Anthropic ships Claude Fable 5.2 and Mythos 5.2 while Meta launches Muse Ember at $0.10

Meta introduced Muse Ember on 9 November 2026 as the compact, low-latency member of its proprietary Muse model family. Accepting native text, image, audio and video inputs across a 1M-token context window, the model targets multi-step coding agents, visual perception pipelines and voice interfaces. Developers can call Muse Ember through the Meta Model API, Muse Code and Google Cloud Vertex AI, while consumer rollouts extend the model to Meta AI on Apple Watch and WearOS smartwatches.

Efficiency and evaluation discrepancies

On the Meta Model API, Muse Ember costs $0.10 per million uncached input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. According to Meta and Artificial Analysis, Muse Ember matches Anthropic’s Claude Haiku 5.5 at 36 on the Artificial Analysis Intelligence Index while finishing the benchmark suite at an average cost of roughly $0.14 ($0.136) per task against $0.411 for Haiku 5.5. Meta’s benchmark charts on the research blog are rendered interactively in HTML rather than as downloadable images. One figure differs across sources: Meta’s research table lists 37.8% on AutomationBench-AA, whereas Artificial Analysis’s summary post reports 68.1% on the same benchmark.

Key numbers

• $0.10 input, $0.004 cached input, and $0.50 output per million tokens across a 1M-token window
• 1608 Arena Elo on Code Arena WebDev, ranking first among small models
• 61.6% on Agents’ Last Exam and 40.9% on Terminal-Bench 4.0
• 80.6% on PerceptionBench with tool use and 46 on Coding Agent Index-AA

Google Cloud announced two eighth-generation Tensor Processing Unit systems on 9 November 2026, TPU 8tx and TPU 8i Max, bringing its eighth-generation accelerator lineup to four distinct configurations alongside TPU 8t and TPU 8i. TPU 8tx pairs TPU 8t silicon with fifth-generation AMD EPYC host processors over Google’s Boardfly optical interconnect to deliver hardware-encrypted confidential computing for regulated inference and mid-scale training. TPU 8i Max scales up the inference-focused 8i board with 2 times the compute, 2 times the on-chip SRAM and 2 times the HBM bandwidth to serve million-token context windows and dense mixture-of-experts workloads.

Agent Compute Engine and TPU Sync

Alongside the boards, Google introduced software layers aimed at stateful agentic traffic, where models pause while external code executes or tools return data. Agent Compute Engine suspends and resumes idle agent sessions across TPU pods, while open-source TPU Sync pools HBM, host DRAM, local storage and pod networking into one unified key-value (KV) cache tier. Google reports that TPU Sync cuts KV-cache transfer latency and time to first token by more than 90% on multi-turn agent workloads. Cloud Pathways and dynamic pod slicing also reached general availability on Google Kubernetes Engine (GKE), delivering up to 5 times faster cluster startup and 4.5 times faster fault recovery. Google did not publish hourly cloud instance rates in the announcement.

NVIDIA unveiled GR00T 2 and Isaac Lab 3.0 at the Conference on Robot Learning (CoRL) in Austin on 9 November 2026. Rather than mapping camera frames directly to joint torques like standard vision-language-action models, GR00T 2 is a world-action model built on NVIDIA Cosmos: a single diffusion transformer jointly denoises predicted future video frames and multi-step robot action chunks. Predicting how objects and hands move in pixel space lets the policy transfer physical intuition from human videos with zero robot action labels.

Training scale and onboard latency

NVIDIA pretrained GR00T 2 across 402,730 hours of egocentric human video, 10,184 hours of handheld Universal Manipulation Interface (UMI) gripper trajectories spanning 2,897 tasks, and 23,500 hours of robot teleoperation. Alongside the paper, NVIDIA released Human-1M-100K, an open 100,000-hour subset of the egocentric corpus on Hugging Face, with model code scheduled for December 2026. On real-world bimanual and humanoid benchmarks, GR00T 2 completes 40 of 50 tasks zero-shot, reaches 71% success from a single robot demonstration, and hits 95% after five demonstrations and 30 minutes of post-training on one RTX 5090 GPU. A distilled variant, GR00T 2 Turbo, executes onboard a Unitree G1 humanoid’s Jetson Thor computer at 124 ms per prediction.

ByteDance Seed introduced Seedream 5.5 on 9 November 2026, upgrading its flagship image generation and editing model for commercial design pipelines. Standard diffusion editors gradually drift in colour, facial identity and background geometry after several conversational revisions; Seedream 5.5 constrains latent updates to edited regions so unmasked pixels remain stable across up to 100 sequential editing rounds. The model also outputs native four-channel RGBA images with clean alpha transparency around hair, glass and UI stickers without requiring a separate background-removal pass.

Realism and commercial evaluations

To reduce synthetic skin sheen and garbled signage, ByteDance Seed retrained the model’s layout and texture stages for dense Chinese and English typography as well as natural lens grain. On Good-Same-Bad (GSB) human preference evaluations published by the team, Seedream 5.5 outperforms GPT Image 2.5 Sunburst, Google’s Nano Banana 2.1 and Seedream 5.0 Pro on text-to-image realism and bilingual text structure, with only image-to-image realism against Nano Banana 2.1 (-6.18%) favouring the rival model. Seedream 5.5 is rolling out across Jimeng, Doubao, Doubao Work and enterprise API platform Volcano Ark; ByteDance Seed did not disclose international per-image API pricing in the blog post.

Epic Games released the experimental Unreal Engine 5.9 MegaJam Preview on 9 November 2026 ahead of the Epic AI MegaJam running from 12 to 15 November 2026. According to Epic Games and CEO Tim Sweeney, a small internal engineering team built the preview branch in roughly one week by directing swarms of autonomous coding agents across the engine’s C++ and Python codebase rather than writing every integration manually.

Agent harnesses, Canvas and browser runtime

The preview embeds an experimental Unreal Assistant panel that connects external coding harnesses, including Claude Code, Codex, Cursor and OpenCode, directly to an auto-launching Unreal Model Context Protocol (MCP) server. Epic also expanded Unreal’s Python scripting APIs across MetaSounds audio graphs, Niagara particle systems, Procedural Content Generation (PCG), physics constraints and UMG user interfaces so agents can inspect and mutate whole levels programmatically without custom C++ glue. An integrated Unreal Canvas workspace generates concept art and converts 2D prompts into textured 3D meshes inside the editor, while a revived WebAssembly and WebGPU export target packages playable Unreal projects to run directly in desktop browsers. The preview is free to download through the Epic Games Launcher, GitHub source access and Linux archives, though Epic warns that experimental APIs may change before stable 5.9.

Skild AI released Steel (Scalable Throughput-optimized Engine for Embodied Learning) on 9 November 2026 as a GPU rigid-body dynamics solver inside Newton 1.8 Release Candidate, the open-source physics engine co-developed by NVIDIA, Google DeepMind and Disney Research under the Linux Foundation. Traditional GPU simulators run fastest when thousands of parallel environments contain identical robots and identical object counts, but stall under warp divergence when training generalist policies across mixed arms, hands, quadrupeds and humanoids.

Heterogeneous batching and throughput

Steel eliminates that bottleneck by grouping articulations with matching joint counts across disparate scenes and dynamically routing narrow-phase collision contacts per environment. Skild AI also coupled the rigid-body solver with Material Point Method (MPM) sand, finite-element cloth and deformable cables, alongside an interactive steel-webgpu solver that runs directly in web browsers. Skild AI’s 20-task benchmark chart is rendered in HTML on its blog rather than as a static image: across 20 heterogeneous robotics environments, Steel reaches 3.8 times the geometric-mean simulation throughput of NVIDIA PhysX, compared with 2.9 times for MuJoCo Warp.

Zyphra released ZONOS3-Mini and its companion ZAC 1 neural audio codec under the Apache 2.0 licence on 9 November 2026. The text-to-speech and zero-shot voice-cloning model uses a sparse mixture-of-experts (MoE) backbone with 800M total parameters and 100M active per token, pairing Gated DeltaNet linear-recurrence layers with standard attention in a 3:1 ratio. Zyphra pretrained the model on 112 H100 GPUs in 2.5 days before applying Direct Preference Optimisation (DPO) and Group Relative Policy Optimisation (GRPO) to reduce word skips and unstable prosody.

Low-frame-rate codec and edge runtimes

Instead of generating dense 50 Hz audio tokens, ZONOS3-Mini predicts tokens from ZAC 1, a 44.1 kHz codec operating at 15 Hz across one w2v-BERT 2.0 semantic codebook and nine residual vector-quantisation (RVQ) acoustic codebooks, conditioned on Zyphra’s MECT-B2 speaker encoder. On one H100 80GB GPU across 55 subsets of Zyphra’s ZTTS2-Eval suite, ZONOS3-Mini achieves an 80 ms median time to first audio at eight concurrent requests and peaks at 457 seconds of generated audio per second, advancing the quality-versus-latency Pareto frontier against Qwen3-TTS, Higgs Audio v3 and Fish S2 Pro. Reference inference implementations ship for SGLang-Omni, CUDA, Apple Silicon MLX, C++/ggml and in-browser WebGPU.

Automattic, the company behind WordPress.com and WooCommerce, launched Spacefast on 9 November 2026 to turn prototypes generated inside AI chats and coding agents into hosted full-stack web applications. Founded inside Automattic by former Beeper engineer Batuhan and backed by infrastructure serving more than 15 billion monthly pageviews, Spacefast connects through Model Context Protocol (MCP) plugins or a single terminal publish command to ChatGPT, Claude, Cursor, Meta’s Muse, Grok Bot and OpenCode.

Managed backend and plan tiers

Unlike static preview sandboxes that vanish when a chat session closes, each Space provisions a persistent public or access-controlled URL, a managed MySQL database, file upload storage, environment secrets, automatic version history with one-click rollbacks, and pinned on-page visual comments that feed revision instructions straight back to the user’s coding agent. Spacefast offers a free Personal plan with $1 of monthly compute credit, up to 20 Spaces and 1 GB of disk storage; a Go tier at $5 a month ($5 usage credit, 100 Spaces, 10 GB storage); a Pro tier at $15 a month ($15 usage credit, 500 Spaces, 25 GB storage and up to five team members included before $15 per extra seat); and custom Enterprise contracts.

Expo open-sourced Exact on GitHub under the MIT licence on 9 November 2026, introducing an experimental cross-platform UI stack created by co-founder Charlie Cheever specifically for human-and-agent co-authoring. Instead of mixing imperative component trees with styling workarounds, Exact separates presentation into a constrained declarative language called Contract while keeping state and data logic in TypeScript or Rust.

Compiler pipeline and native painters

An ahead-of-time compiler lowers each Contract specification into a compact execution plan. On the web, that plan runs through a roughly 20 KB DOM runtime with zero virtual-DOM diffing overhead. On desktop and mobile targets, a shared Rust layout kernel computes geometry and hands draw commands directly to native platform painters across macOS AppKit, iOS UIKit and GPU-rendered Linux windows, with no embedded webview and no asynchronous JavaScript bridge. Because every UI node maps deterministically to its Contract source, Exact ships with a headless agent driver that lets coding agents build screens, inspect live layout bounds, trigger user actions and assert visual states programmatically. Windows and Android painters are listed on the repository roadmap rather than complete in the initial preview.