Top developments

Google made Gemini 4 Argon generally available at an introductory $2 per million input tokens and $10 per million output, with cached input 95% off. The unusual part is the end date: rates double on December 1. Per Artificial Analysis the launch pricing runs about 60% of GPT-6 Astra's per-task cost, and the deadline turns a discount into a seven-week window to evaluate before the real price arrives. Google pairs the pricing with a 1M-token output limit, a Fairwind Program expansion offering a version without cyber guardrails to vetted defenders, and free Antigravity access for 10,000 scientists: pricing and researcher goodwill, deployed together.

GreenWAM trains one world foundation model for both prediction and policy across 20 robot families, evaluated on WorldArena 1.0 with wins over pi0.5- and Cosmos-3-Edge-derived policies, then distills the result to a single-step real-time policy. World models that ship as deployable policies are what the field has been promising for years; this is the most credible attempt yet to close that gap.

AutoAgent0 is an agentic runtime supervision harness for safe closed-loop driving: a rule-based verifier plus a recovery-planning loop wrapped around existing experts, with no retraining required. It is evaluated on Fail2Drive and NavSafe-infinity. Supervision harnesses that leave the underlying policy untouched are the pragmatic route to deployment, and this is the driving-domain entry in that trend.

Frontier labs

Google DeepMind and Google Cloud will give up to 200 European researchers direct access to frontier science models (AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve, AlphaEarth Foundations) plus a shared $3 million in compute credits, announced at an AI summit in Dublin. The access runs through the EU's RAISE initiative, with the European Commission selecting the recipients. It is an access commitment, not a release, and it reads as a charm offensive aimed at European regulators and talent: a real offer, sized at roughly $15,000 per researcher if split evenly.

Meta's Muse Wonder and Wonder Full remain benchmark claims only: no technical paper yet, carried over from earlier coverage. Independent Artificial Analysis scoring puts Wonder Full at 56 on its Intelligence Index, two points behind Claude Opus 5.5 at 58, with strengths in agentic and long-horizon coding evaluations. The details: Wonder is live via the Meta Model API and Muse Code at $4/$20 per million tokens, while the stronger Wonder Full is limited to Muse Code subscribers with no API pricing announced. Treat Meta's own charts as vendor numbers until third-party evaluations land.

Quiet day otherwise: nothing new from Anthropic, OpenAI, xAI, Mistral, or Cohere in the window.

AV / embodied

LoopWAM does closed-loop policy improvement from mixed-quality deployment rollouts, lifting RoboChallenge success from 23.67% to 43.33%. Learning from your own messy deployment data instead of curated demos is the economically important version of imitation learning.

AXIS-Bench is a quality-aware VLA pretraining study finding that approach-aware episode conditioning beats plain behavior cloning, and that sim pretraining raised real-world success from 9.6% to 31.7%. Data quality conditioning mattering more than data quantity is a result worth internalizing.

LIBERO-RIG is a hardware-setup robustness benchmark: pooled VLA success collapses from 0.932 to 0.509 at maximum severity. The sim-to-real gap is really a lab-to-lab gap, and this quantifies it.

UniW(A)M unifies world and world-action models with a mixture of experts, reaching 88.3% real-world success against a 76.7% baseline; it is under review at ICLR 2027 and belongs to the same lineage as the UniWAM entry already in the Knowledge Vault from October 4.

WorldRecon does generative multi-view mesh reconstruction for large-scale scenes, with 14.9% and 43.1% Chamfer reductions. Reconstruction quality keeps compounding.

Tesla claims 300+ Cybercabs now serving Austin, an eightfold increase in 30 days, with Tesla's AI chief signaling Dallas is next. It is a scaling datapoint with no research content: the company says nothing about rides completed, vehicle utilization, or remote-assistance rates, so treat it as a deployment count, not an operating report.

TIER IV, Taisei, and Taisei Rotec were selected for Japan's MLIT FY2026 Autoflow Road demonstration: an eight-day, three-vehicle trial of multi-vehicle autonomous operations in dedicated logistics lanes, testing platooning, merging, obstacle avoidance, and sensor and roadside-infrastructure simplification. Worth a glance if you track the Japanese AV scene.

Research worth reading

GreenWAM (2610.18774): one world foundation model for prediction and policy across 20 robot families, distilled to a single-step real-time policy.

AutoAgent0 (2610.18866): agentic runtime supervision harness for safe closed-loop driving, no retraining required.

LoopWAM (2610.18873): closed-loop policy improvement from mixed-quality deployment rollouts.

AXIS-Bench (2610.18869): quality-aware VLA pretraining study; episode conditioning beats plain BC.

LIBERO-RIG (2610.18843): hardware-setup robustness benchmark for VLA models.

WorldRecon (2610.18846): generative multi-view mesh reconstruction for large-scale scenes.

ChunkAct (2610.18874): action-chunking methods in the embodied batch.

mmVB (2610.18858): visual tokens on demand for multimodal models.

Tev / Tev-Gym (2610.18868): temporal evaluation environments for embodied agents.

EgoMatrix (2610.18822): egocentric representation learning for embodied AI.

WorldOracle (2610.18826): world-model-based oracle methods.

UniW(A)M (2610.18697): unified mixture-of-experts for world and world-action models.

Plan-Bench (2610.18836): planning benchmark for embodied agents.

Game2Real-Bench (2610.18834): game-to-real transfer evaluation benchmark.