AI Digest

Category

Google releases Gemini 4 Carbon, its strongest model, to developers and enterprises

TLDR AI

Google is claiming top spots on agentic coding and automation benchmarks at a lower cost per task than any other external model.

  • Available through the API in AI Studio and Gemini for business, and rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search; priced at $2/$10 per million input/output tokens, the same as Gemini 4 Argon.
  • Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next best external model.
  • The release adds frontier cybersecurity capabilities through the Fairwind Program and previews /swarm, an Antigravity mode where named agents split a task and coordinate on a shared message board.
Read on blog.google →
Benchmarks
BenchmarkGemini 4 CarbonClaude Fable 5.2GPT-6.2 AstraMuse WonderKimi K3.1 ProΔ vs best other
GPQA diamond96.596.096.8—94.2-0.3
FrontierMath-Tiers-1-3-v2-Private90.595.192.682.178.2-4.6
FrontierMath-Tier-4-v2-Private95.1100.0100.070.765.9-4.9
OTIS Mock AIME 2024-2025100.099.7100.099.4100.0+0.0
SimpleQA Verified77.794.078.5—62.2-16.3
Chess Puzzles78.079.078.0—45.0-1.0
EBR-bench87.690.592.431.437.1-4.8
Mystery Game Puzzles73.090.095.051.042.0-22.0
ARC-AGI98.598.598.5——+0.0
SimpleBench92.694.979.2——-2.3
APEX-Agents84.280.172.266.165.5+4.1
ARC-AGI-294.293.895.8——-1.6
PostTrainBench41.044.6—24.2—-3.6
LMCA65.770.6—56.857.6-4.9
DTBench98.499.2—95.293.3-0.8
Furniture Assembly77.590.089.254.255.8-12.5
FrontierSWE61.269.9———-8.7

Δ = Gemini 4 Carbon minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

Anthropic launches its IPO roadshow, offering 825M shares at $70–$78

TLDR AI

The company has moved from preparing for an IPO to setting a price range and a Nasdaq ticker.

  • Anthropic plans to offer 825 million shares priced at $70–$78 each.
  • It applied to list on Nasdaq under the ticker ANT, and most proceeds from newly issued shares will go to Anthropic.
  • The Rundown reports the IPO is expected next week.
Read on anthropic.com →

Anthropic says it now considers its AI R&D capability threshold met

The Rundown AI

A frontier lab is formally saying its models are reaching the point where they can meaningfully speed up AI research.

  • The statement appeared in Anthropic's updated IPO filing.
  • Anthropic based the judgment on what it expects from upcoming models.
Read on sec.gov →

AI 2027 authors move their superhuman AI researcher forecast back to 2027

The Rundown AI

The forecasters behind AI 2027 had pushed their dates into the 2030s and now say a recent METR estimate changed their minds.

  • Their biggest reason is METR's estimate that Anthropic now improves about 2x as fast, mostly thanks to AI. The forecasters' models gave a pace like that under 4% odds of showing up this early.
  • Daniel Kokotajlo gives a 50-50 chance of a Superhuman AI Researcher by May 2027, while the other five forecasters land between August and November.
  • The group predicts AI that can automate all white-collar work before 2028 and says that if its forecast holds, a pause should be the AI industry's top priority.
Read on blog.aifutures.org →

OpenAI launches Businesses in ChatGPT, with brand profiles and customer-facing agents

TLDR AI

Brands get an official way to manage how they show up inside ChatGPT and to talk to shoppers there.

  • Companies can claim verified profiles, provide product feeds with pricing and availability, and track how often ChatGPT mentions them.
  • Eligible US e-commerce and online service businesses can launch Business Agents, and brands can now pay to make their agent the landing spot for ads in ChatGPT.
  • The launch is in beta, and profiles are currently free to claim for eligible companies.
Read on openai.com →

France unveils the Paris Call for AI Safety and Security, with OpenAI, Anthropic and Mistral signing on

The Rundown AI

Frontier labs are publicly committing to emergency off switches in their most advanced models.

  • Signers commit to building emergency off switches into their most advanced models.
  • Signers include OpenAI, Anthropic and Mistral.
Read on elysee.fr →

OpenAI acquires AI travel-planning startup Odessia

TLDR AI

The deal looks like an acqui-hire and suggests OpenAI is interested in travel.

  • Odessia is a months-old AI trip-planning startup founded by Sonder founder Francis Davidson, who is joining OpenAI to work on travel.
  • OpenAI has not announced any travel-specific products or plans.
Read on techcrunch.com →

More than 35,000 Grok-built apps were open to anyone with the link

TLDR AI

The exposure shows the privacy risks of AI app builders that share by link by default.

  • The exposed apps included SpaceXSI's own staff tools.
  • One exposed app contained an ICU patient's lab results.
Read on blog.claroc.io →

Anthropic adds Claude Sheets, Whiteboards and Data, plus embedded partner apps

TLDR AI

Claude now covers spreadsheets, whiteboards, dashboards and meetings, so teams can stay in one app.

  • Sheets builds working spreadsheets with formulas and exports them to Excel, Whiteboards lays plans out on a shared board, and Claude Data (formerly Dashboards) is now generally available.
  • Interactive connectors run apps like Adobe and Atlassian directly inside Claude, and plugins from Salesforce, NetSuite and Workday (in beta) ship with built-in skills for sales, finance and HR work.
  • Claude Meetings, which joins calls to take notes and update documents, is coming soon, with a Google Meet beta for Team and Enterprise users.
Read on claude.com →

Microsoft releases MAI-Transcribe-2 speech recognition at $0.10 per audio hour

TLDR AI

Microsoft AI says it beats Gemini, OpenAI and Whisper transcription on speed, accuracy and price.

  • New features include diarization, configurable transcription styles and word-level timestamps.
  • Microsoft says it handles a broader range of real-world audio than Gemini 3.5 Transcribe, GPT-Transcribe, Whisper V3-Large and ScribeV2.
Read on microsoft.ai →

Google Cloud previews Gemini for healthcare

TLDR AI

A Gemini agent built for providers and payers, with HIPAA-supporting governance.

  • The preview includes purpose-built healthcare skills and connectors to EHR and claims systems.
  • It targets healthcare providers and payers, with HIPAA-supporting governance.
Read on cloud.google.com →

Adobe previews MCP connectors that let outside AI agents edit Photoshop, Illustrator and Premiere files

The Rundown AI

Adobe's core creative apps will open up to users' own AI agents.

  • The MCP connector tools are coming in early 2027.
  • Adobe also launched Firefly Graph in beta, a builder for repeatable, shareable AI workflows.
Read on blog.adobe.com →

ExecuTorch WebGPU runs PyTorch models on the GPU in the browser

TLDR AI

PyTorch models can run in the browser without being converted to another framework's format.

  • ExecuTorch now runs PyTorch models in the web browser on the GPU via WebGPU.
  • Models don't need converting to another framework's format.
Read on pytorch.org →

Tether's BotPsy is a 0.8B open-weights agent model built for on-device tool use

Hugging Face Blog

A sub-billion-parameter model that beats a model 2.5x its size on most agentic benchmarks and runs on-device.

  • Built on Qwen3.5-0.8B and post-trained with SFT, multi-teacher on-policy distillation and two rounds of RL. It raises BFCL v4 from 23.4 to 44.0 and τ-bench from 23.2 to 51.5.
  • It beats Qwen3.5-2B-Think on five of six tool-use benchmarks, trailing only on Claw-Eval (46.5 vs 48.9).
  • It will be released under Apache 2.0 in bf16, Q8 and Q4, as safetensors and GGUF for llama.cpp.
Read on huggingface.co →

Google DeepMind's AlphaProof Code formally verifies security properties of Signal's protocol

TLDR AI

AI-driven formal verification now reaches real cryptographic implementations, not just models of them.

  • AlphaProof Code is a multi-agent AI framework for mathematical verification.
  • It verified key security properties of Signal's cryptographic protocol at the implementation level.
Read on deepmind.google →

Meta's e-means beats k-means in a tenth of the epochs

Hugging Face Daily Papers

Meta offers a drop-in replacement for k-means, which is used everywhere in retrieval, quantization and data curation.

  • It uses an exponential-moving-average, mini-batch optimization procedure, with scaling laws that set decay rate and batch size without hyperparameter sweeps.
  • It beats Lloyd k-means' 25-epoch loss in 1.4 to 2.6 epochs, for 10-45x GPU wall-clock speedups over state-of-the-art k-means implementations.
  • It partitioned 10 billion DINO embeddings into 16 million clusters in 2.5 epochs.
Read on huggingface.co →

Tongyi Lab's PhysOPSD uses executable physics code as privileged context to make generated video more physically plausible

Hugging Face Daily Papers

It turns simulation programs into dense feedback for post-training video models.

  • PhysOPSD-Engine translates real-world physical scenarios into verifiable simulation programs, which serve as privileged context for on-policy self-distillation.
  • With only 3K samples, it beats SFT and RL baselines across diverse backbones while preserving general video quality.
Read on huggingface.co →

Zeroing under 0.05% of an LLM's weights matches full fine-tuning in post-training

Hugging Face Daily Papers

Deletion Tuning gets fine-tuning-level results with tiny masks and much less forgetting.

  • The method learns a binary mask over transformer gate projections and matches full fine-tuning across RL with verifiable rewards, SFT and on-policy distillation.
  • The masks are hundreds to thousands of times smaller than full checkpoints.
  • Under RL, prior-task accuracy stays near base level, while full fine-tuning, LoRA and random sparse fine-tuning lose up to 5.4 points.
Read on huggingface.co →

Tree-Ising repairs LLM reasoning loops by editing bits in the model's weights, without gradients

Hugging Face Daily Papers

A cheap optimization approach beats a frontier agent's weight-space repair at a tiny fraction of the cost.

  • It splits the repair into independent QUBO (Ising) problems, one per output row, restricted to sign flips and exponent shifts in BF16 weights.
  • On Qwen3.5-0.8B it reaches 54.6% on MATH-500, versus 51.7% for the GPT-6 Astra agent's edit, using four forward-pass requests and about $0.90 of GPU time.
  • On Qwen3.5-2B it lifts greedy GSM8K accuracy from 21.6% to 69.4%.
Read on huggingface.co →

RoboScript lets robots learn manipulation tasks in context by rehearsing them in simulation

Hugging Face Daily Papers

In-context learning beats trained robot policies without any gradient updates.

  • A frozen model starts from one demonstration, rehearses the task in a simulator and writes the feedback into a Task Memory instead of its weights.
  • It reaches 83.4% mean success on all 50 RoboTwin 2.0 tasks, more than 20 points above the strongest trained action policy.
  • It transfers to a physical dual-arm robot, and the code is open source.
Read on huggingface.co →

Infrathon tests frontier agents on 24-hour AI infrastructure optimization tasks

Hugging Face Daily Papers

Long-horizon evaluation shows that short benchmarks can get model rankings wrong.

  • It covers six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours; 135 of the 150 runs kept improving in the second half of their budgets.
  • Claude Opus 5.5 raised the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%.
Read on huggingface.co →

Goodfire argues individual neurons are more interpretable than assumed

TLDR AI

Researchers may be able to study model computations without always relying on sparse autoencoders.

  • Goodfire's Neuron Scientist agents generate and test descriptions of what individual neurons do.
  • The work suggests neurons can serve directly as an interpretable basis for studying language models.
Read on goodfire.com →

Weight arithmetic between code models suppresses reward hacking without retraining

TLDR AI

Subtracting model weights can isolate and dial back a specific bad behavior.

  • Subtracting code models trained with loose and strict rewards identifies a direction that controls inefficient solutions.
  • Extrapolating beyond the strict model cut resource-limit failures from 14.5% to 8%, though too much extrapolation breaks outputs.
Read on x.com →

Embodied First Exam benchmarks coding agents controlling 47 robots

Hugging Face Daily Papers

Even the best coding agent succeeds on only about a third of realistic robot-control tasks.

  • It has 1,156 tasks in 35 suites on 47 robots, and agents can choose direct control, writing control code, or training a policy.
  • On 456 core tasks, the strongest of eight agents, GPT-6 Astra in Codex, reaches 34% in the Standard Setting.
Read on huggingface.co →
BenchmarkGemini 4 CarbonClaude Fable 5.2GPT-6.2 AstraMuse WonderKimi K3.1 Pro
GPQA diamond96.596.096.8—94.2
FrontierMath-Tiers-1-3-v2-Private90.595.192.682.178.2
FrontierMath-Tier-4-v2-Private95.1100.0100.070.765.9
OTIS Mock AIME 2024-2025100.099.7100.099.4100.0
SimpleQA Verified77.794.078.5—62.2
Chess Puzzles78.079.078.0—45.0
EBR-bench87.690.592.431.437.1
Mystery Game Puzzles73.090.095.051.042.0
ARC-AGI98.598.598.5——
SimpleBench92.694.979.2——
APEX-Agents84.280.172.266.165.5
ARC-AGI-294.293.895.8——
PostTrainBench41.044.6—24.2—
LMCA65.770.6—56.857.6
DTBench98.499.2—95.293.3
Furniture Assembly77.590.089.254.255.8
FrontierSWE61.269.9———

Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).