Top developments
SemiAnalysis's GDPx benchmark puts a hard number on the agentic gap: the best frontier model scores just 12 percent. GDPx, published November 2, tests models across 11 high-value domains (chip design, biology, finance, legal, quant trading, recursive self-improvement, hardware, materials, alignment, software engineering, cybersecurity) chosen for where frontier AI is expected to earn its first trillion dollars. Claude Opus 5.5 leads at 12.0 percent, Gemini 4 Carbon is second at 10.9 percent, GPT-6 Astra third at 9.6 percent, and finance and legal return 0 percent across all 13 tested models. The low absolute numbers are the story: even the best models are far from automating high-value knowledge work, and the domain-by-domain breakdown (Carbon leads biology and software engineering, Astra leads recursive self-improvement and materials) is more useful than the aggregate ranking. GDPx coverage
Autoware 2.0 is out, bringing end-to-end AI into the leading open-source driving stack. The Autoware Foundation's new major version adds support for data-centric, end-to-end AI stacks and hybrid AI/rule-based configurations alongside the established rule-based modules, plus a Reference Design framework (deployment profile plus reference software stack) meant to make commercial adoption concrete, and foundation models published on Hugging Face. TIER IV is spearheading the ecosystem push with Arm, AWS, and Red Hat. This is genuinely useful infrastructure for open AV research: the stack the community actually runs now speaks the same end-to-end language as the frontier labs. TIER IV announcement
Weekend leaks point to OpenAI's GPT-6.2 Astra and Anthropic's Claude Mythos 5.5 / Fable 5.5 landing imminently, but none of this is confirmed. GPT-6.2 Astra is reportedly in gray-scale testing in the Codex backend (handling about 20 percent of traffic) after OpenAI skipped the planned 6.1, and Anthropic's website source code was found referencing Claude Mythos 5.5 (a security-focused edition for vetted institutions) and Fable 5.5 (the general flagship, reportedly with 1M-token context and already appearing in cloud catalogs). Treat all of this as unverified leak, not announcement, though the rumored release window of November 3-4, days after Gemini 4 Carbon topped the charts, fits the pattern of this three-way race. Leak evidence check
Frontier labs
Gemini 4 Carbon's chart-topping comes with an asterisk: almost nobody can use it yet. Carbon scores 59 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57), with standout cyber numbers (88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym), but access is limited to vetted cyber defenders in Google's Fairwind Program, with no date given for wider availability. A one-point lead is close to a tie, and a model nobody can use cannot win customers. The real test is how long the lead holds and when it opens up. Gemini 4 Carbon coverage
AV / embodied research
Tesla's Texas robotaxi fleet hit 1,204 registered vehicles on October 31, 55 behind Waymo's 1,259. The latest batch added 124 Model Ys, flipping the mix to 631 Model Ys and 573 Cybercabs; Avride (344) and Zoox (63) trail far behind. Registry counts are not cars carrying riders, but the trajectory is the story: Tesla is closing a gap that looked structural months ago, alongside its Robotaxi hours extension to 2 a.m. Texas fleet tracker report
MicroVision expanded its collaboration with FieldAI across government and industrial autonomous systems (November 1), following successful tests of its lidar on FieldAI ground vehicles and drones. A small vendor-collaboration note, not a technical result, but it shows the defense and industrial autonomy pipeline continuing to consolidate around proven perception hardware. MicroVision announcement
Research worth reading
ObjectGuidedMap: object-aware BEV map prediction under heavy occlusion. A multitask architecture built on Mask2Map that jointly predicts BEV maps and 3D objects, with cross-task fusion at both the BEV-feature and instance-query levels, so map elements can borrow context from the traffic participants occluding them. Gains are largest exactly where it matters: plus 21 percent mAP in heavily occluded scenes on nuScenes and Argoverse 2. arXiv
Mechanistic Interpretability for End-to-End Self-Driving (CoRL 2026). Sparse autoencoders trained on the activations of a single-camera end-to-end planner recover interpretable structure aligned with route intent (98.3 percent unsupervised recovery), speed, and braking; paired counterfactual edits show the features are visually grounded (traffic lights, stop signs); steering individual latents produces predictable, monotonic planner changes; and an adapter on SAE latents cuts trajectory selection error by 3.44 percent. The rare interpretability paper that also ships a measurable planning improvement. arXiv
GATS: geometry-measured token selection for efficient sparse latent world models. Vision-centric driving models encode thousands of image tokens per frame, most of them sky and empty road; GATS uses LiDAR/radar point projections to select a small fraction of tokens before the vision backbone runs, cutting end-to-end FLOPs up to 18x. At a 25 percent token budget it beats the dense baseline on nuScenes (0.520 m vs 0.573 m 3-sec L2) and NAVSIM (88.9 vs 85.9 EPDMS) at 6.6x fewer FLOPs, and on 739 intent-critical scenarios it covers 95.8 percent of hazardous vehicles with a 0.09 percent collision rate (vs 0.34 percent dense). Directly relevant to fitting capable driving models onto vehicle compute. arXiv
Other notable
A viral video of a Waymo stuck sideways exiting an Atlanta parking lot (November 1), with remote support reportedly unable to get the car to accept its suggestion, is a reminder that odd edge cases and the remote-assist loop remain a real part of the driverless experience, even as fleets scale. Waymo incident coverage
ECARX and Arm signed a non-binding MoU (November 2) to co-develop high-performance AI compute platforms for mobility and physical AI. Early-stage partnership paper; watch for actual silicon, not the MoU. ECARX announcement