AI News

AI news brief: Friday, October 30, 2026

27 stories, most important first.

OfficialBusiness & Funding · · Updated

Cohere closes CAD $7B Series E at CAD $35.7B valuation

Sovereign AI platform Cohere raised a CAD $7B Series E at a CAD $35.7B valuation, with 165% year-over-year ARR growth.

Why it matters: Regulated buyers get a well-funded vendor that can run models on premises, which matters for firms weighing data residency against managed APIs.

2 sources
ReportedSafety & Security ·

DeepSeek and Moonshot reportedly train models on GPT-6 Astra outputs

Reports say DeepSeek and Moonshot AI use GPT-6 Astra outputs to train new models after Anthropic tightened defenses against distillation.

Why it matters: Model-output theft is emerging as a competitive and security issue, which could lead to stricter access controls and terms of service across frontier model providers.

1 source
ReportedPolicy & Regulation ·

Three California AI employment laws converge on January 1, 2027

FEHA, CCPA and SB 947 apply together from January 1, 2027, covering discriminatory automated decision systems, pre-use notices and human corroboration of discipline.

Why it matters: Employers using AI in hiring or discipline face notice, opt-out and record-keeping duties, with fines for noncompliance.

1 source
ReportedSafety & Security ·

OpenAI expands ChatGPT violence safeguards and commits to public safety reporting

OpenAI says it is strengthening violence prevention in ChatGPT through expanded safeguards, earlier interventions, and a new public safety reporting commitment.

Why it matters: Companies that deploy ChatGPT may get questions from their risk and compliance teams about how the platform handles violent content. The public safety reporting commitment gives buyers a new item to check during vendor reviews.

1 source
ReportedResearch ·

Drew Breunig bakes a working NES emulator into Qwen3.5-4B's weights

Drew Breunig says he embedded a working emulator for the NES 6502 CPU in the weights of Qwen3.5-4B, which can run NES ROMs such as Super Mario Bros very slowly.

Why it matters: The project shows that fine-tuning can encode deterministic program execution into a model's weights. The demo runs very slowly, so it is a research curiosity rather than a practical way to run software today.

1 source
ReportedResearch ·

Microsoft paper shows history-aware routing helps stateful agents choose cheaper models

A Microsoft paper finds that letting a router read session history improves agent task success while deciding which turns go to a cheaper model.

Why it matters: Companies routing agent traffic to cheaper models should judge them on completed sessions, since per-token prices omit retries, extra calls and resent history. Switching models mid-session can also add cache costs.

1 source
ReportedSafety & Security ·

Anthropic says Claude now produces a misuse assessment after each response

Anthropic says the per-response assessment has surfaced misuse its classifiers were not designed to catch, and that the assessment improves as the model does.

Why it matters: If model-generated misuse assessments become standard, platform teams may need to build them into risk controls. The report, however, gives no detection rates.

1 source
ReportedProducts ·

Claude API raises maximum request size from 32MB to 128MB

The Claude API's maximum request size rose from 32MB to 128MB, and the per-image limit rose to 3000px for requests with more than 20 images.

Why it matters: Larger payloads let document-heavy and multi-image workloads run in fewer requests, which can simplify pipelines for document review and visual analysis.

1 source
RumorModels ·

Rumors say Qwen4 launches next week after Qwen's CTO remarks

Reports say Qwen's CTO pointed to a launch next week and a code change for the Qwen4 family has landed, but neither is confirmed.

Why it matters: If a 27B variant ships, it could matter for teams that run models locally, but nothing official has been published yet.

1 source