Three model launches, Wayve’s Mercedes-Benz production agreement, Unit 42’s continuous AI security service, and Snorkel’s funding round.

Stories and sources

Anthropic releases Claude Opus 5.5

Opus 5.5 launches with revised pricing and a gradual GitHub Copilot rollout.

OpenAI launches GPT-6 Sol and Luna

The models reach the API, ChatGPT Work and Codex, with separate Copilot plan eligibility.

Xiaomi releases MiMo 2.6 and updates MiMo Code

Xiaomi reports training gains and adds tool sequencing and multimodal file reading.

Wayve and Mercedes-Benz sign production agreement

Driving assistance is planned for future vehicles within two years.

Unit 42 launches continuous AI security testing service

Palo Alto Networks announces worldwide annual subscriptions.

Snorkel announces $350 million Series E

The AI data provider reports a $3.5 billion valuation and plans expanded capacity.

This episode uses an AI-generated narrator.

Grok coding and transcription releases, OpenClaw updates, HackerOne’s planned Mythos integration, and Lovable’s acquisition of Sutro.

Stories and sources

SpaceXAI releases Grok 4.7

The coding and knowledge-work model is available through the API, Cursor and Grok Build, with separate pricing and access conditions for its Fast variant.

SpaceXAI details Grok Voice Transcribe 2.0

The company reports multilingual transcription improvements; its documentation disagrees about the default model.

OpenClaw adds update validation and agent setup features

Version 2026.9.5 adds candidate-update checks, specialist setup and plugin installation without Gateway restarts.

HackerOne announces planned Mythos integration

HackerOne plans Mythos support for repository audits and code reviews, with validation controls.

Lovable announces Sutro acquisition

Sutro’s founder and three engineers join Lovable; the announcement does not disclose a price.

This episode uses an AI-generated narrator.

FormalFlow’s supervised proof work, Mila and Mozilla’s local-AI initiative, Microsoft’s agent security changes, Talos’s Qilin findings and CoreWeave’s enlarged debt offering.

Stories and sources

FormalFlow paper reports supervised AI proof formalization

Researchers report a machine-checked component theorem, with human oversight of mathematical meaning.

Mila and Mozilla announce local-AI foundation initiative

The partners plan open interfaces and locally installable implementations, with initial working implementations expected within six months.

Microsoft Agent Framework updates MCP security handling

Python 1.19.0 changes session scoping and archive handling.

Talos reports indications of AI-generated Qilin scripts

Talos gives a qualified assessment of possible AI-assisted code generation.

CoreWeave prices enlarged $3.7 billion debt offering

The AI cloud provider scheduled settlement for September 22, subject to closing conditions.

This episode uses an AI-generated narrator.

Qwen's multimodal tools, Anthropic's evaluation partnership and life-sciences programme, PrismML's compressed model, Wispr's Canto, and OpenAI's incident disclosures.

Stories and sources

Alibaba details Qwen3.8-Omni-Flash and accompanying tools

Alibaba describes multimodal inputs, text output and accompanying tools for audio-video agent workflows.

• Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery. — Alibaba Cloud
• Qwen3.8-Omni-Flash — QwenCloud

Anthropic and Accenture announce embedded evaluation partnership

Evaluators will work inside Anthropic; funding and operating arrangements remain under development.

• Partnering with Accenture on embedded evaluation — Anthropic
• Accenture and Anthropic Partner to Build Team of Embedded Evaluators at Anthropic — Accenture

PrismML releases Ternary Bonsai 2 27B

PrismML releases compressed model weights and reports benchmark comparisons with the full-precision base model.

• Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint — PrismML
• PrismML Launches Bonsai 2 27B, Its Most Capable Model Yet — PrismML via PR Newswire

Wispr introduces Canto and deploys it across Flow

Wispr reports dictation benchmark results, including limitations on challenging and public datasets.

• Canto: a speech model built for the real world — Wispr
• What's new — Wispr

Anthropic opens Life Sciences Verification Program applications

The beta offers verified teams biology access under differentiated grants.

• Introducing the Life Sciences Verification Program — Anthropic

OpenAI publishes reporting framework and six misalignment disclosures

OpenAI describes individual training and evaluation incidents and plans faster disclosure.

• Our framework for reporting model misalignment — OpenAI

This episode uses an AI-generated narrator.