AI and system design news, explained
When something changes in AI or system design, we run it ourselves and explain what it means for the systems you build, in plain words, with the sources linked.
Qwen4-Max explained: what changed from Qwen3.8-Max, what it costs, and whether you can download itQwen4-Max is the new flagship AI model from Alibaba's Qwen team, released on 16 November 2026. It keeps the 2.4 trillion parameter size of Qwen3.8-Max but uses 71 billion of them for each token instead of 95 billion, reads up to one million tokens, and costs 25% less per token. Unlike Qwen3.8-Max, its weights are not available to download; you use it through Qwen's cloud.Qwen4-MaxQwen 4QwenAlibabaread the explainer · 8 min read
Old Bitly links hijacked: why bit.ly links in old Burger King, McDonald's and Trump tweets went somewhere elseOn Sunday 15 November 2026, short links in tweets that Burger King and McDonald's posted in 2011 started sending people to pages that had nothing to do with burgers. Old links in Donald Trump's early tweets did the same. Nobody broke into those accounts. The links themselves had been created again on Bitly's side and pointed somewhere new. Bitly said an issue let bad actors repurpose old, inactive links, and it disabled them. Here is what happened, why a short link can change at all, and what every link service should do about it.Bitlybit.lyURL shortenerlink hijackingread the explainer · 10 min read
Lovable apps exposed passports and private messages: what the study found and how to check your appA Hamburg security firm looked at 159,041 apps built with Lovable, the AI app builder, and found apps whose databases handed out names, emails, passport numbers and private chats to anyone who opened them, with no login. Nothing was hacked in the usual sense. The data sat behind a door nobody had locked. Here is what the study found, the one database setting behind most of it, and a test you can run to see the difference yourself.LovableLovable securitySupabaserow level securityread the explainer · 9 min read
The National Weather Service outage of 13 November 2026 explained: was the weather radar down, and what brokeOn Friday 13 November 2026, the same afternoon the FAA stopped flights across the United States, the National Weather Service warned that its web pages might show old products or none at all, and screenshots of a "national radar outage" spread online. By that evening the NWS said its radars and its network were working normally. We checked its own public data feed: the forecast offices kept writing forecasts at the usual rate all day. Here is what broke, what did not, and what engineers can learn from it.National Weather Service outageNWS radarweather.govoutageread the explainer · 12 min read
The Anthropic IPO explained: dates, price range, ticker ANT, and what the S-1 filing says about computeAnthropic, the company behind Claude, is selling shares to the public for the first time. Its filing with the US Securities and Exchange Commission sets a price range of $70 to $78 for 825 million shares on Nasdaq under the ticker ANT, with the price expected on 18 November 2026 and trading on 19 November. We read the filing ourselves and pulled out the dates, the terms, who keeps control, and the numbers an engineer will find most interesting: what it costs Anthropic in compute to earn each dollar.Anthropic IPOAnthropicClaudeIPOread the explainer · 17 min read
DeepSeek's application APIs explained: web search, files and links in one call, and what an answer costsOn 13 November 2026 DeepSeek opened the features of its own chat app to developers. With one or two extra fields in an API request, the deepseek-flash model can now search the web, read the links a user pastes, answer from uploaded PDFs and Office files, and run a safety review on its own answers. I read every guide DeepSeek published with it, ran DeepSeek's open-source demo code, and priced DeepSeek's own example answers. Two details will catch teams out: the usage field that looks like the search count is not what you are billed for, and one field must come back to the API unchanged or the model forgets the page it read from a link.DeepSeekDeepSeek APIdeepseek-flashweb search APIread the explainer · 16 min read
Claude outage, 13 November 2026: what broke, what Anthropic said, and how to keep your app workingOn 13 November 2026 Claude went down twice. Both outages are over: when I checked at 04:51 UTC on 14 November, Anthropic's status page said all systems were operational. From 09:58 UTC (15:28 in India) people had trouble signing in to claude.ai and Claude Code for about two hours, and the Console was listed as affected too. From 13:09 UTC a second, wider outage brought errors to claude.ai, the Claude API, Claude Code and the connectors for another two and a half hours. This page sets out what Anthropic's status page said, what each outage did and did not touch, how often Claude has had incidents over the last 92 days, and how to build your own app so the next one does not take it down with it. I tested the last part with Anthropic's real Python SDK.Claude outageClaude downClaude CodeClaude APIread the explainer · 8 min read
llama.cpp 1.0 explained: what the first stable version changes, the new FP8 KV cache, and how to upgradellama.cpp, the open-source engine that runs large language models on laptops, phones and servers, reached version 1.0.0 on 13 November 2026. The big change is a promise: from now on a breaking change to its C API means a new major version. The release also adds an FP8 KV cache, which halved the memory of our test cache, and tightens the security of llama-server.llama.cppggmllocal LLMLLM inferenceread the explainer · 10 min read
The FAA nationwide ground stop of 13 November 2026 explained: what happened, the timeline and the causeOn Friday 13 November 2026 the FAA stopped every departing flight at every airport in the United States for 4 hours and 23 minutes, from 3:53 pm to 8:16 pm Eastern time. The FAA says communication issues hit several of the centers that control planes between airports. We rebuilt the night from the FAA's own advisories, and explain what a ground stop is, why it is the safe choice, and what engineers can learn from it.FAA ground stopair traffic controloutageincident responseread the explainer · 18 min read
Microsoft 365 outage, 12 November 2026: why Outlook and Exchange went down, and what else brokeOn 12 November 2026 at 16:17 UTC (21:47 in India) users began losing access to their Outlook mail in Exchange Online, the Microsoft service that stores and delivers Outlook mail, calendars and contacts. Twelve hours later it was still not fully fixed. Microsoft listed 18 other services as affected, from Teams calendars to Copilot and Intune. This page explains what Microsoft has said so far, what broke and what kept working, what you can do while it lasts, and why an outage like this goes on for so long. I also tested Microsoft's servers from India while it was happening.Microsoft 365 outageOutlook downExchange OnlineMicrosoft Teamsread the explainer · 10 min read
Gemini 4 Hydrogen explained: price, thinking levels, free tier and how it compares with Carbon and ArgonGemini 4 Hydrogen is the fast, lower-cost member of Google's Gemini 4 family, released on 12 November 2026. It costs $1.50 per million input tokens and $7.50 per million output tokens at any length up to its 1 million token window, it is on the free tier of the Gemini API, and it adds a fourth thinking level called xhigh. Google pitches it as the model that does the everyday work while Carbon and Argon handle the hardest reasoning.Gemini 4 HydrogenGemini 4GoogleGemini APIread the explainer · 10 min read
OpenAI Control Room explained: how OpenAI stops a running AI job within 35 secondsControl Room is the system OpenAI uses to stop its own research AI jobs when a safety or security problem shows up. Every job gets an id, every inference engine asks Control Room before it answers a request, a stop takes effect within 35 seconds, and if Control Room itself goes down, the engines refuse everything. OpenAI described it on 9 November 2026.OpenAIControl RoomAI kill switchAI safetyread the explainer · 18 min read
Gemini Decisions API explained: Google's model that answers with probabilities, and what it really costsOn 11 November 2026 Google released the Gemini Decisions API and a new model, Gemini Decision 1. It never writes text. You send one input, such as a support ticket, a photo or a call recording, with up to 1,000 questions, and it returns a probability for each possible answer; Google says in about a fifth of a second. I read the launch post and the docs, counted the real tokens for an egg photo I had already sent to OpenAI's similar Decisions API in October, and worked out where this API saves money and where it does not.Gemini Decisions APIGemini Decision 1GoogleGemini APIread the explainer · 8 min read
OpenAI Codex Security explained: how it scans code for vulnerabilities, and what changed in version 0.10.0Codex Security is OpenAI's tool for finding security bugs in a repository with a model that reads the code, checks each finding against the source and drafts a fix. Version 0.10.0, released on 12 November 2026, lets an interrupted scan resume, puts one budget over a batch of scans, and explains why a fix was blocked.Codex SecurityOpenAIapplication securitySASTread the explainer · 9 min read
AWS Emulator explained: test S3, DynamoDB and other AWS calls on your laptop, with no AWS accountAWS has published AWS Emulator, its own program that answers AWS API calls on your machine or in CI, so tests can create buckets, tables and queues without an AWS account or a bill. It is open source, free, and honest about its gaps: it does not check IAM policies, keeps everything in memory, and publishes a fidelity score per service.AWS EmulatorAWSlocal testingintegration testingread the explainer · 8 min read
.NET 11 explained: C# 15 unions, the MCP server template, faster tooling, and what to do about .NET 8 and 9Microsoft released .NET 11 on 10 November 2026. It brings union types in C# 15, an MCP server template in the SDK for building tools that AI apps can call, OpenAPI 3.2 by default in ASP.NET Core, and faster everyday commands. The same day, .NET 8 and .NET 9 reached end of support, so anyone still on them should plan a move to .NET 10 or .NET 11..NET 11C# 15ASP.NET Coreunion typesread the explainer · 8 min read
MiniMax-H3-Image explained: the open-weight 16-megapixel image model, its price, hardware and licenceMiniMax-H3-Image is an image model from MiniMax whose weights you can download. It makes new images from text and edits up to 16 reference images, at about 16 megapixels, in 6 sampling steps. It came out on 10 November 2026. Before you build on it, read the licence: it does not cover the United States, the United Kingdom, the European Union or South Korea.MiniMax-H3-ImageMiniMaxMiniMax H3image generationread the explainer · 7 min read
Claude Fable 5.2 and Mythos 5.2 explained: one model, two sets of safeguards, and a lower priceClaude Fable 5.2 and Claude Mythos 5.2 are the same underlying model from Anthropic, released on 9 November 2026. Fable 5.2 has safeguards that block most cybersecurity and research biology tasks, so anyone can use it; Mythos 5.2 does not, so only vetted organisations get it. Both cost $8 per million input tokens and $40 per million output tokens.Claude Fable 5.2Claude Mythos 5.2AnthropicLLM pricingread the explainer · 6 min read
Meta Muse Ember explained: the small Muse model built for the work agents repeat all dayMuse Ember is the smallest and cheapest model in Meta's Muse family, released on 9 November 2026. It is built for the steps an AI agent repeats all day, such as tool calls, routing, extraction and reading long documents. It costs $0.10 per million input tokens and $0.50 per million output tokens, reads text, images, audio and video, and takes up to 1 million tokens in one request.Muse EmberMetaMuse Sparksmall language modelsread the explainer · 18 min read
Revolut self-hosts GLM-5.3 for 2,000 engineers: the GPUs, the settings and what the $0.30 really meansRevolut, the European digital bank, now gives its 2,000 engineers an open model, GLM-5.3 from the Chinese lab Z.ai, running on 56 NVIDIA B200 GPUs it controls. It says its AI coding bill is now a tenth of what it paid Anthropic and OpenAI. I went through the talk where Revolut said this, checked the model's files, worked out how it fits on the GPUs, and priced our own coding agents' token mix to see what "30 cents per million tokens" really compares with.GLM-5.3Revolutself-hosting LLMsNVIDIA B200read the explainer · 9 min read
Pi coding agent explained: the minimal open-source agent harness developers are switching toPi is an open-source coding agent that runs in your terminal. It gives a model four tools by default, read, bash, edit and write, saves every session as a tree you can branch, and leaves out features like sub-agents and plan mode on purpose so you can add only what you need. It was on the Hacker News front page on 8 November 2026 with a developer's post about moving to it from Claude Code.Picoding agentagent harnessopen sourceread the explainer · 9 min read
Gemini Omni 1.2 Flash explained: 15-second AI video, conversational edits and a 30% lower priceGemini Omni 1.2 Flash is Google's fast video model. It turns a text prompt, images or short clips into a video with sound, now up to 15 seconds long, and lets you change the clip by simply asking for a change in the next message. Google made it generally available on 5 November 2026 at about $0.07 per second of 720p video, which is roughly 30% less than Omni 1.1.Gemini OmniGemini APIAI video generationGoogle DeepMindread the explainer · 18 min read
Claude Code's "don't ask again" saved passwords in a project file: 1,122 reached GitHub. Check yoursWhen you pick "Yes, and don't ask again" on a Claude Code permission prompt, it saves the command as a rule in .claude/settings.local.json, and if you typed a password into that command, the password is saved too. Apiiro counted 1,122 public GitHub copies of the file holding a real secret. I simulated the five ways the file meets git, ran gitleaks and TruffleHog over a fake one, and wrote a check that finds what both missed.Claude CodeClaude Code securitysettings.local.jsonsecrets in gitread the explainer · 10 min read
libghostty 1.0 explained: the terminal engine inside JetBrains IDEs, now with a stable C APIlibghostty is the terminal emulator from the Ghostty app, cut out as a small C library that any program can embed. Version 1.0 of libghostty-vt shipped on 6 November 2026, and it promises a stable interface: a program built against it today keeps working with later 1.x releases. The terminal in JetBrains IDEs already runs on it.libghosttyGhosttyterminal emulatorC libraryread the explainer · 10 min read
Agent memory explained: Anthropic's components of agent memory, dreaming, and what the tests showedAgent memory is what an AI agent learns from its own work and keeps between sessions, saved as files it can read later. On 6 November 2026 Anthropic published how it designs memory for Claude: the agent writes short notes while it works, and a separate pass called dreaming rereads old transcripts to fix wrong notes, build linked wiki pages and keep the always-loaded part small.agent memoryAI agentsAnthropicClauderead the explainer · 10 min read
MaxxBench explained: why DeepSWE coding scores fell by up to 25 points when the tasks changed a littleMaxxBench is a coding benchmark from Unlap that takes 89 tasks from DeepSWE, a popular benchmark for AI coding agents, and changes each one a little, such as asking for lowercase filenames. Models that tied at about 74% on DeepSWE ended up 23 points apart. Reading the agents' reasoning, the authors found many of them guessing what the hidden tests wanted instead of doing what the prompt said.MaxxBenchDeepSWEcoding agentsbenchmarksread the explainer · 17 min read
Claude Code sent users' email addresses to websites: why it happened, what the fix covers, how to block itClaude Code gives the model your signed-in email address. When the model wrote its own curl commands, it sometimes put that address in the User-Agent header, the line that tells a website which program is calling, so the address landed in other sites' logs. Anthropic shipped a stopgap (a temporary fix) and promises the real fix early in the week of 9 November. I ran the stopgap's published check on 16 commands, scanned this Mac's own Claude Code logs, and wrote a stricter hook you can install today.Claude CodeClaude Code privacyUser-AgentAI agent privacyread the explainer · 17 min read
Tencent Hy4, explained: what Tencent shipped, what it costs, and what running it yourself would takeHy4 is the newest large AI model from Tencent, the Chinese company behind WeChat and QQ. It went live on 4 November 2026 on Tencent Cloud and OpenRouter, and, unlike its August preview, it reads images as well as text. Tencent says it will publish the weights (the model files anyone can download) but gives no date. I read Tencent's own post, checked what OpenRouter and Hugging Face actually list, and worked out what an agent task costs and what servers the open preview model needs.Tencent Hy4Tencent HyHy4 previewopen-weight modelsread the explainer · 19 min read
Weight-space programming, explained: GPT-6 Astra edited Qwen's weights with no training; I checked the diffOn 2 November 2026 four researchers at MIT CSAIL showed that OpenAI's GPT-6 Astra can read a neural network's weights (the numbers a model learns in training) and change them by reasoning, with no training at all. Their biggest result: Astra changed 131,072 of Qwen3.5-0.8B's weights and its math score went from 7.9% to 48.1%. I downloaded both models, compared them byte by byte, and read how their published answers end.weight-space programmingGPT-6 AstraQwenQwen3.5-0.8Bread the explainer · 17 min read
Kimi K3.1 Pro, explained: what Moonshot shipped, what it costs, and why the thinking you send back adds upKimi K3.1 Pro is the newest large AI model from Moonshot AI, a company in China. It went live on 23 October 2026 in Moonshot's API and its coding tool Kimi Code, and Moonshot says the weights (the model files anyone can download) will be public by 4 November. It is the same model as Kimi K3, with 2.8 trillion parameters (the numbers a model learns in training), trained further, and it costs a third less than K3 on Moonshot's API. I read Moonshot's own pages, ran its prompt code and tokenizer on my Mac, and worked out what an agent task costs when, as Moonshot requires, you send the model's thinking back on every step.Kimi K3.1 ProMoonshot AIKimi K3open-weight modelsread the explainer · 11 min read
CVE-2026-111497 (Node2Shell), explained: the Brotli bug in Node.js, .NET and Envoy, and how to checkCompressed data is unpacked all the time without anyone asking: a web reply, an upload, a font. A bug in Google's Brotli decoder lets one crafted compressed stream write past the end of a buffer, which can crash the program, leak its memory or run an attacker's code. It is being exploited, it is on CISA's must-patch list with a 10 November deadline, and Node.js, Bun, .NET, Envoy, Istio and Python have shipped fixes.CVE-2026-111497BrotliNode2ShellBroccoli2Shellread the explainer · 14 min read
Gemini 4 Carbon, explained: Google's security model, who gets it first, and how to test a code fixGoogle announced Gemini 4 Carbon on 28 October 2026 as its strongest model, built for finding and fixing security holes in code. Only a group of cyber defenders can use it for now, so I checked our API key, then built the small test harness that decides whether any model's security fix actually works.Gemini 4 CarbonGoogleGemini APIsecure codingread the explainer · 20 min read
Neo4j Context Engine, explained: an ontology that tells AI agents what your words mean, and a lab you can runOn 19 October 2026 Neo4j announced Context Engine, a product that gives AI agents a map of what a company's words mean and where its data lives. It is in private preview, so I could not use it. I built a small version of the same idea with open tools, asked 12 business questions three ways, and measured what happened.Neo4j Context EngineNeo4jGraphRAGknowledge graphread the explainer · 21 min read
NVIDIA Nemotron 3.5 Bolt and Super VL, explained: a 74B agent model in 22 GB, and what it needs to runOn 15 October NVIDIA released two open Nemotron 3.5 models for AI agents: Super VL, a 120B model for servers, and Bolt, a 74B model squeezed to 22 GB so it can run on one local GPU. I read the Bolt file's own header without downloading it, counted its 74 billion parameters and its 2-bit experts myself, and worked out what memory it really needs: more than the "24GB" in its name.Nemotron 3.5Nemotron 3.5 BoltNemotron 3.5 Super VLNVIDIAread the explainer · 16 min read
JEV-35B, explained: an open decision model you can download, what it needs, and a test of its prompt trickJEV-35B is a model on Hugging Face that never writes text. You give it a situation, one question and up to 256 answers, and it returns a probability for each answer in about a fifth of a second, by its maker's measurement. Unlike the hosted decision APIs from TypeSafe, OpenAI and Microsoft, the weights are free to download under Apache-2.0. It is too big for my 24 GB Mac, so I checked its own config files and tried its prompt layout on a small chat model: with that layout the small model changed 8 to 11 of 24 answers when I reversed the options.JEV-35BAutoTrustdecision modelsclassificationread the explainer · 16 min read
Intel GPU D-series (D270), explained: 480 GB of memory on one card, and what that buys you for AI inferenceIntel's new data center GPU, the D-series (formerly Crescent Island), puts up to 480 GB of memory on one air-cooled PCIe card, five times the 96 GB of the NVIDIA card Intel compares it with, at about the same memory bandwidth (how many bytes the chip can read each second): 1.5 TB/s. I measured the rule that decides what that means on my own Mac, then worked it out for the card: more users and longer conversations held in memory, not faster answers.Intel GPU D-seriesIntel D270Crescent IslandAI inference hardwareread the explainer · 14 min read
EmbeddingGemma 2, explained: one open model for text, code, pictures and sound, tested on a laptopOn 6 October 2026 Google released EmbeddingGemma 2, a free, open model that turns text, code, pictures, video and sound into lists of 768 numbers, so that things with similar meaning get similar numbers. That is the engine behind search by meaning and RAG. I ran it on a MacBook Air: it found the right text for 7 of 8 plain-English questions and the right picture for 6 of 6, at about 47 ms per sentence, and cutting its vectors to a sixth of their size kept the answers but squeezed the scores together.GoogleEmbeddingGemma 2embeddingssemantic searchread the explainer · 13 min read
REA (Reverse Engineer Anything), explained: what the viral GitHub repo does, and what it missed in a real testREA is a free, open-source tool that lets an AI coding agent such as Claude Code or Codex look inside an app you do not have the source code for, and explain how a feature works. It went from 3 points on Hacker News in July to more than 109,000 GitHub stars this week. I ran it on a real open-source app: every place it pointed to was right, but it did not record any of the 8 places where the app sends a reply back to a window. Its own output warns that a missing finding proves nothing.REAreverse engineeringMCPAI agentsread the explainer · 12 min read
Microsoft-Decision-1, explained: a model that scores answers instead of writing them, and a test to runOn 9 October 2026 Microsoft released Decision-1, a model that never writes text. You give it some text, a question and a fixed list of answers, and it returns a probability for each answer. It costs $0.042 per million input tokens, output is free, and it runs in Microsoft Foundry and on OpenRouter. I could not call it myself, so I ran the test behind Microsoft's main robustness claim on two other models: qwen2.5:3b, a small chat model, changed its answer on 4 of 24 tickets when I reversed the options, and it was at least 99.9% sure both times.MicrosoftMicrosoft-Decision-1decision modelsclassificationread the explainer · 8 min read
OpenAI's Decisions API, explained: I ran it on 33 egg photos, and the top answer nearly binned a clean eggOpenAI's Decisions API, in public beta since 6 October 2026, does not write a reply. You send a photo or text, a question and a fixed list of answers, and it returns a probability for each. I ran it on 33 egg photos: it answered in about 0.2 seconds, and one clean egg came back 47% dirty, 44% clean. Here is what that means for your code.OpenAIDecisions APIclassificationconfidence thresholdsread the explainer · 20 min read
Claude Haiku 5.5, explained: a tenth of the price up to 100,000 tokens, and what breaks when you switchClaude Haiku 5.5 is Anthropic's new small, fast model, released on 7 October 2026. Up to 100,000 tokens of prompt it costs a tenth of Haiku 4.5 per token. I read Anthropic's pages, ran the same 12 support tickets through both models, and worked out where the price jumps and which old code stops working.Claude Haiku 5.5AnthropicClaudeLLM pricingread the explainer · 21 min read
DeepSeek V4.1 Flash, explained: an open model that does less work reading your prompt, and what it costsDeepSeek V4.1 Flash is a large AI model from the Chinese company DeepSeek that anyone can download under the MIT licence or call through DeepSeek's API for a low price. It came out on 10 September 2026. I read DeepSeek's own pages, ran its prompt code and tests on my Mac, and worked out the bill and the hardware.DeepSeek V4.1 FlashDeepSeekopen-weight modelsmixture of expertsread the explainer · 17 min read
Mistral Large 4, explained: a 1 trillion parameter model with open weights promised, and what it takes to runMistral Large 4 is a very large AI model that companies will soon be able to download and run on their own servers, if they can afford the machine. Mistral, a French AI company, opened an early preview on 6 October 2026 and promises the files by the end of October. I checked what I could reach today, ran the same code on a small Mistral model, and worked out the hardware and the bill.Mistral Large 4Mistral AIopen-weight modelsmixture of expertsread the explainer · 16 min read
OpenAI Codex explained, and how it compares with Claude CodeCodex is OpenAI's AI coding agent: a program that reads your code, runs commands and makes changes on its own, in a terminal, a code editor, the ChatGPT app or OpenAI's cloud. This page explains what Codex is in October 2026, what it costs, and how it differs from Claude Code. I also tried it with a free local model and no OpenAI account. It found the bug, then failed to save the fix.CodexOpenAIClaude CodeAI coding agentsread the explainer · 14 min read
Qwen 3.8 explained: what is new, and running the 27B model with OllamaQwen 3.8 is the newest family of AI models from Alibaba's Qwen team. Three of its models have open weights: files you can download and run yourself. I ran the smallest one, Qwen3.8-27B, with Ollama, a free tool that runs models on your own machine, and recorded every step. This page covers what is new, what it needs, and what I measured.Qwen 3.8Qwenopen-weight modelsOllamaread the explainer · 19 min read
Claude Code mods explained: build your first mod, and what it can reachA Claude Code mod is a small piece of JavaScript or TypeScript that runs inside Claude Code, Anthropic's AI coding tool, and changes how it looks and what it does. It comes inside a plugin, an add-on package. I built one on my Mac, ran it, and read what a mod is allowed to touch.Claude CodeClaude Code modsAnthropicpluginsread the explainer · 16 min read
ChatGPT 6.1 Sol explained: what GPT-6.1 Sol is, what it costs, and when to use itGPT-6.1 Sol is the middle of OpenAI's three GPT-6 models, released on 29 September 2026. You pay per token, a small piece of text: $2 for every million tokens you send and $10 for every million it writes. That is 20 times the price of Luna, the cheap model, and one fifth of Astra, the most capable one. I used it with my own OpenAI account and compared it with Luna on 300 support tickets.GPT-6.1 SolChatGPTOpenAILLM costread the explainer · 16 min read
Claude Code, explained: what changed this week, and a real bug fix in four turnsClaude Code is Anthropic's AI coding agent. On 1 October 2026 it gained Claude Mods, a way to change how the tool itself behaves. I read the changelog and docs, then gave it a failing test on my own Mac and recorded every step, every token and the cost.Claude CodeAnthropicAI agentscoding agentsread the explainer · 12 min read
Gemini 4 Argon, explained: Google's new frontier model, and why I could not call it yetGoogle announced Gemini 4 Argon on 30 September 2026, with a 1 million token output limit and a price of $2 in and $10 out per million tokens. It is only open to a small group of cyber defenders for now, so I checked our own API key, got a 404, and turned that into a lesson on choosing models in production.Gemini 4 ArgonGoogleGemini APILLM pricingread the explainer · 11 min read
OpenAI dots, explained: how an always-on AI agent keeps working after you close the chatOn 29 September 2026 OpenAI launched dots: ChatGPT agents that have their own cloud computer and keep working after you leave. Here is what a dot is, who can use one, and how you would design an agent like this yourself.OpenAIAI agentsChatGPTdurable executionread the explainer · 12 min read
Kolibri, explained: Aleph Alpha's open model computes like a 3B model but needs memory like a 78B oneAleph Alpha released Kolibri on 3 October 2026: downloadable weights under Apache 2.0, built for German and English. It would not run on our 24 GB Mac, which turned out to be the most useful lesson, so we ran its tokenizer instead and worked out what it needs from your hardware.KolibriAleph Alphaopen-weight modelsmixture of expertsread the explainer