The AI log, frontier · open weights · science · agents and coding
10 Nov 2026
Gemini 4 Carbon, AlphaProof Code, MiniMax H3 Image, Scribe v3
Carbon goes public, DeepMind verifies Signal's code in Lean, MiniMax opens an image model, Agent Substrate, Scribe v3 and a weight backdoor.
By Ali Asif, published
What shipped
- Google
Gemini 4 Carbon is now available
Google's strongest model opens to paying developers and enterprises at $2 and $10 per million tokens until November 30, then $4 and $20.
- Google DeepMind
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
Agents wrote over six million lines of Lean proofs for Signal's compiled cryptographic code in four weeks, with humans reviewing about 1%.
- MiniMax
MiniMax-H3-Image
A 33.1B-parameter diffusion transformer for text-to-image and editing, distilled to 6 sampling steps with output up to about 16 MP, released under a community license.
- NVIDIA
Agent Substrate
An open source runtime donated to the CNCF Sandbox that checkpoints idle agent sessions and restores them onto free Kubernetes workers.
- ElevenLabs
Introducing Scribe v3 and v3 Live
Speech-to-text models with 27% fewer word errors than Scribe v2 across 102 languages on ElevenLabs' own test, plus realtime diarization.
- Microsoft Research
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft shows that reprogramming fewer than 10 dormant neurons can backdoor an agent, and argues for provenance checks and limited agent permissions.
Six primary pages I opened today: a frontier model going public, a formal verification result, an open image model, an agent runtime, a speech model and a security paper. Everything below comes from those pages, not from my own testing.
Gemini 4 Carbon
Google calls Carbon its strongest model, with frontier performance in cybersecurity and long-horizon agentic work. It is on the Gemini API paid tier in AI Studio and Android Studio, in Gemini for business, and for Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. The introductory price matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens. From November 30 it moves to $4 and $20.
The benchmark claims are mostly relative. Google reports first place on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than twice lower cost per task than the next external model, and second on a multimodal agent benchmark. Cyber capability stays limited to the Fairwind Program. The part I trust most is the caveat: Google says Argon suits most software and knowledge work, because Carbon spends more tokens exploring and verifying, and that prompt injection robustness is still an area for improvement.
AlphaProof Code and Signal
DeepMind used agents powered by Carbon to write more than six million lines of Lean proofs about the compiled code behind Signal's cryptographic protocol. The work took four weeks, and agents flagged about 1% of the codebase for human review. Humans wrote and reviewed every specification, and Lean checks every proof. As a test, the team reintroduced two old bugs into an isolated copy, and the system produced a machine-checked proof that each buggy version was insecure. The verification artifacts are open source. The page does not say whether the framework itself is available.
MiniMax H3 Image
This is an open weights model for text-to-image and editing in one checkpoint. The diffusion transformer has 33.1B parameters and is distilled to 6 sampling steps with no guidance setting. Outputs reach about 16 MP, and editing takes up to 16 reference images. The weights are about 68.65 GB, and the text encoder is a separate 66.7 GB download. The license is a community license that needs separate written authorization for some uses, so read it before building on it.
Agent Substrate
NVIDIA describes an open source runtime, donated to the CNCF Sandbox, that runs many agent sessions on a small pool of warm Kubernetes pods. Idle sessions are checkpointed and restored onto any free worker. The 10x density claim is NVIDIA's. APIs are still changing, and GPU suspend and resume is only a proposed design.
Scribe v3
ElevenLabs reports 27% fewer word errors than Scribe v2 across 102 languages on an internal FLEURS-based test, and Scribe v3 Live adds realtime diarization. No pricing is stated.
Lazarus
Microsoft Research shows that reprogramming fewer than 10 dormant neurons can backdoor an agent without retraining. The attack needs write access to the weights, which is why the recommendations center on provenance and limited agent permissions.
What I'd watch next
I want to see Carbon's token use on real coding work before November 30, and whether the AlphaProof Code framework opens up. I would also watch how Agent Substrate handles GPU restores.
Sources
- blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
- deepmind.google/blog/introducing-alphaproof-code
- github.com/MiniMax-AI/MiniMax-H3-Image
- github.com/agent-substrate/substrate
- elevenlabs.io/blog/scribe-v3
- www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents
Topics: NVIDIA, Google DeepMind, Google, Gemini, DeepMind, ElevenLabs, Kubernetes, Microsoft