This observation covers 7:49 am on 10 November to 7:49 am on 11 November 2026 in Australia/Sydney time and includes six technical changes verified against primary and independent material.
Gemini 4 Carbon moves to paid API and product access
Google initially made Gemini 4 Carbon available only to selected cybersecurity defenders on 28 October. The new fact on 10 November is access through the paid Google AI Studio API and enterprise services, with a staged rollout to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Introductory pricing is US$2 per million input tokens and US$10 per million output tokens, scheduled to rise to US$4 and US$20 from 1 December. Artificial Analysis has independently run its composite evaluation, recording a one-million-token context window, an index score of 60 and long task times. Google’s other capability and cost comparisons remain specific to their respective evaluation settings.
Claude Fable 5.2 gains third-party evaluation after release
Anthropic released Claude Fable 5.2 on 9 November and restricts Claude Mythos 5.2, which uses the same underlying model, to organisations verified for cybersecurity or life-sciences work. Both have a one-million-token context window, a 128K maximum output and always-on adaptive reasoning. Fable is available through the Anthropic API, AWS, Google Cloud and Microsoft Foundry at US$8 per million input tokens and US$40 per million output tokens. By this run on 11 November, Artificial Analysis had completed an independent evaluation, scoring the maximum-effort setting at 65 while recording a cost of US$26.54 per task and a long time to first token. Anthropic’s claims about speed, safety and research results remain developer-reported.
LongCat-2.5 progresses from hosted service to MIT-licensed weights
This series previously recorded LongCat-2.5 chat and API access on 2 November, when weights had not been delivered. On 10 November, Meituan LongCat published the main model repository, MIT licence, model card and 208 safetensors shards on Hugging Face. The model card describes a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active parameters, supporting thinking and non-thinking modes. No Hugging Face inference provider serves it. Coding, agent, visual and quantisation results are developer evaluations without a complete independent reproduction.
MiniMax opens H3-Image generation and editing weights
MiniMax released H3-Image on 11 November with a unified generation and editing checkpoint. The model card lists a 33.1-billion-parameter single-stream diffusion transformer, with a BF16 transformer of about 68.65 GB and a further 66.7 GB text encoder reused from MiniMax-H3. It supports text-to-image generation, up to 16 reference images and output of about 16 megapixels without an upscaler. The weights use the MiniMax H3-Image Community License rather than Apache 2.0. Claims of sub-10-second 4K generation and comparative speed, quality and pricing are developer-reported and have not been independently reproduced.
Alpamayo 2 Edge weights and inference code become downloadable
NVIDIA released Alpamayo 2 Edge on 10 November as a two-billion-parameter in-vehicle vision-language-action model. Its GitHub repository provides Apache 2.0 inference code, while the Hugging Face weights use OpenMDW-1.1. The pinned checkpoint needs about 5.3 GB of storage; examples require Linux and an NVIDIA GPU with BF16 support, with at least 16 GB of GPU memory recommended. The model consumes six-camera observations, vehicle history and navigation instructions and produces causal reasoning plus 64 future poses. DRIVE AGX Thor latency and closed-loop reinforcement-learning gains are NVIDIA measurements without an independent in-vehicle deployment reproduction. No hosted API is available.
MKA multimodal knowledge-agent evaluation goes live
Google DeepMind and Artificial Analysis released the Multimodal Knowledge Agent evaluation on 10 November. It tests agents that process drawings, engineering diagrams, clinical scans, Office files and spreadsheets and produce reports, slide decks or spreadsheets. Artificial Analysis independently runs 21 models with the same Stirrup tools, sandbox and a limit of 250 turns. The full 200-task set spans 10 industry verticals but remains private; Google released only 10 example tasks on Hugging Face. The published methodology and leaderboard expose operating conditions and results, while the private full set prevents complete external reproduction.
Sources
Google, Gemini 4 Carbon is now available, 10 November 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
Artificial Analysis, Gemini 4 Carbon (high) Intelligence, Performance & Price Analysis, no publication date shown; accessed 11 November 2026, https://artificialanalysis.ai/models/gemini-4-carbon
Anthropic, Claude Fable 5.2 and Mythos 5.2, 9 November 2026, https://www.anthropic.com/claude-fable-and-mythos-5-2
Anthropic, Claude Platform release notes, 9 November 2026, https://platform.claude.com/docs/en/release-notes/overview
Artificial Analysis, Claude Fable 5.2 (max) Intelligence, Performance & Price Analysis, no publication date shown; accessed 11 November 2026, https://artificialanalysis.ai/models/claude-fable-5-2
Meituan LongCat, LongCat-2.5, 2 November 2026, https://longcat.ai/blog/longcat-2.5
Meituan LongCat, LongCat-2.5 model repository, updated 10 November 2026, https://huggingface.co/meituan-longcat/LongCat-2.5
MiniMax, MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era, 11 November 2026, https://www.minimax.io/blog/minimax-h3-image
MiniMax, MiniMax-H3-Image model repository, no publication date shown; accessed 11 November 2026, https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
NVIDIA Technical Blog, Running AV Reasoning Models Within Onboard Compute and Latency Limits, 10 November 2026, https://developer.nvidia.com/blog/running-av-reasoning-models-within-onboard-compute-and-latency-limits/
NVIDIA, NVlabs/alpamayo2-edge, updated 10 November 2026, https://github.com/NVlabs/alpamayo2-edge
NVIDIA, Alpamayo2-Edge model repository, no publication date shown; accessed 11 November 2026, https://huggingface.co/nvidia/Alpamayo2-Edge
Google DeepMind, Introducing the Multimodal Knowledge Agent benchmark, 10 November 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/
Artificial Analysis, Multimodal Knowledge Agent (MKA) Benchmark, no publication date shown; accessed 11 November 2026, https://artificialanalysis.ai/evaluations/mka
Google, Multimodal Knowledge Agent public tasks, updated 10 November 2026, https://huggingface.co/datasets/google/multimodal-knowledge-agent
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.