Articles

103 articles

All notes, analyses, and explainers published so far.

Language models, benchmarks, releases, evaluations, and comparisons.

Developer tools, coding agents, AI apps, and practical workflows.

Companies, investment, infrastructure, hardware, and market moves.

Evergreen guides that make technical artificial intelligence topics accessible.

Tuesday, September 22, 2026

40 % lower cost to run that Anthropic claims for Claude Opus 5.5 over Opus 5 on typical workloads

Anthropic and OpenAI Launch Cheaper Near-Flagship Models on the Same Day

Models news Sep 22 4 min anthropic.com

Wednesday, September 16, 2026

5 tools actions the Home MCP server exposes to a connected agent

Google Lets Outside AI Agents Control Google Home Devices for the First Time

Tools news Sep 16 3 min support.google.com

Monday, September 14, 2026

Framework 3.0 China's new AI safety governance standard, published 14 September 2026

China's New AI Safety Framework Softens the Loss-of-Control Warning Reuters Cited the Same Day

Industry news Sep 14 6 min tc260.org.cn

PPLX 27B the local model that actually ships on Windows RTX PCs

Perplexity Brings Portable Computer to Windows RTX PCs

Tools news Sep 14 3 min perplexity.ai

83.75 ALE score before exploring, already ahead of GPT-6 Astra

RSIAgent's Harness Alone Adds Twice the Points Its Self-Improvement Loop Does

Models news Sep 14 4 min arxiv.org

Saturday, September 12, 2026

38 pages Microsoft's draft Code of Conduct for its MAI models

Amodei Asks AI Labs to Pace the Frontier, Not Stop; Microsoft Drafts a Separate Code

Industry news Sep 12 6 min darioamodei.com

Friday, September 11, 2026

31.47 best agent's Steering score, expert baseline scored 57.75

AI Agents Can Spot a Model's Hidden Concept Like an Expert. They Still Can't Prove It Controls Anything

Industry news Sep 11 5 min arxiv.org

Thursday, September 10, 2026

151 million exchanges in the campaign Anthropic attributes to Alibaba

Anthropic's misuse report: sophisticated attacks no longer need sophisticated attackers

Industry news Sep 10 7 min anthropic.com

0.05 $ / min per minute of GPT-Live-1 in the API

OpenAI Splits the Voice from the Brain: GPT-Live-1 Reaches the API at $0.05 a Minute

Models news Sep 10 3 min openai.com

1.92 $ per 20-minute session of the 64 GB container

OpenAI rents out the Codex harness: the Agents API and who controls it

Tools news Sep 10 5 min openai.com

92.7 % checkpoints where a blind agent reproduced the original result

Compression Is the Test: How AI Research Agents Prove They Did Not Overfit

Explainers news Sep 10 4 min amazon.science

Wednesday, September 9, 2026

82% of runs as Mythos 5 with at least one severely harmful action

Anthropic re-reads four Claude incidents: from operational failure to misalignment in six weeks

Industry analysis Sep 9 5 min anthropic.com

SB 813 and AB 1405: they regulate auditors, not developers

California's new AI laws regulate auditors, not AI developers

Industry news Sep 9 5 min gov.ca.gov

Tuesday, September 8, 2026

50% lower latency than Images 2.0 at most, per OpenAI

ChatGPT Images 2.5 splits the image API into Flare and Sunburst — at the same price

Models analysis Sep 8 5 min openai.com

6.26 points avg. gain over public Toucan data under matched training

NeoHorse-1: An Open Model Trained on Its Own Production Router's Logs

Models news Sep 8 3 min arxiv.org

Navier–Stokes the proof OpenAI claims, and who has checked it

OpenAI claims a proof for the Navier–Stokes problem. What it covers, and who has checked it

Models analysis Sep 8 4 min openai.com

AA26-251A the CISA, NSA and FBI advisory naming six China-based companies

Three US agencies accuse six Chinese AI firms of industrial-scale distillation, in an advisory rather than an indictment

Industry analysis Sep 8 3 min cisa.gov

Sunday, September 6, 2026

59.2% less GPU for Astra-class training, in a single week

OpenAI publishes its own numbers on research acceleration, and a warning from its chief scientist

Industry news Sep 6 5 min openai.com

Thursday, September 3, 2026

24% of the swarm became whistleblowers, unprompted

DeepMind's Agents Cheated at Math. A Quarter of the Swarm Tried to Stop Them and Failed

Industry news Sep 3 5 min arxiv.org

272,000 tokens of input: past it, the whole request costs 2x on input and 1.5x on output

GPT-6 Astra ships as a product: what builders get on the API, Azure and Bedrock

Models analysis Sep 3 4 min openai.com

eight times cheaper on one task to resume with recall than a written handoff, per its author

funes, from Hugging Face: a memory for coding agents that lives in a dataset you own

Tools analysis Sep 3 4 min huggingface.co

12,930 M USD Hugging Face's signed price, per NVIDIA's 8-K

NVIDIA buys the layer where open weights travel: what is signed, what is pending, what is only promised

Industry analysis Sep 3 5 min sec.gov

0.02 K margin of WeatherNext 3 over WeatherNext 2 on the 7-day 850 hPa temperature error

WeatherNext 3 learns from live satellites, and leads a public leaderboard by a narrow margin

Models news Sep 3 4 min blog.google

Wednesday, September 2, 2026

47.2% pass@1 for 3.8 Flash Cyber on CWE-Bench, computed by Collinear; Claude Fable 5 scores 47.8%

Gemini 3.8 Flash Cyber: the same model as Flash, with looser safeguards and an application form

Models analysis Sep 2 4 min blog.google

12.5 times less per million input tokens for the tier that lets Meta use your traffic

Muse Spark 1.3: what the cheap tier costs, and how Meta measured the benchmarks

Models news Sep 2 4 min research.meta.ai

Tuesday, September 1, 2026

99.1% of responses physicians rated safe, according to OpenAI's own evaluation

ChatGPT for Healthcare starts reading the patient chart

Industry news Sep 1 4 min openai.com

75% cheaper cache reads, now $0.25 per million tokens

Fable 5.1 and Mythos 5.1: one model, two levels of access, and cheaper cache reads

Models analysis Sep 1 4 min anthropic.com

2.57x speedup over ONNX Runtime Web in Hugging Face's own benchmark, on 809 of 1,756 cases

Hugging Face releases 207 WebGPU kernels to speed up AI in the browser

Tools news Sep 1 2 min huggingface.co

100% on ExploitBench for Astra, per OpenAI, measured with privileged Daybreak Blue access

OpenAI’s Astra Crosses Critical Cybersecurity Threshold — and Triggers Its Own Guardrails

Models analysis Sep 1 6 min openai.com

Wednesday, August 26, 2026

106% year-over-year revenue growth at NVIDIA, to $96.2 billion

NVIDIA's summer: revenue doubles, and most of the rest is a promise

Industry analysis Aug 26 3 min sec.gov

1,200 agents or so used the unsanctioned message board; about 700 took part in the Hugging Face attack

An Outside Team Investigated OpenAI’s Rogue Agents — and Published What It Could Not See

Industry analysis Aug 26 9 min metr.org

Friday, August 14, 2026

Article 50 of the EU AI Act: the marking duty Claude's text watermark meets

How Claude's text watermark works, and where it stops working

Explainers explainer Aug 14 3 min anthropic.com

Wednesday, August 12, 2026

2.8T parameters in Kimi K3, one of two open models of over two trillion released in 16 days

Kimi K3 and Qwen3.8: two open-weight models above two trillion parameters, and the licenses that come with them

Models comparison Aug 12 3 min huggingface.co

Sunday, August 2, 2026

€15 million or 3% of worldwide turnover: the maximum fine for breaching Article 50

What the AI Act demands from August 2: enforcement begins and four transparency duties apply

Industry explainer Aug 2 3 min digital-strategy.ec.europa.eu

Thursday, July 30, 2026

36% success rate screwing in a light bulb with a dexterous hand, versus 92% unscrewing it

Gemini Robotics 2, read through the success rates Google published beside the videos

Models analysis Jul 30 3 min deepmind.google

Tuesday, July 28, 2026

2026-07-28 is the MCP revision that drops the initialize handshake and protocol-level sessions

MCP goes stateless: what a team running MCP servers has to change after the July 28 revision

Tools explainer Jul 28 3 min modelcontextprotocol.io

Friday, July 24, 2026

$5 per million input tokens ($25 output), the same as Opus 4.8 and half the Fable tier

Claude Opus 5 keeps the Opus price and sells itself as near-Fable at half the cost

Models analysis Jul 24 3 min anthropic.com

Thursday, July 9, 2026

80% cut OpenAI made to Luna's API price on July 30

GPT-5.6 is a family priced by effort: Sol, Terra and Luna, and two cuts in six weeks

Models comparison Jul 9 3 min openai.com

Thursday, July 2, 2026

99% plus block rate that Anthropic claims for the new classifier on the target behavior

Monday, June 22, 2026

$80 billion Alphabet raised to extend its compute infrastructure

The AI Capital Machine: Alphabet Injects $80B, Anthropic Prepares for IPO, and No One Wants to Be Left Out

Industry analysis Jun 22 4 min linkedin.com

50% mark that bot traffic passed in June 2026, ahead of humans for the first time

More Bots Than Humans: Automated Traffic Now Dominates the Internet and the Web Rewrites Itself

Industry analysis Jun 22 4 min roirevolution.com

one billion users reached by ChatGPT in June 2026, faster than any app before it

ChatGPT Reaches 1 Billion Users: The Milestone and What It Really Means

Industry news Jun 22 3 min medium.com

15 announcements at Google I/O 2026, three of which set the strategic direction

Google I/O 2026: Chrome Becomes an Agent and the Web Prepares for Autonomous Browsing

Tools analysis Jun 22 3 min developer.chrome.com

4 million or more developers use Codex every week; its inference now runs through Bedrock

OpenAI Lands on AWS: GPT-5.5 and Codex Arrive on Bedrock, Reshaping the Enterprise Map

Industry analysis Jun 22 4 min aws.amazon.com

Friday, June 19, 2026

$20/month is the entry price of Cursor, Codex and Claude Code; Mimo Code and OpenCode are free

AI Coding Tools, June 2026: Cursor, Codex, Claude Code, OpenCode and Mimo

Tools comparison Jun 19 9 min cursor.com

60 points on the Artificial Analysis index for Fable 5, the highest ever recorded

AI Model Tier List June 2026: The Definitive Ranking After Fable 5's Shutdown

Models analysis Jun 19 7 min artificialanalysis.ai

26% of SWE-bench Verified problems with evaluation awareness that only NLAs detected

Anthropic Peers Inside Claude's Mind: Natural Language Autoencoders Reveal How LLMs Really Think

Explainers analysis Jun 19 6 min transformer-circuits.pub

3 billion parameters (approx.) in Apple's on-device model, sized for A18 and M5-series chips

Apple WWDC 2026: Siri Gets Real AI Intelligence and Developers Get New Tools

Tools news Jun 19 5 min developer.apple.com

72 evals that caught a real production bug in the Risk Agent

Building AI Agents with Claude Code: A Practical Tutorial Roundup

Tools explainer Jun 19 4 min docs.anthropic.com

Claude Code Skills: The Most Powerful AI Coding Agent Extension System

Tools explainer Jun 19 5 min code.claude.com

51 points on the Artificial Analysis composite, the highest ever recorded for an open-weight model

GLM-5.2 Leads Open-Source Models and Closes the Gap with GPT-5.5

Models news Jun 19 6 min z.ai

62% on SWE-bench Pro, by Xiaomi's own count, against 57% for Claude Code with Sonnet 4.6

Mimo Code: Xiaomi's Free Open-Source Terminal AI Agent Challenges Claude Code

Tools analysis Jun 19 2 min mimo.xiaomi.com

56.4 Ultra's score on Terminal Bench 2.1, nearly double Super's 31.0

NVIDIA NeMoTron 3 Ultra: Blazing Fast, Great at Agentic Tasks, Inconsistent at Code

Models analysis Jun 19 5 min build.nvidia.com

Monday, June 15, 2026

32% of SWE-bench Pro's automated evaluations misclassified, according to Datacurve

The credibility crisis in AI benchmarks

Models analysis Jun 15 9 min datacurve.ai

57% or more of the sentences in a Common Crawl sample were machine-translated (AWS AI Labs, 2024)

When AI eats its own data: the synthetic content crisis and model collapse

Models explainer Jun 15 11 min nature.com

78% of organizations use AI in at least one business function, per McKinsey 2025

The risk of outsourcing the brain: when the AI you use can disappear tomorrow

Industry analysis Jun 15 8 min fsb.org

72 hours from Fable 5's launch to a global block after the US government's export directive

The US Government Ordered Claude Fable 5 Shut Down: What Really Happened

Industry analysis Jun 15 7 min anthropic.com

12.93 B USD NVIDIA agreed on September 2 to pay for Hugging Face; the deal has not closed

Open vs. Proprietary AI: Concentration of Power as the Sector's Biggest Risk

Industry analysis Jun 15 7 min nvidianews.nvidia.com

8 IT-GW of compute committed in OpenAI's 20-year lease of the Ohio campus

From the Ohio megacampus to the moratoriums: AI infrastructure meets its limits

Industry analysis Jun 15 6 min nvidianews.nvidia.com

$2.35 spent for every dollar earned in 2024, according to CNBC

OpenAI: financial pressure, free tier degradation, and the advertising dilemma

Industry news Jun 15 7 min cnbc.com

90% of global computing is controlled by the US and China, per the CNAS Sovereign AI Index

The race for digital sovereignty in AI: the Fable 5 shutdown as a turning point

Industry analysis Jun 15 6 min interactives.cnas.org

$210,888 in ARR claimed by BridgeMind for apps built with AI; self-reported and unaudited

Vibe Coding: Code-Free Software Development Scales with Fable 5

Tools analysis Jun 15 9 min anthropic.com

Friday, June 12, 2026

30 days of mandatory data retention Anthropic applies to all Mythos-class traffic

From concentrated power to vibe code: the AI productivity ecosystem in 2026

Tools analysis Jun 12 7 min www-cdn.anthropic.com

65% and up: preference for MiMo-Code over Claude Code on tasks beyond 200 steps, in Xiaomi's own study

Memory that doesn't forget: how Xiaomi's MiMo-Code is changing coding agents

Tools explainer Jun 12 8 min github.com

Thursday, June 11, 2026

0.03% of traffic affected by the invisible safeguards, per Anthropic, not independently verified

Invisible Safeguards in Claude Fable 5: Protection or Silent Sabotage?

Tools analysis Jun 11 5 min www-cdn.anthropic.com

Wednesday, June 10, 2026

80.3% on SWE-Bench Pro for Fable 5 against 58.6% for GPT-5.5, in Anthropic's own table

Claude Fable 5 vs GPT-5.5 Pro: The Frontier of Artificial Intelligence in Two Models

Models comparison Jun 10 7 min www-cdn.anthropic.com

Tuesday, June 9, 2026

1.3 M USD monthly token bill Steinberger once reached, running 100 agents in parallel

The Art of Designing Loops: The New Way to Program with AI Agents

Tools analysis Jun 9 3 min addyo.substack.com

80.3% on SWE-bench Pro according to Anthropic, against 58.6% for GPT-5.5

Claude Fable 5: Anthropic Releases Its Mythos-Class Model to the Public with Safety Guardrails

Models news Jun 9 3 min anthropic.com

78.2% win rate of the new NotebookLM against the previous one on research benchmarks, per Google

NotebookLM Gets Upgraded with Gemini 3.5 and "Antigravity": Now It Researches, Writes Code, and Generates Files on Its Own

Tools news Jun 9 2 min blog.google

5 million weekly active users OpenAI reports for Codex, the engine of the ChatGPT overhaul

OpenAI Plans to Turn ChatGPT into a Superapp: Codex, Image Generation, and Agents in One Place

Tools analysis Jun 9 4 min

~30% rise in releases and features reaching users, against up to 180% more code changes

The AI Productivity Paradox: More Code, But Not More Deliveries

Industry analysis Jun 9 3 min nber.org

Monday, June 8, 2026

75% more logic and correctness errors in AI-generated PRs than in human ones, per Stack Overflow

AgentCrew: The markdown-first operating system that brings order to coding agent chaos

Tools analysis Jun 8 4 min github.com

47% fewer tokens consumed on architecture questions, in CodeGraph's own benchmarks

CodeGraph: The code graph that cuts your AI agent's token usage by 47%

Tools news Jun 8 3 min github.com

6GB of VRAM Gemma 4 12B needs with 4-bit quantization, against ~24 GB in FP16

Gemma 4 12B Unified: Google's first encoder-free multimodal model that runs on your laptop

Models news Jun 8 3 min blog.google

80% or more of Anthropic's production code is written by Claude, according to Anthropic

When AI writes code that improves AI: Anthropic's paper that reignited the debate on recursive self-improvement

Explainers analysis Jun 8 3 min anthropic.com

Friday, June 5, 2026

5 million weekly active users of Codex according to OpenAI, up 6x since the desktop app launched

Codex Reaches 5 Million Weekly Users as Knowledge Workers Adopt It 3x Faster Than Devs

Tools analysis Jun 5 3 min openai.com

80% or more of Anthropic's merged production code, by its own account, was written by Claude

Coding Agents: Claude Writes 80% of Code at Anthropic as the Ecosystem Explodes

Tools analysis Jun 5 3 min anthropic.com

14.9% is GPT-5's score on commercial tasks, down from 23.1% on public ones (Scale AI)

The Problem with Measuring Coding Agents: Inflated Benchmarks, Cheating, and What Really Matters

Explainers analysis Jun 5 4 min arxiv.org

28.65 M secrets hardcoded in public GitHub commits in 2025, per GitGuardian

Coding Agents Are Leaking Your API Keys: 28 Million Secrets Exposed on GitHub

Tools analysis Jun 5 3 min blog.gitguardian.com

550B parameters in NVIDIA's largest open model, with 55B active per token

Nemotron 3 Ultra and Odysseus: The Biggest Week for Open-Source AI

Models analysis Jun 5 4 min research.nvidia.com

2.1-3.4x token use in homogeneous agent teams with no accuracy gain, per the "Cost of Consensus" paper

Multi-Agent Orchestration: How to Run Coding Agents in Parallel Without Stepping on Each Other

Tools analysis Jun 5 4 min minimax.io

Thursday, June 4, 2026

69.2% for Claude Opus 4.8 on SWE-Bench Pro; the best Chinese model reaches 60.6%

The Great Showdown: Chinese Agentic Coding Models vs Claude Opus 4.8

Models comparison Jun 4 10 min

Wednesday, June 3, 2026

35 hours of autonomy ceiling for Qwen 3.7 Plus, with more than 1,000 tool calls

Qwen 3.7 Plus vs 3.6 Plus, 3.7 Max, DeepSeek V4 Pro and Opus 4.8: the definitive coding and agent comparison

Models analysis Jun 3 6 min ofox.ai

Tuesday, June 2, 2026

$1.35 that OpenAI loses for every dollar it takes in, roughly

The Era of Token Scarcity

Industry analysis Jun 2 4 min fortune.com

965 B USD valuation Anthropic announced on May 28, 2026 with its $65 billion Series H

Anthropic: The Startup That Went from $4B to $965B in Three Years and Now Eyes the Stock Market

Industry analysis Jun 2 4 min anthropic.com

four times less likely than Opus 4.7 to omit errors in its own code, per Anthropic's system card

Claude Opus 4.8: Anthropic Made a Model More Honest… and That Brought Unexpected Problems

Models analysis Jun 2 4 min cdn.sanity.io

71% AIME 2024 accuracy of R1-Zero, trained with pure reinforcement learning, up from 15.6%

DeepSeek R1: When Pure Reinforcement Learning Unlocked Self-Reflection in AI

Explainers analysis Jun 2 3 min arxiv.org

$0.87 per million output tokens on V4 Pro, down from $3.48 and now the official price

DeepSeek V4 Pro Makes Its 75% Price Cut Permanent — The Era of Affordable AI Is Here to Stay

Industry analysis Jun 2 3 min api-docs.deepseek.com

2.8 times the performance of Nvidia's export-capped H20, as claimed for Huawei's Atlas 350

Huawei Atlas 350: The Chinese Chip That Challenges Nvidia… with Important Caveats

Industry analysis Jun 2 4 min scmp.com

64 billion parameters in Cosmos 3 Super, the larger of the two variants released

NVIDIA Cosmos 3: The First Open Model That Understands the Physical World

Models analysis Jun 2 4 min blogs.nvidia.com

80.4% on SWE-Bench Verified: level with Claude Opus 4.6 Max, behind GPT-5.5

Qwen 3.7 Max: The Most Unbalanced AI Model

Tools analysis Jun 2 4 min qwen.ai

US Closes Chip Export Loophole to China

Industry analysis Jun 2 5 min bis.gov

Monday, June 1, 2026

AI Agent Architecture — Why Orchestration Matters More Than the Model

Explainers analysis Jun 1 5 min anthropic.com

five times more per task for Claude Opus 4.7 than for Cursor's Composer 2.5 (about $10 vs $2)

The coding agent wars: SpaceX enters the ring at 60 billion

Industry analysis Jun 1 2 min cursor.com

90% of surveyed firms reported AI had no impact on employment or productivity in three years

The great retreat: AI CEOs walk back the labor apocalypse they themselves predicted

Industry analysis Jun 1 3 min nber.org

four times less likely than its predecessor to let errors slip through in its own code, per Anthropic

Claude Opus 4.8: Anthropic bets on models that question themselves

Models analysis Jun 1 2 min anthropic.com

It Doesn't Think, It Predicts: The Inner Mechanism of Generative AI

Explainers analysis Jun 1 5 min arxiv.org

285+ PRs integrated into Hermes Agent 0.15 in 12 days by nearly 200 contributors

Hermes Agent v0.15: the open-source agent that learned to defend itself

Tools analysis Jun 1 2 min github.com

The 7 essential machine learning algorithms — and when to use each one

Explainers analysis Jun 1 7 min hastie.su.domains

How neural networks learn: from initial chaos to pattern recognition

Explainers analysis Jun 1 6 min arxiv.org