Gemini batches up to 1,000 decisions as Runway opens robotics weights
Google’s new API can evaluate up to 1,000 questions in one multimodal request, and Runway is releasing the weights for its Praxis-1 robotics model. This issue also covers Baidu’s preview release, shared planning, live speech translation, and how teams coordinate and evaluate agents.
01
Models and APIs
4 stories
2026-11-11Google DeepMind
Gemini Decisions API evaluates up to 1,000 questions per request
Google introduced the Decisions API, powered by Gemini Decision 1. A request can combine text, images, audio, video and PDFs in a shared input of up to 256K tokens, then ask up to 1,000 independent questions about it. The API handles yes/no decisions, multiple choice and rubric scores, returning probabilities or confidence scores.
Google measured 150 to 250 milliseconds of end-to-end latency in tests with contexts up to 16K tokens. The shared input is billed once per request. Standard pricing is $0.10 per million input tokens, with no output charge. A free tier in Google AI Studio supports prototyping.
2026-11-11Baidu ERNIE
Baidu brings ERNIE 5.5 Flash Preview to Qianfan
Baidu made ERNIE 5.5 Flash Preview available through its Qianfan API. It is the first post-trained version of Flash Base, whose pretrained weights were already released. The model has 280 billion total parameters, 14 billion active parameters and a one-million-token context window. Baidu plans to release the preview weights by November 25.
In an 11-hour demonstration, the model worked inside an agent framework to build an Unreal Engine 5.8 project in C++. It turned problems such as floating characters and misplaced cameras into checks it could run in code. The demo used external scene and character assets.
2026-11-11Runway
Runway releases weights for its Praxis-1 robotics model
Runway released the Praxis-1 weights, inference code and fine-tuning code under Apache 2.0. The 27-billion-parameter model predicts both what a robot will see next and what action it should take. The release supports full fine-tuning and LoRA adaptation.
Runway reports a 46.2% success rate across 1,200 episodes on NVIDIA’s RoboLab-120 benchmark. FLUX 3 Action reached 42.9% in the same table. The figures measure task success in that benchmark.
2026-11-10AWS
Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock
AWS added Moonshot AI’s Kimi K3.1 Pro Highspeed to Amazon Bedrock. It supports image input, tool calls, structured outputs and streaming. Developers already calling Kimi K3 through Bedrock can keep the same endpoint and API and switch to the new model ID.
Applications can select a service tier on each request. AWS says Flex, an option for background summaries and evaluations, costs half as much as Standard and can take longer to respond. Bedrock caches repeated prefixes automatically and lets developers explicitly mark the stable part of a prompt.
02
Products and developer tools
5 stories
2026-11-11Google Labs
Google Labs introduces Tabletop for shared planning
Google Labs introduced Tabletop, a real-time collaboration experiment. Users can add links, files and ideas, then use a prompt or starter template to build an interactive itinerary, dinner poll, expense splitter or tournament bracket.
Once the space is ready, the organizer can share an invitation link in a group chat so members can use the tools on the same page at the same time. Tabletop is rolling out via a waitlist for people aged 18 and older in the United States.
2026-11-11Adobe
Adobe previews edits across shots and photos you can explore
Adobe showed ten research prototypes at MAX Sneaks. Project Scene Sweep tracks recurring people and locations across a video sequence so creators can apply a change consistently as the shots and camera angles change.
Project Sight Span turns a single image into a scene users can move through to explore compositions beyond the original frame. Adobe is making selected prototypes available as early research previews in its new Labs Playground.
2026-11-11DeepL
DeepL Aure follows changes in tone during live speech translation
DeepL introduced Aure for real-time speech translation. Its earlier voice preservation technology sampled a speaker at the start of the conversation. Aure continuously uses the source audio and translated words to adjust the delivery, aiming to carry changes in pace, emphasis and emotion into the translated speech.
Aure applies automatically in DeepL Voice for online meetings, in-person conversations and the Voice API. It initially supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. Other languages continue to use the existing voice models.
2026-11-11GitHub
GitHub opens issue types and custom fields to personal repositories
GitHub made issue types and fields generally available for repositories owned by personal accounts. Individual developers and open-source maintainers can sort feedback using types such as Bug, Feature and Task or define their own.
A personal account can create up to 25 custom issue fields. They support text, numbers, dates, single selection and multiple selection. Fields can be pinned to particular types or made available on issues without a type.
2026-11-11AWS
SageMaker adds usage-based software charges for bidirectional streaming
AWS added usage metering for SageMaker AI bidirectional streaming inference. Transcription or voice model providers selling through AWS Marketplace can report consumption during a live session and set software charges based on a unit such as a second of audio processed.
The provider defines the unit. If a product meters audio by the second, 12 minutes of metered audio is 720 units. The length of time the connection stays open alone does not determine the billable audio. Provisioned SageMaker endpoint infrastructure is billed separately.
03
Research and benchmarks
4 stories
2026-11-11Anthropic
Anthropic compares one coordinator with per-user agents
Anthropic compared ways for agents to coordinate across four simulated settings: a shared token budget, clinic scheduling, group bookings and a code merge queue. With Claude Opus 5 and Sonnet 5, one agent coordinating for the whole group produced better group outcomes in all four settings than separate agents coordinating on behalf of each user.
The shared-budget test used 25 scenarios. Relative to the best possible outcome, the Opus 5 team of individual agents achieved 30%, and the single coordinator achieved 64%. Anthropic released three of the four environments as MAMUBench.
2026-11-11AWS
AWS introduces UAEF for cross-framework agent evaluation
AWS introduced its Universal Agent Evaluation Framework. Adapters convert messages, tool calls and agent handoffs from frameworks including LangGraph, LangChain and Strands Agents into one trace format. Teams can score those traces using a selection of 36 built-in metrics across seven dimensions.
Teams can evaluate an existing trace or run an agent against a dataset. The framework records the model, prompt and tools used in each run so results can be compared with a baseline. AWS says 12 organizations in telecom, healthcare and financial services used it to evaluate agents.
2026-11-11OpenRouter
OpenRouter measures the input cost of three tool interfaces
OpenRouter asked Claude Sonnet 5.5 for the latest release and the three newest open bug issues in one GitHub repository. All three approaches answered correctly. The CLI used 1,421 input tokens, the skill used 3,315, and the MCP setup with its default tools used 59,417.
Definitions for the 46 default MCP tools took about 19,436 tokens per request. The full release object added a long list of download metadata. A follow-up with fewer tools and trimmed results used 5,357 input tokens for the task and answered correctly in every run.
2026-11-11OpenRouter
OpenRouter compares three precisions across six endpoints for one model
OpenRouter compared fp16, fp8 and fp4 across six endpoints serving Qwen3.8 27B. All six scored between 98.3% and 99.2% on 240 tasks with exact answers, including arithmetic, tracing code, sorting and finding information in long inputs.
On 40 separate multiplication problems with no working shown, one fp4 endpoint got 26 right. The other five got 30 to 33, including 30 on the other fp4 endpoint. OpenRouter says the sample was too small to attribute that gap to precision. Running the full set, including the 40 problems, cost $0.58 to $1.24 per endpoint.