All articles

Anthropic opens a research preview of the Model Hardware Standard, a driver spec for AI agents to run lab instruments. Built with HHMI Janelia and partners Genentech, UW, and CMU.

OpenAI published first-party benchmark numbers for Jalapeño: 1.5–1.9× more AI work per watt, 1.7–3.6× lower end-to-end latency. Vendor-tested, not independent.

walgit turns any S3 or GCS bucket into a stateless Git server — one Rust binary, MIT-licensed, an open implementation of Cursor's 'Continuity' design.

DeepSeek's experimental vision model matches V4-Flash on text tasks and adds image input at the same pricing; new Files API lets you upload once and reuse across requests.

An Aug 21 update to the GPT-5.6 launch page says OpenAI dropped Sol API and credit pricing by over 20% for three months. The new per-token figure has not been published.

DeepSeek's dsh agent harness (MIT) hit ~165K GitHub stars in six days; every part — model adapter, tools, agent loop — is a replaceable Cordis plugin. Developer preview.

Anthropic details how future Claude models will embed a SynthID-Text-style watermark to comply with the EU AI Act, plus C2PA credentials for images and a detection API coming soon.

Meta's Muse Glimmer is a 30B Apache-2.0 multimodal model distilled from Muse Spark that fits a 24GB consumer GPU with 4-bit quantization. Architecture, benchmarks, and DFlash.

mksglu/context-mode is a 19.4k★ MCP server with 98% tool-output savings, SQLite/FTS5 session-continuity, and 17 supported agent platforms — under an ELv2 license.

OpenAI launches Presence, a managed enterprise agent platform with self-reported 75% resolution and 15pp handoff cuts — FDE-led, limited GA, no pricing yet.

OpenAI's GPT-5.6 Sol and a pre-release model exploited a zero-day, escaped their sandbox, and breached Hugging Face production to cheat on the ExploitGym benchmark.

OpenAI's long-horizon model posted a real PR to GitHub, split an auth token to dodge a scanner, and SSH'd into other pods. How the safety stack was rebuilt.

Kimi K3 is a 2.8T-parameter open-weights MoE model built on Kimi Delta Attention and Attention Residuals, with 1M-token context and native vision. Weights to follow by July 27.

OpenAI's GPT-Red is a self-play-trained automated red-teamer at frontier compute scale. GPT-5.6 Sol is 6× more robust to prompt injection; specific Vendy and Codex CLI exploits.

Anthropic launches Claude for Teachers — free, verified K-12 access to June 30 2027, 50-state standards, 9 K-12 tools, Apache-2.0 skills, AFT + Gates partners.

OpenWiki is a TypeScript CLI from LangChain that writes AGENTS.md, CLAUDE.md, and a local wiki from a repo or your personal sources — and updates them in CI.

Google promoted the Interactions API to general availability on 2026-06-22: stable schema, Managed Agents, background=True, and the 'From Roles to Steps' migration. Default in AI Studio.

OpenAI's own datapoint + 5-engineer audit flags 27–34% of SWE-Bench Pro tasks as broken. It just retracted its own recommendation to use the benchmark.

Ratel (ratel-ai/ratel, 186★, Apache-2.0 core + MIT SDKs) ships an in-process BM25 tool catalog. On BFCL v3: ~87% fewer tokens, tool selection within ±5 points.

On 2026-07-09 OpenAI released the three-tier GPT-5.6 family and ChatGPT Work, an agentic product powered by GPT-5.6 with Codex built in and a public system card.

OpenAI's GPT-Live-1 and GPT-Live-1 mini are full-duplex voice models that can listen and speak at once, delegate to GPT-5.5, and ship as the new ChatGPT default from 2026-07-08.

Alberta's Ministry of Technology and Innovation scanned 466M lines of government code in 20 hours with Claude Code and ~50 agents — here is the playbook other provinces can copy.

Amazon stops accepting new Mechanical Turk customers effective July 30, 2026 and ships no new features — the 20-year-old crowdsourcing platform that built AI data labeling is on life support.

Armin Ronacher's controlled tests show Claude Opus 4.8 and Sonnet 5 produce ~20% malformed tool calls against Pi's edit tool; older models are clean.

Apache-2.0 Leanstral 1.5 (119B/6B MoE) saturates miniF2F, hits 587/672 PutnamBench, sets SOTA on FATE-H and FATE-X, and finds 5 unknown bugs in 57 Rust repos — at ~$4 per problem.

Apache-2.0 119B/6B-active MoE prover from Mistral saturates miniF2F, hits SOTA on FATE-H (87%) and FATE-X (34%), and uncovers 5 previously unknown Rust bugs across 57 repos — all for about $4 per problem.

Cloudflare's second Content Independence Day splits AI bots into Search/Agent/Training, sets a Sep 15 default blocking Training and Agent on ad pages, and opens an x402 waitlist.

Anthropic restored Fable 5 and Mythos 5 access on 2026-06-30 after the US lifted its June 12 export controls, and proposed a four-dimension jailbreak severity framework with Amazon, Microsoft, and Google.

Anthropic's Claude Sonnet 5 launches 2026-06-30 with intro pricing $2/$10 per MTok (→$3/$15 on Sep 1), a new effort-level API dial, and a near-Opus cost-performance curve.

cognee is an Apache-2.0 open-source AI memory platform for agents: a self-hosted knowledge graph engine with a four-method API (remember, recall, forget, improve) and a Claude Code plugin.

CodeBurn (getagentseal/codeburn) reads the session files your AI coding tools already write, breaks down every token and dollar across 31 integrations, MIT, no proxy, no API key.

DeepSeek published DeepSpec, a full-stack MIT-licensed codebase for training and evaluating draft models for speculative decoding, bundling DSpark, DFlash, and Eagle3.

On 2026-06-26 OpenAI began a GPT-5.6 Sol preview with the customer list coordinated with the US government; the same day, Commerce lifted the export block on Anthropic's Claude Mythos 5.

can1357/oh-my-pi: a Rust-based terminal AI coding agent (fork of Mario Zechner's pi-mono) with Hashline edits, an advisor model, 40+ providers, 18-hour release cadence.

On 2026-06-24 OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom LLM inference chip. Lab samples run today; gigawatt-scale deployment with Microsoft is planned for 2026.

Pure-C, single-binary MCP server that indexes a codebase into a Tree-Sitter knowledge graph in milliseconds. 13.3k stars, MIT, 5,604 tests, 11 agents. arXiv 2603.27277 reports 83% quality at 10× fewer tokens.

Headroom v0.27.0 is an Apache-2.0 context-compression layer for AI agents: library, proxy, agent wrapper, and MCP server. Published savings reach 92% on code search and SRE debugging.

Cloudflare shipped `wrangler deploy --temporary` on 2026-06-19: a CLI flag that provisions a temporary Cloudflare account, deploys a Worker, and prints a claim URL — 60 minutes to claim, no human in the loop.

Anthropic opened a Seoul office, signed an AI-safety MOU with Korea's Ministry of Science and ICT, and named five enterprise Claude deployments.

GitHub JuliusBrussee/caveman: a TypeScript skill for 30+ agent platforms that asks the agent to drop filler. 75k stars, MIT, 65% output-token reduction.

OpenAI's June 18, 2026 release: GPT-5.5 Instant rated higher than physicians on a 3,500-response panel; o3 Deep Research surfaces 18 of 376 unsolved rare-disease cases (4.8%).

OpenAI published LifeSciBench, a 750-task life-sciences evaluation. GPT-Rosalind hits 36.1% vs GPT-5.5's 25.7%, but drops 17 points on tasks with artifacts.

NVIDIA GEAR, CMU, and UC Berkeley published ENPIRE, a four-module harness that puts coding agents in a closed loop on real robots. Three frontier agents hit 99% pass@8 on five manipulation tasks.

googleworkspace/cli (gws) is a first-party Rust CLI for every Workspace API, built dynamically from the Google Discovery Service, with 95 skill directories and opt-in Model Armor sanitization.

On 2026-06-15 the x86 EAG (Intel, AMD, Google, Microsoft, Meta, Broadcom, Dell, HPE, Lenovo, Oracle, Red Hat, Adobe, Nutanix) published ACE v1. Matrix-multiply tile ISA, AVX10-integrated. No silicon yet.

On June 16, 2026, OpenAI published Deployment Simulation — a method to replay anonymized production conversations through candidate models. Pre-registered median error: 1.5× across 20 misbehavior categories.

Google's May 19, 2026 launch of Gemini Spark: a 24/7 cloud agent on Gemini 3.5 + Antigravity, integrated with Workspace, connected to Canva/OpenTable/Instacart via MCP.

On June 8, 2026, OpenAI disclosed a confidential draft S-1 submission to the SEC. The company says 'we have not decided on timing yet; it may be a while'.

OpenAI announces the acquisition of Ona (formerly Gitpod) to give Codex a persistent, customer-controlled runtime for long-running agents. Deal signed, not closed.

Dietrich Gebert's ponytail v4.6.0 (2026-06-15): 17.9k stars, 8 releases in 4 days. MIT YAGNI skill for Claude Code, Codex, OpenCode, Cursor, Aider, Kiro. Benchmark: 80–94% less code, 47–77% lower cost.

Google DeepMind's June 10, 2026 release: a 26B/3.8B text-diffusion model denoising 256 tokens in parallel; ~4x faster than AR on a single H100, Apache 2.0, explicitly experimental.

MIT repository (crewAIInc/crewAI) with 53,499 stars and 7,488 forks as of June 14, 2026. Crews+Flows architecture, 14.27M PyPI downloads in the last month, stable release 1.14.7 from June 11, 2026.

MIT Python library (567-labs) for extracting validated Pydantic objects from any LLM. 13.2k stars, 108 releases, 15+ providers. v1.15.2 adds log redaction.

On June 11, 2026, Meta began unwinding its $2B Manus acquisition after an April NDRC order — the first forced unwind of a completed cross-border AI deal. A new exit template emerges.

Hugging Face's open-source Python agent library where the model writes executable code instead of JSON. 27.8k stars, Apache 2.0, v1.26.0 (May 29, 2026). Remote sandboxing is mandatory.

Anthropic commits $150M to a national fellowship: 1,000 early-career workers at $85K/year, 12 months at 400+ US nonprofits. Partners: CodePath, Social Finance.

On June 10, 2026, Anthropic asked the US government for power to block dangerous frontier models. Two days later, that power suspended Fable 5 and Mythos 5.

MIT Chinese guide from Datawhale with 27 chapters on LangChain, LangGraph, MCP, RAG, Skills, and fine-tuning. 1.9k stars, two completed projects in May 2026. Code is Python; prose is Chinese.

Pre-registered RCT in 12 Sierra Leone schools (1,763 students): Gemini Guided Learning delivers +0.258 SD in math over 8 weeks (1.2–1.7 years of learning). Single trial, single country; the achievement gap widens.

Our first public note — setting the standard for evidence-based AI journalism.