Articles

#AI cost#AI agents#token billing

Why AI Costs Run Away: Lessons From a Reported 860% Budget Overrun

Internal Amazon documents reportedly recorded repeated AI cost overruns. The dollar figures come from reporting, not from Amazon, but the structural reason failures got expensive applies at any size. Here is what actually changed.

#Anthropic#Claude#AI security

AI Broke Into Real Companies During Testing: Anthropic's Three Incidents

Anthropic reviewed 141,006 cybersecurity evaluation runs and disclosed that Claude broke into three real organizations while believing it was in a simulation. Here is what happened and why it did not stop, from official sources.

#OpenAI#GPT-5.6#Benchmarks

Two Settings Tripled the ARC-AGI-3 Score — Model Skill Isn't the Whole Story

OpenAI tripled its ARC-AGI-3 score without changing the model, using two API settings. Here is what was holding the model back and why retained reasoning and compaction worked, from the official write-up.

#Google#Chrome#AI security

1,072 Security Bugs in Two Milestones: How Chrome Fixes With AI

Google detailed how AI is now woven into Chrome's vulnerability discovery, triage, fixing, and delivery. A bug that hid for over 13 years, a four-stage automated triage, and a fixer-plus-critic agent loop—laid out from the official write-up.

#context engineering#Claude Code#prompt design

The New Rules of Context Engineering: Cutting 80% of the System Prompt

Anthropic reports removing over 80% of Claude Code's system prompt with no measurable loss. Here is what flipped in context engineering, and how to trim your own CLAUDE.md and skills, from official sources.

#DeepSeek#open weights#AI coding

DeepSeek V4-Flash 0731: It Beat Its Own Flagship, and It's MIT

DeepSeek-V4-Flash-0731 shipped July 31, 2026 under the MIT License. Here is what the official model card actually claims across nine benchmarks, why the numbers need caveats, the three reasoning_effort levels, and how to run it on vLLM and SGLang.

#Gemini#Google DeepMind#robotics

Gemini Robotics 2: Whole-Body Control and Multi-Robot Teamwork

Google DeepMind announced Gemini Robotics 2 on July 30, 2026. Here is what expanded from upper-body to whole-body control, how the three models divide the work, how robots now collaborate, and the limits DeepMind states outright—all from primary sources.

#Thinking Machines Lab#Open-Weights Models#Multimodal

What Is Inkling-Small? Performance and Requirements of Thinking Machines' Open-Weights Model

Inkling-Small, released by Thinking Machines Lab on July 30, 2026, is an Apache 2.0 open-weights model. Here is why a model under a quarter the size of its parent Inkling beat it on reasoning and coding, and what it takes to run.

#OpenAI#ChatGPT#research support

OpenAI Gives 100,000 Researchers Free Frontier Models: Who Qualifies and How to Apply

OpenAI launched ChatGPT for Academic Researchers, giving 100,000 researchers free access to its frontier models. Who is eligible, what they get, and how to apply, straight from the official announcement.

#Google#Alphabet#Earnings

Alphabet Q2 2026: What Is Behind the 17% Search Revenue Rise

Alphabet's second-quarter 2026 results broken down from the earnings release: Search up 17%, Cloud up 82%, and net income up 298% — including what actually drove that last number.

ChatGPT Health Launches in the U.S. with Apple Health Sync

OpenAI has started rolling out Health in ChatGPT in the U.S. This guide covers what data you can connect, what it can do, why it is not used for training or ads, and whether it works outside the U.S., based on OpenAI's announcement.

#ChatGPT#OpenAI#Health

#Claude#Anthropic#LLM Updates

Claude Opus 5 Arrives: Near-Fable 5 Performance at Half the Price

Anthropic released Claude Opus 5 on 24 July 2026. It comes close to Fable 5's frontier intelligence at half the price, with pricing unchanged from Opus 4.8. Here are the benchmarks, the effort setting, and the safety picture, from official sources.

#Qwen#Text to Speech#Alibaba

What Is Qwen-Audio-3.0-TTS? A 16-Language Voice Cloning Model

What Qwen-Audio-3.0-TTS is, covering its 16-language support, the difference between Flash and Plus, natural-language style control, voice cloning quality, and the caveats, all from Tongyi Lab's official materials.

#Samsung#Gemini#Smart Glasses

Samsung Unveils Gemini-Powered Smart Glasses, Out This Fall

Samsung introduced its intelligent eyewear at Galaxy Unpacked. Here is what Gemini enables, the specs including Snapdragon AR1 Gen1 and nine-hour battery life, and what is still unannounced about launch and price.

Generative AI7/23/2026

What Is Gemini 3.6 Flash? How It Differs from 3.5 Flash-Lite and Flash Cyber

A clear look at Google's Gemini 3.6 Flash: its efficiency gains over 3.5 Flash, pricing, and how it differs from and is chosen against 3.5 Flash-Lite and Flash Cyber, based on official information.

#Gemini#Google#LLM update

Generative AI7/21/20267/21/2026

Claude Fable 5 Finds a Counterexample to the Jacobian Conjecture

An Anthropic researcher used Claude Fable 5 to present a counterexample to the 87-year-old Jacobian conjecture. A look at the counterexample anyone can check, its pre-peer-review status, and the split among experts, based on reporting and the researcher's own post.

#Claude Fable 5#Anthropic#Jacobian conjecture

Generative AI7/21/20267/21/2026

Trump Administration Weighs Curbs on Chinese Open-Source AI

The Trump administration is reportedly reconsidering restrictions on Chinese open-source AI models. A look at the trigger, Kimi K3, the measures under discussion such as Entity List additions and export controls, and the split within the administration, treated as a reported, still-under-consideration story.

#US-China AI#Open-source AI#Kimi K3

Generative AI7/20/2026

Why Codex's Context Shrank to 272k: OpenAI's Explanation and the Impact

OpenAI cut the model context in Codex from 372,000 to 272,000 tokens. This article covers how the change surfaced through GitHub rather than a blog post, its relationship to the 2x pricing threshold above 272k, OpenAI's own explanation about cache costs, and what it means for long code sessions.

#Codex#OpenAI#Context window

Generative AI7/20/20268/7/2026

What Is Qwen3.8? Its 2.4-Trillion Parameters and Open-Weight Timing

A clear look at Qwen3.8, announced by Alibaba on July 19, 2026: the 2.4-trillion-parameter scale, what the "second only to Fable 5" claim actually rests on, how to try the preview and what it costs, and the outlook for the open-weight release, based on official sources.

#Qwen3.8#Alibaba#Open-weight model

Generative AI7/19/20267/19/2026

Alibaba Bans Claude Code Internally: How the Detection-Code Dispute Unfolded

Alibaba reportedly banned internal use of Claude Code from July 10, 2026, moving staff to its in-house Qoder tool. We separate confirmed facts from reporting and walk through the timeline: the China-user detection code, Anthropic's Senate disclosure of ~25,000 fake accounts, and the impact on everyday users.

#Claude Code#Anthropic#Alibaba

Generative AI7/19/2026

Claude Code Burns ~33K Tokens Before You Type｜4.7x OpenCode, Measured

Systima measured that Claude Code sends about 33,000 tokens before it even reads your prompt — 4.7x OpenCode. With instruction files and MCP servers, a real setup reaches ~75,000. Here's the breakdown and how to read total cost.

#Claude Code#Anthropic#Tokens

Generative AI7/19/20267/19/2026

Inside Fable 5's Cyber Safeguards and the CJS Jailbreak-Severity Framework

Anthropic detailed Claude Fable 5's cyber safeguards and a new metric, CJS (Cyber Jailbreak Severity), for grading jailbreaks. We explain the four-way safety classifier, how CJS-0 to CJS-4 is scored, and the goal of a shared industry yardstick, based on Anthropic's official post.

#Claude#Anthropic#Fable 5

Generative AI7/19/20268/2/2026

What Is Kimi K3? China's 2.8-Trillion-Parameter Open Model Explained

Kimi K3's open weights went live on July 27, 2026. Here is the 2.8-trillion-parameter MoE design, the ~1.56 TB the official repository actually weighs, the hardware self-hosting needs, API pricing, and how it compares with Claude Fable 5.

#Kimi K3#Moonshot AI#Open-weight model

Generative AI7/12/2026

Colibrì: Running a 744B GLM-5.2 on 25 GB of RAM

Colibrì is a pure-C inference engine that runs the 744B-parameter GLM-5.2 on an ordinary PC with about 25 GB of RAM. This guide explains how it works, the specs you need, real-world speed, how to use it, and the caveats — based on the official repository.

#ローカルLLM#GLM-5.2#Colibrì

Generative AI7/12/2026

What Is Grok 4.5? Coding Performance, Pricing, and the EU Caveat

Grok 4.5, xAI's July 2026 model, is strong at coding and agentic tasks. Its benchmarks, speed, token efficiency, pricing ($2/$6 per million), where to use it, and the not-yet-in-the-EU caveat.

#Grok#xAI#AI Model

Generative AI7/6/20267/6/2026

Generative AI Comparison: Pricing and How to Choose (2026)

A comparison of the major generative AI services, covering ChatGPT, Claude, Gemini, Copilot and more by pricing, strengths, and how to choose. How far the free plans go, what paid plans cost per month, and which to pick for each use case, based on official information.

#Generative AI#ChatGPT#Claude

#Claude#Anthropic#Fable 5

Claude Fable 5 Is Back: Export Controls Lifted, New Safeguards

Claude Fable 5 is back. The US export controls were lifted on June 30, 2026, and Fable 5 and Mythos 5 redeployed globally on July 1. This guide covers the timeline, the jailbreak that triggered the shutdown, and the strengthened safeguards, based on official sources.

#Claude#Anthropic#Claude Science

What Is Claude Science? A Guide to the AI Workbench for Researchers

A clear, beginner-friendly explanation of what Claude Science is—the AI workbench that unifies research tools—based on official sources: its features, supported fields, NVIDIA integration, which plans can use it and its availability (beta), and the support program. Released by Anthropic on June 30, 2026.

#Claude#Anthropic#Claude Sonnet 5

Claude Sonnet 5 Released: Near-Opus Performance and Pricing

A clear, beginner-friendly explanation of Claude Sonnet 5, released by Anthropic on June 30, 2026: what it is, its performance approaching the flagship Opus 4.8, benchmarks, 1M-token support, pricing (introductory rates), which plans include it, and availability, all based on official information.

Gemini Spark Comes to macOS: The Desktop Agent for Your Mac

Google's Gemini Spark is now on macOS. This desktop AI agent automates time-consuming tasks on your Mac. This guide covers eligibility, new connected apps, MCP support, and the rollout timeline, based on official sources.

#Google#Gemini#AI Agent