AI Round-up: 7 October 2026

Mistral introduced a public API preview of Large 4, its trillion-parameter multimodal model, with weights promised for later this month rather than available now. OpenAI published mathematical manuscripts and proof artefacts from an unreleased frontier model; verification varies across the collection, so the release is not blanket confirmation of every claimed result. Anthropic expanded its Cyber Verification Program with tiered access to advanced Claude models for vetted security work, while Microsoft examined the practical burden of AI-driven vulnerability discovery.

Developer updates included OpenAI's Decisions API beta and revised API usage tiers, plus GitHub's ReviewBench for AI code reviewers. News also covered AI infrastructure demand, HSBC's reported AI-related job-cut plans and coordination around business-facing AI agents. YouTube mixed safety scrutiny with coding and Claude-skills demonstrations, while Reddit discussions focused on coding-assistant throughput, usage limits and agent workflows.