishes the CalCompute public-computing consortium, and takes effect January 1, 2026.

It is a transparency statute, not a liability one: nothing in SB 53 creates new causes of action for harms caused by model outputs. What it does create is a written, dated, publicly posted artefact that companies can be measured against, and an incident-reporting channel into state government, at a scale threshold high enough that it currently catches a handful of labs rather than the deployer ecosystem.

GitHub blocked creation of new server-side Copilot Extensions (the GitHub App based flavor) at 8:00 AM PST on Wednesday and will disable all existing ones on 10 November 2025 after a brownout window from 3-7 November. The replacement is Model Context Protocol servers, steered through the GitHub MCP Registry, with per-server allowlisting for IDEs arriving from 28 October. Client-side VS Code Copilot Extensions and ordinary GitHub Apps are untouched.

The stated rationale is the part that matters: one MCP server works with GitHub Copilot, Claude Code and any other MCP-compatible host, removing the build-once-per-platform tax that proprietary extension APIs imposed. GitHub is a major client vendor trading its own integration format for Anthropic's emergent protocol — the clearest signal yet that MCP is becoming the default tool interface for coding agents rather than an Anthropic-flavored option.

Alibaba launched Qwen3-Max at its Apsara conference in Hangzhou: over 1 trillion parameters, trained on roughly 36 trillion tokens, and available only through the Alibaba Cloud API with no weights published — the first Qwen to cross a trillion parameters and the largest to stay proprietary. Alibaba reports 69.6 on SWE-bench Verified for the Instruct variant and says the Thinking variant, still training at launch, reaches 100% on AIME 2025 and HMMT once tool use and scaled test-time compute are added. It also advertises strong Tau2-Bench results for multi-turn tool use.

The same keynote open-sourced Qwen3-Omni-30B-A3B on Hugging Face and ModelScope, previewed four Wan 2.5 generation models, and upgraded Qwen3-Coder and Qwen3-Image-Edit. Alibaba Cloud CEO Eddie Wu repeated a RMB 380 billion three-year plan for AI and cloud infrastructure and said the company has open-sourced over 300 models since 2023. Read together with Tuesday's release, the split is deliberate: permissive weights at 235B for vision, a closed API at 1T for the text frontier.

Qwen published Qwen3-VL-235B-A22B in Instruct and Thinking variants: a mixture-of-experts model with 235 billion total parameters and roughly 22 billion active per token, released under Apache 2.0 on Hugging Face with weights weighing about 471 GB. Native context is 256K tokens, expandable to 1M, with second-level timestamp grounding for long video, 3D grounding for spatial reasoning, OCR across 32 languages, and GUI action planning for computer and mobile interfaces. Qwen reports DocVQA 96.5 in Thinking mode plus MathVista 85.8 and MathVision 74.6, and claims the Instruct build matches or beats Gemini 2.5 Pro on major visual perception benchmarks.

The same day, Qwen shipped Qwen3Guard under Apache 2.0 at 8B, 4B and 0.6B parameters, an openly licensed moderation series in the Llama Guard mould. The license carries no usage cap, no field-of-use restriction and no revenue threshold, so commercial fine-tuning and redistribution are permitted outright. Smaller dense and MoE Qwen3-VL sizes followed in October, which matters because the 235B flagship is not something most teams can hold in VRAM.

NVIDIA and OpenAI announced a letter of intent on Monday for at least 10 gigawatts of NVIDIA systems across OpenAI's data centers, with NVIDIA intending to invest up to $100 billion progressively as each gigawatt is deployed. Huang framed the scale as roughly 4 to 5 million GPUs; NVIDIA's own August earnings maths prices one gigawatt of capacity at $50-60 billion, about $35 billion of it NVIDIA hardware. First phase is targeted for the second half of 2026 on the Vera Rubin platform.

The structure is circular: OpenAI pays NVIDIA cash for chips, NVIDIA buys non-controlling equity at then-current valuations, and the models are trained and served on the same vendor's racks. On Tuesday OpenAI, Oracle and SoftBank added five Stargate sites — Lordstown, Ohio; Milam and Shackelford counties in Texas; Doña Ana County, New Mexico; and an unnamed Midwest site — bringing Stargate to nearly 7 gigawatts of planned capacity and more than $400 billion of investment over three years. OpenAI's announcement pegs its user base at over 700 million weekly active users.