Models & Agents

Models & Agents GPT-5.6 ships with new multi-agent and programmatic tool-calling APIs plus major cost-per-task gains that directly affect what production agents can sustain today. What You Need to Know: OpenAI released GPT-5.6 with expanded API support for programmatic tool calling and multi-agent orchestration, alongside three new models and 18 pelican variants tied to reasoning levels. ChatGPT W...

Models & Agents OpenAI just made real-time voice agents available to every ChatGPT Plus and Pro user, with free rollout starting now. What You Need to Know: GPT-Live, the next-generation voice model, is now live across iOS, Android, and web for paid tiers, with free users coming in the next few days. OpenAI also audited SWE-Bench Pro and retracted its recommendation after finding 30% of tasks brok...

Models & Agents OpenAI's next frontier model arrives Thursday while a 2.7-trillion-parameter Chinese model readies for open release. What You Need to Know: Sam Altman confirmed GPT-5.6 Sol launches this week. MiniMax plans a 2.7T-parameter M3 Pro model for Q3 open-sourcing focused on complex reasoning. Simon Willison's Claude Fable 5 review of sqlite-utils 4.0 surfaced four additional blockers plu...

Models & Agents Real-time voice agents just got cheaper and more capable—OpenAI split its Realtime API into specialized models with lower latency. What You Need to Know: OpenAI released GPT-Realtime-2.1 and a mini reasoning variant optimized for voice, cutting p95 latency by at least 25% via better caching. Tencent open-sourced Hy3, a 295B MoE with 21B active parameters and 256K context that hits...

Models & Agents A 1.6-trillion-parameter open MoE model with native 1M context just launched, giving builders a new domestic-trained option for long-context work. What You Need to Know: Meituan released LongCat-2.0, a 1.6T-parameter Mixture-of-Experts model activating ~48B parameters per token with native 1M context via LongCat Sparse Attention, trained end-to-end on domestic AI ASIC superpods. .....

Models & Agents Coding agents just saved a production library release by catching five release-blocking bugs that the author missed, at an estimated $149 cost. What You Need to Know: Simon Willison used Claude Fable to complete the sqlite-utils 4.0 stable release, uncovering transaction-handling flaws and documentation gaps that would have broken SemVer guarantees. LlamaIndex shipped a public lega...

Models & Agents Specialized agent stacks are cutting construction document review cycles from 60 days to 10 by replacing general-purpose models with perception-semantics-agent layers trained on domain data. What You Need to Know: Trunk Tools released a three-layer architecture that extracts, relates, and acts on millions of pages of construction documents with 95% reported accuracy. ... AI Disclos...

Models & Agents Local browser agents just became practical and private — WebBrain runs entirely in Chrome or Firefox using your own models. What You Need to Know: WebBrain delivers an open-source, MIT-licensed browser agent that reads pages and automates multi-step tasks via Ask and Act modes. Simon Willison shipped a one-shot CLI coding agent built on his llm library. Safety research introduced P...

Models & Agents Chinese labs just shipped a free, open-weight coding agent that undercuts Western tools on price while removing export-control risk. What You Need to Know: Z.ai released ZCode, a desktop agentic IDE built around the newly open-sourced GLM-5.2 model, with cross-device remote control via WeChat and Feishu. SenseNova-U1-8B-MoT-Infographic-V2 arrived as an Apache-2.0 image model specia...

Models & Agents Export controls lifted on Claude Fable 5 and Mythos 5, letting Anthropic restore global access tomorrow with tighter cybersecurity classifiers. What You Need to Know: Anthropic is redeploying Claude Fable 5 globally after US government talks, routing some coding tasks back to Opus 4.8 while new classifiers block more misuse. OpenAI introduced GeneBench-Pro, a benchmark focused on a...

Models & Agents Meituan just open-sourced a 1.6T MoE coding agent that ran the top OpenRouter leaderboard for two months while training entirely on Chinese ASICs. What You Need to Know: LongCat-2.0 delivers 59.5 on SWE-bench Pro (above GPT-5.5) with a 1M context window and zero-cost cache hits under an MIT license. Builders now have a commercially usable near-frontier agent model they can run or f...

Models & Agents OpenAI's $20B Cerebras chip purchase has effectively removed high-throughput ASIC inference capacity from the market for everyone else. What You Need to Know: Cerebras' near-term inference supply is now pre-allocated to a single hyperscaler, pushing smaller teams off the waitlist indefinitely. New agent memory runtimes and research agents launched today show continued movement on t...

Models & Agents — Weekly Recap (Week of June 28, 2026) Pull up a chair, this is Models and Agents, episode 95, for June 28, 2026. Let's see what happened in the AI world today. ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube: https://www.youtube.com/watch?v=6sdadppZU58 📝 Full show notes, transcript & sources: re...

Models & Agents OpenAI’s GPT-5.6 family launches in limited preview, giving builders three new tiers for balancing capability, speed, and cost under tighter government oversight. What You Need to Know: OpenAI released the GPT-5.6 series (Sol flagship, Terra balanced, Luna efficient) with a robust safety stack and 750 tokens/sec inference planned for July, but only to a small set of vetted partners...

Models & Agents OpenAI's internal rollout shows agents handling complex, cross-functional work at scale, giving builders an early view of what production agent systems will soon need to match. What You Need to Know: OpenAI reports Codex agents are already doing longer-running, cross-team tasks across every department. YesWeHack launched autonomous agents for on-demand penetration testing. ... AI D...

Models & Agents GPT-5.5 Instant now handles intent and constraints more reliably while rolling out to all users this week. What You Need to Know: OpenAI released an updated GPT-5.5 Instant that improves intent understanding, complex constraint handling, and recommendation quality. OpenAI also announced its first custom AI chip, Jalapeño, built with Broadcom for ChatGPT and agent workloads. ... AI...

Models & Agents Claude now joins teams as a persistent, org-wide Slack entity—the third major shift in LLM interaction after websites and desktop apps. What You Need to Know: DFlash delivers up to 15x throughput on NVIDIA Blackwell by drafting whole token blocks in parallel instead of autoregressive token-by-token generation. Qwen released two new AgentWorld MoE models trained to simulate tool, te...

Models & Agents OpenAI just shipped a full cyber defense stack—GPT-5.5-Cyber plus Codex Security and Patch the Planet—so defenders can now scan, validate, and patch at machine speed inside existing workflows. What You Need to Know: OpenAI expanded Daybreak with a dedicated GPT-5.5-Cyber model, Codex Security plugin for Codex, and the Patch the Planet program for open-source maintainers. Alibaba re...

Models & Agents Open-weight GLM-5.2 is drawing Silicon Valley attention as Chinese labs close the capability gap with a new publicly available model. What You Need to Know: GLM-5.2 arrives as the latest open-weight release from a Chinese startup, with early reports highlighting strong performance that has caught the eye of Western labs. ... AI Disclosure: This podcast is curated by Patrick but use...

Models & Agents — Weekly Recap (Week of June 21, 2026) Welcome back to Models and Agents, episode 88, for June 21, 2026. There's signal and noise in AI every day. ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 📝 Full show notes, transcript & sources: read the episode page 🌐 Part of the Nerra Network — explore every show at nerran...

Models & Agents Virtuals just wired Leyten’s distributed GPU engine into its agent network to run GLM-5.2 at scale. What You Need to Know: Virtuals’ integration lets agents tap GLM-5.2 across a distributed GPU fabric. A major security report details active exploits against Langflow and related frameworks. ... AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis f...

Models & Agents Custom CUDA kernels now keep vector search inside the GPU for agentic RAG, cutting PCIe round-trips that silently throttle long-horizon agents. What You Need to Know: OpenAI reports GPT-5.5 Instant now matches its frontier models on health queries for free users. A new GPU-resident Top-K kernel delivers deterministic microsecond tail latencies for retrieval. ... AI Disclosure: This...

Models & Agents NVIDIA just released an open frontier model built from the ground up for long-running agents. What You Need to Know: NVIDIA released Nemotron 3 Ultra, an open model explicitly designed for persistent agent workloads. DeepSeek previewed V4 series MoE models that support million-token contexts with major efficiency gains. ... AI Disclosure: This podcast is curated by Patrick but uses...

Models & Agents OpenAI’s LifeSciBench brings 750 expert-authored tasks from real biotech and pharma workflows into AI evaluation. What You Need to Know: OpenAI released LifeSciBench, a benchmark spanning seven biological research workflows developed with 173 scientists. GPT-Rosalind outperforms GPT-5.5 across all workflows, while GPT-5.4 drove a full medicinal chemistry project from literature to...

Models & Agents OpenAI’s new deployment simulation technique replays real user requests against unreleased models to surface undesired behaviors before launch. What You Need to Know: OpenAI released research on Deployment Simulation that estimates real-world failure rates by running candidate models on de-identified past conversations. Anthropic published fresh data showing Claude Code task value...