Google launches Gemini Decisions API and Antigravity browser control as Runway open-sources Praxis-1

Google Antigravity rolled out built-in Browser Use across the Antigravity desktop App and Antigravity CLI (v2.30.0+ when signed in with a personal Google Account), documented on 11 November 2026 a day after Google Antigravity’s /swarm multi-agent preview. Developers can ask in plain language or invoke the /browser slash command to test a running localhost server, click through responsive viewports, fill forms, inspect console errors and failed network requests, and verify visual UI fixes before an agent completes its turn.

Side-panel actuation, parallel tabs and sandbox rules

Rather than requiring a separate external Chrome setup in the desktop App, the main coding agent delegates live UI work to an isolated browser subagent running inside the embedded side-panel webview. Each browser subagent runs in its own context window on a Gemini Flash model by default so full-page DOM snapshots and screenshots do not consume main-thread context or higher-tier model quota; if Flash quota runs out, the agent pauses and asks the user before switching models.

While acting, the embedded browser displays an animated step-status cursor without stealing keyboard focus from the editor or hijacking a foreground tab the developer is reading. Multiple browser subagents can also operate concurrently in dedicated parallel tabs for cross-environment audits, governed by configurable Browser Actuation Rules for allowed and denied URLs, download paths, and manual sign-in handoffs, while CLI sessions execute browser tasks through Chrome DevTools.

Google DeepMind released AlphaGenome 2 on 11 November 2026 through its updated alphagenome Python package (alphagenome.models.v2) and free non-commercial research API, alongside the zero-code and programmatic AlphaGenome Atlas covering pre-computed in silico mutagenesis and AlphaGenome Variant Impact (AVI) prioritisation scores across 9 billion human coding and non-coding variants. Whereas AlphaGenome 1 scored variants across fixed reference assays simultaneously, AlphaGenome 2 zooms into specific cellular contexts to model why identical single-letter DNA mutations trigger disease in certain cell types, such as heart muscle or immune cells, while sparing others.

Single-cell expression conditioning and imputed modalities

AlphaGenome 2 pairs a 1-megabase (1 Mb) conditional U-Net and transformer sequence backbone with a dedicated cell expression encoder (encode_cells). Researchers can select composite ontology reference conditions or pass their own raw human single-cell RNA-seq (scRNA-seq) unique molecular identifier count matrices (up to 256 cells per API call via cell_conditioning) to generalise regulatory predictions to previously unseen cell states.

The v2 release adds predicted single-cell chromatin accessibility (SC_ATAC), denoised and protocol-specific single-cell gene expression, CpG methylation variant scoring (MethylationScorer), cross-modality imputation for unmeasured biosample tracks, and batched input-gradient calls (predict_interval_gradients_batch) for enhancer-to-gene attribution across human and mouse genomes. Google DeepMind has not announced commercial API pricing tiers.

SpaceXSI introduced Grok Bot 1.0 on 11 November 2026 via its official X channels alongside Elon Musk and the dedicated @bot product handle, packaging Grok into persistent, shareable workplace teammates rather than single-user chat threads. Instead of restricting prompts and custom setups to one individual’s private session, teams can configure task-specific Bots and share them directly across colleagues inside an organisation.

Workspace channels, live meetings and adaptive memory

According to the launch posts, Grok Bot 1.0 operates natively inside Slack channels and Microsoft Teams workspaces and can be invited directly into live team meetings as an active participant. Across recurring threads and calls, shared Bots retain institutional context and learn team conventions over time so coworkers do not have to restate project background on each assignment.

SpaceXSI framed the 1.0 release around delegating multi-step workplace tasks to a shared colleague inside existing enterprise communication tools. SpaceXSI has not published a standalone technical documentation page, enterprise seat or token pricing, seat limits, model version details, or compliance and data-retention controls alongside the X launch posts.

OpenAI rolled out Finances inside the ChatGPT desktop application on 11 November 2026, a day after launching Businesses in ChatGPT. Announced via the official @ChatGPT account, the dedicated desktop workspace is available immediately after updating the desktop app across Free, Go, Plus, and Pro subscription tiers in the United States.

Spending breakdowns, portfolio tracking and custom tools

Finances gives individuals a structured workspace inside the desktop client to review recurring expenses, identify opportunities to save, and examine personal investment holdings in one view rather than uploading ad-hoc CSV exports into standard chat threads. Users can also build custom interactive financial tools tailored to their own budgets, savings targets, or portfolio allocation questions directly within the app.

Making the workspace available down to the Free and Go tiers broadens access beyond paid Plus and Pro subscribers from day one in the US market. OpenAI did not publish a companion technical blog post detailing supported brokerage or bank aggregation providers, supported desktop operating systems, encryption boundaries for connected financial records, or rollout dates outside the United States.

Fish Audio introduced Live 3 on 10 November 2026 across the Fish Agent Platform, its developer API, and enterprise deployments, launching an English full-duplex voice model built to listen continuously while speaking. Instead of stitching separate voice-activity thresholds and rule-based interruption timers onto a cascaded speech-to-text and text-to-speech pipeline, Fish Audio trained conversational timing directly into the speech model on recordings of human dialogues while keeping a separate reasoning layer for agent instructions, knowledge bases, and external tool calls.

Learned backchannels, mid-order revisions and barge-in

Live 3 distinguishes brief listener backchannels from genuine interruptions. When a caller says “yeah” or “mm-hmm” while an agent reads back store hours or an order summary, Live 3 continues speaking without cutting off its own turn; when the caller interjects to change days or items, the model stops immediately and yields the floor.

During chunked inputs such as phone numbers or multi-item food orders with mid-sentence self-corrections, Live 3 can decide to emit a short confirmation in as little as 186 milliseconds and wait through pauses until the caller finishes deciding, drawing on Fish Audio’s library of more than 2 million voices and custom clones. Fish Audio has not published standalone per-minute API pricing or multilingual release dates for Live 3.

Runway open-sourced Praxis-1 under the Apache 2.0 licence on 11 November 2026, releasing base and embodiment-tuned checkpoints on Hugging Face alongside reference training and inference code on GitHub (runwayml/praxis-1-oss), a day after Google DeepMind unveiled Gemini Robotics 2.5. Built to overcome the scarcity of teleoperated robot trajectories by transferring physical intuition from general human video pretraining, Praxis-1 pairs a 27.2B-parameter Mixture-of-Experts chunk-causal video-diffusion trunk (54.4 GB in BF16, using frozen Wan2.2 VAE and Qwen3.5-2B encoders) with a 0.3B-parameter action expert mid-trained across 18 robot embodiments.

Joint video-action chunking and RoboLab-120 results

On each control call, Praxis-1 ingests a language instruction, recent camera frames, and current proprioceptive pose to jointly predict a future video chunk and 1.6 seconds of motor commands (24 action steps at 15 Hz), adapting across single-arm Franka DROID rigs, bimanual YAM arms, Google RT-1 SimplerEnv, and low-cost SO-101 hardware via six lightweight per-rig projection tensors and NVFP4 quantised checkpoints. Runway’s benchmark charts on the research page are interactive web components rather than static image files: fine-tuned on DROID, Praxis-1 leads all world action models and open-weight vision-language-action policies on NVIDIA’s 120-task RoboLab-120 simulation suite, ahead of FLUX 3 Action, pi 0.5, and GR00T N1.6, trailing only zero-shot frontier VLM harnesses AMP1 (86.0%) and SenseNova-RoboRSI (57.1%).

Key numbers

• 46.2% success rate (555 of 1,200 episodes) on NVIDIA RoboLab-120, ahead of FLUX 3 Action at 42.9%
• 27.2B MoE video-diffusion trunk plus 0.3B action expert (54.4 GB BF16 across 17 safetensors shards, plus NVFP4 variants)
• 24 action steps at 15 Hz (1.6 seconds of motion) predicted per chunk alongside future scene video

Baidu released ERNIE 5.5 Flash Preview on 11 November 2026 as the first post-trained model built on ERNIE 5.5 Flash Base. Available immediately via API on Baidu AI Cloud’s Qianfan platform, the Mixture-of-Experts language model pairs 280 billion total parameters with 14 billion active parameters per token across a 1-million-token context window. Baidu stated that it plans to release the model weights under the Apache 2.0 licence at a later date, without naming a release date or publishing international USD token pricing.

Autonomous Unreal Engine 5.8 C++ and procedural Blender CAD builds

Trained specifically for multi-hour agent harnesses across game engines and parametric 3D software, ERNIE 5.5 Flash Preview completed an 11-hour autonomous coding run that merged multi-generation assets and wrote an entire playable third-person wuxia game (Yunxiu Xing) in C++ inside Unreal Engine 5.8. Because the model is text-only and cannot view rendered frames directly, it translated playtest glitches such as floating characters and camera offsets into automated code assertions, and separately generated nearly 100 playable 2D and 3D mobile mini-games from design documents using deterministic state-space verification.

In 3D modelling demonstrations, coordinated ERNIE 5.5 Flash Preview agents wrote Python scripts in Blender 5.2 (bpy) and FreeCAD to assemble a parametric airport and dual-deck airliner rendered with Cycles and OptiX, reconstruct the Horyu-ji temple complex, block out a 2,100-object, 1.82-million-triangle previsualisation scene, and model articulated W16 engine and lunar rover assemblies.

At ElevenLabs Summit NYC on 11 November 2026, ElevenLabs opened free early access to ElevenAI, a workplace assistant for employees built on its 10 November 2026 Scribe v3 speech-to-text release, while previewing Interaction Engine and rolling out enterprise upgrades across its customer-facing ElevenAgents platform. ElevenAI provides cross-app voice dictation, consent-aware live and in-person meeting notes with per-attendee action items, real-time call guidance, and context handoffs between human staff and automated agents.

Interaction Engine, Memory, Routines and NVIDIA sovereign scale

Interaction Engine coordinates conversational turn-taking and dynamic model routing inside ElevenAgents, handling routine triage on lower-cost models and escalating complex steps such as billing refunds to frontier models. Alongside it, Memory distils returning authenticated callers’ preferences and passing remarks into persistent structured records alongside a company’s CRM, while Routines automates pre- and post-call workflows such as mortgage pre-screening callbacks, waitlist SMS outreach, and CRM updates.

On the infrastructure side, ElevenLabs integrated NVIDIA Nemotron 3.5 Super VL into ElevenAgents, reporting leading customer-support task completion at 70% lower cost than rival open models and top-three platform speed when reasoning is toggled off. ElevenLabs is also adding 15 MW of on-premises NVIDIA Vera Rubin compute alongside its Blackwell Ultra systems, deploying a sovereign landline scheduling agent for Poland’s national e-Health platform, and launching a 550-partner ElevenLabs Partner Program (including a McKinsey and QuantumBlack alliance) plus the ElevenLabs Marketplace.

Perplexity published Escaping SPACE: Part II on 11 November 2026, expanding red-teaming of its open-source Sandboxed Platform for Agentic Code Execution (SPACE), which isolates Perplexity Computer agents inside root-privileged guest virtual machines. Adding Claude Mythos 5.1, Claude Opus 5.5, and GPT-6 Astra to the nine models tested in Part I (bringing the total to 12 models), researchers evaluated two boundaries: escaping hypervisor isolation to steal a host AWS credential, and bypassing egress network policies (after patching DNS-spoofing flaws discovered in Part I and disclosing similar bugs to 8 of 10 third-party sandbox vendors).

Boundary-focused prompting vs CTF flags and cyber-model trade-offs

Under standard Capture the Flag (CTF) prompts, nine non-GPT models spent at most 4% of their actions probing the hypervisor (and GPT-6 Astra, GPT-5.6 Cyber, and GPT-5.6 Sol spent 18% to 22%), instead hunting for exposed files or intercepting the guest spaced-svc vsock daemon. Rewriting prompts to explicitly target virtualisation vulnerabilities shifted effort toward hypervisor probes by roughly 60 percentage points on average and pushed Claude and GPT runs near the 120-minute ceiling (up from 21.5 average minutes for Opus 5.5), while SPACE’s VM-host isolation still held across every run.

Perplexity also found that general-purpose frontier models outperformed cybersecurity-specialised derivatives on broad sandbox tasks: base GPT-5.6 Sol succeeded in 3 of 3 source-assisted Part I network-bypass runs where cyber-tuned GPT-5.6 Cyber scored 0 of 3, and general-purpose GPT-6 Astra similarly outperformed cybersecurity-focused Claude Mythos 5.1.

OpenRouter launched its Extractions API in alpha (/api/alpha/extractions, published on 10 November 2026 and announced broadly on 11 November 2026), introducing a dedicated endpoint that pulls structured, typed JSON values directly from unstructured text, images, and PDFs without prompt-engineered JSON wrappers. At launch, the API routes across three specialised extraction providers: Cloudflare (including cloudflare/clairic, demonstrated extracting six invoice fields for $0.00011), Inception, and ScaleDown.

Nine schema field types, Unicode spans and calibrated null floors

Developers pass a document in state (or image and PDF references in assets) alongside a map of named fields across nine built-in types: text, number (with configurable decimals and units), date, time, datetime, list, duration, estimate, and multi-step forecast. Rather than returning raw strings alone, single-value extractions return per-field confidence scores, spell_confidence on OCR text read from images, Unicode code-point character spans ([start, end)) pinpointing the exact source substring on supported text models, and p_null, the calibrated probability that a requested field is absent from the document.

When p_null crosses 0.5 (or a custom per-field null_threshold), the API returns an explicit null alongside a best-effort guess value so downstream pipelines can enforce deterministic confidence floors and route only uncertain fields to human review, while estimate and forecast fields express uncertainty through mean, median, p10, and p90 distributions.