OpenAI Decisions API opens in public beta
OpenAI turned decision routing into a first-class API primitive for picking models, tools, or actions in near real time, with claims of up to 10x faster decisions than GPT-6 Luna via Responses.
Technical AI signals · Daily at 8 PM ET
Tuesday, October 6, 2026
Today split between OpenAI productizing fast fixed-choice routing with the Decisions API, Mistral pushing out a 1T-parameter Large 4 preview, and a dense layer of agent eval, control, and infrastructure work underneath.
OpenAI turned decision routing into a first-class API primitive for picking models, tools, or actions in near real time, with claims of up to 10x faster decisions than GPT-6 Luna via Responses.
Mistral put a trillion-parameter model with 49B active parameters behind its API now and says open weights are planned for late October, making it the clearest new frontier-scale model artifact of the day.
Three new papers push the agent-eval stack forward at once: OSWorld-Pro adds 300+ tasks with 2,800+ subgoals, UndoBench isolates recovery from task competence across enterprise workflows, and DeskForge generates dense desktop supervision from real apps.
Overmind packaged a concrete recipe for domain agents by building context graphs from code and traces, curating train/eval data, and reporting a 0.12% quote-invention rate on 4,000+ contract questions.
Google DeepMind released a 740M open model that maps text, images, video, and audio into one 768-dimensional space, giving operators a small self-hostable multimodal retrieval building block.
Paid tiers are being reduced from five to three, and the new Grow tier now requires $500 in total API payments instead of $1,000.
GLM-5.3 is available on Amazon Bedrock for enterprise coding and agentic use.
Applying the batch I/O rule to Codex made listing threads about 80% faster.
The specification lets programs report idle, working, waiting, finished, or failed states to the terminal.
The updated open-weights model is 27B parameters, tops the new Hugging Face Decision Index 0.3 benchmark, and costs half as much as v1 at $0.02 per million input tokens.
The interview covers few-shot learning, training robots across different embodiments, and the company’s claim that its latest model shows emergence not explicitly trained for.
H-JEPA is built on SIGReg/LeWM and uses a stable training recipe where semantic abstractions emerge across levels.
The platform lets agents discover, run, and pay for 2,500 APIs at runtime with pay-per-call pricing.
GitHub says it is rebuilding Git infrastructure while GitHub keeps running.
The guide addresses version compatibility across host kernels, GPU drivers, and other Kubernetes cluster components.
NVIDIA describes green contexts for running multiple independent GPU components in one process.
The post describes networking and data movement as GPU-controlled operations rather than host-driven services.
OpenAI and Ironclad are training and evaluating agents on complex contracting workflows.
OpenAI says an internal frontier model solved open problems in mathematics and shared Lean proof formalizations on GitHub.
ReviewBench evaluates code review agents on representative GitHub pull requests using multi-source ground truth, calibrated evaluation, and production-aligned metrics.
Goodfire found an internal signal that accompanies reward hacking and built probes that detect it in real time at scale.
Commenters speculate that a frontier model helped produce the accelerator design.
Commenters quote the model as a sparse Mixture-of-Experts system with 501 billion total parameters and 23 billion active parameters.
Commenters describe Polars as a fast replacement for pandas and praise its query-planner-style behavior.
Commenters are skeptical of zeroth-order training methods for neural networks.
Commenters say Claude discovered an algorithm that refutes the 3SUM, APSP, and Exact Triangle hypotheses.
Commenters discuss Gleam's move away from Erlang source output and into Erlang abstract form.
Commenters say recent GLM versions make it easier to reverse engineer closed protocols from binaries.
Kandinsky 6.0 Video comes in 3B and 29B versions that generate 5-second clips with synchronized 44 kHz audio and lip-sync.
SearchJev scores legal search decisions directly and uses SLCD to learn calibrated probabilities from uncertain supervision.
OpenRUA uses only terminal access to a robot's native software stack instead of bespoke harness layers.
RealtimeWAM uses one-step action generation and asynchronous inference to avoid multi-step denoising and sequential waiting.
LMBuild evaluates whether generated 3D structures are buildable and functional by representing them as assemblies with parts, joints, materials, and sequences.
ASCENT trains an agent's weights on its own execution trajectories during deployment.
The paper shows long action chunks fail around subskill transitions and proposes switching based on transition timing.
Amin Vahdat says failures happen multiple times an hour at 100,000-accelerator scale and argues goodput is the useful metric.
Sam Parsons describes three agent-driven payment flows and says tokenization keeps card details out of the agent and merchant PCI scope.
Charles Frye walks through request handling from server IO and tokenization to scheduling, model execution, and detokenization.
Qwen3.8-27B is a post-trained model that the official Qwen API plans to host with 1M context by default.
DeepSeek-V4.1-Flash is a multimodal MoE model with 552B backbone parameters and up to 1 million tokens of context.
Kolibri 1 is a 78B-parameter MoE reasoning model with explicit reasoning mode and tool calling.
Clef is a 27B multimodal model that returns probabilities for typed questions in a single forward pass.
Clef-Flash is a 9B multimodal decision model that returns probabilities instead of free-form text.
ZDTaichu5.0-9B supports text, images, and video and targets spatial reasoning and embodied-AI research.
The dataset contains agentic RL environments for code, cyber vulnerability reproduction, general knowledge work, visual web development, and symbolic music composition.
The dataset is tagged for agents, LLM agents, computer use, and AI safety.
The space is a research article with chapters, embeds, bibliography, and Docker-based deployment instructions.
REA is an MCP for reverse engineering binaries, applications, and runtime behavior.
DeepGEMM unifies FP8, FP4, BF16, fused MoE, and MQA kernels in one CUDA codebase.
The plugin generates STEP, GLB, STL, or 3MF models and can run design-for-manufacturing checks.
The framework records agent steps and replays them without model calls until the app changes.
Claude-Mem preserves context across sessions by capturing tool usage observations and generating semantic summaries.
The skill forces action-first, numbered outputs instead of long lead-ins.
Impeccable ships 24 commands and 60 deterministic detector rules for frontend design.
The project adds ten layout grammars, including Sankey, fishbone, Wardley map, kanban, and UML class.
Agency Agents now has a native app for macOS, Linux, and Windows that installs agents into Claude Code, Cursor, Codex, Gemini, and Osaurus.
openGym runs on self-hosted infrastructure and supports passkey login, synced devices, and offline use.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.