标签：#ai

Online publishers are getting more control over whether their websites appear in Google's AI Search features, thanks to a UK regulatory ruling. The new conduct rule imposed by the Competition and Markets Authority (CMA) requires Google to let website owners keep their content out of features like AI Overviews and prevent it from being used […]

TL;DR Day 2 of AI Native DevCon shifted from agent capability to operating discipline. The strongest sessions focused on how teams can run AI-native delivery with clearer context pipelines, measurable agent behavior, safer execution boundaries, and better organizational ownership. The scale showed up in the numbers too. Across the two days, DevCon brought together 650+ in-person registrations, around 2,000 online registrations, and a packed mix of sessions, workshops, hallway conversations, and practical lessons. Day 2 leaned into workshops. That shift mattered because the second day was less about proving agents can do useful work and more about showing how teams can make that work repeatable. Hey there, welcome back. Rohan Sharma here again continuing the devcon series. Day 1 gave us the framing, including Guy Podjarny ’s core point that skills should be treated like real software assets. Day 2 picked up from there and moved into the operating details. Once agents are inside daily engineering work, platform and product teams need to decide what changes first, who owns those changes, and how the results are measured. Talks that shaped Day 2 Harness engineering beyond code Marc Sloan from Tessl focused on the next gap many teams are hitting. Code context is increasingly structured, but product and design context still lives in external systems such as Figma, Notion, and Linear. Pulling that context live can reduce staleness, but it introduces drift in evals, versioning, and reproducibility. The practical lesson was to stop treating external product and design context as random reference material. Teams need a defined layer between the repository and those external systems, with clear versioning so evaluations can be replayed against known context snapshots. Without that, agents can produce work that looks technically correct while missing the product constraint that actually mattered. That is a very expensive kind of almost-right. From vibes to metrics Simon Obstbau

AI Coding Agents in 2026: From Pair Programming to Autonomous Teams Slug: ai-coding-agents-2026-stack-comparison 1. The Three Categories That Actually Matter The 2024‑2025 hype cycle treated every AI coding tool as a single‑dimensional “best‑of‑list.” 2026 data shows that professional developers now average 2.4 tools per workflow (Stack Overflow Survey 2025). The real decision is architectural: Layer Goal Typical Agent Type Line‑level editing Speed, low latency Editor assistants Repo‑level planning Context depth, multi‑file changes Autonomous agents Enterprise governance Isolation, audit, CI/CD integration Platform agents Choosing a “one best tool” ignores the trade‑off between context window size (how many tokens the model can see) and execution speed (how fast the tool returns a suggestion). A narrow‑window editor assistant excels at instant autocomplete, while a wide‑window autonomous agent can rewrite an entire microservice in a single run. The three‑tier framework aligns the tool’s strengths with the architectural layer where they matter most. 2. Tier 1: Editor Assistants — Speed at the Line Level Tool Market Position Key Feature (2026) Pricing (per developer) Cursor $500 M+ ARR, fastest growth in Q1 2026 Parallel agents update git worktrees; 2‑second latency on 8‑core laptops $15 /mo (individual) – $120 /mo (team) GitHub Copilot 4.7 M paid subscriptions, 75 % YoY growth Agent Mode with multi‑agent workflows; deep VS Code integration $10 /mo (individual) – $100 /mo (enterprise) Windsurf 1.2 M active users, strong UI polish Real‑time code‑style enforcement; limited to 4‑file context Free tier up to 5 k lines, $30 /mo premium Tabnine Enterprise‑only after 2026 pivot Air‑gapped deployment; NVIDIA Nemotron 4‑bit models for on‑prem inference $200 /mo per seat (minimum 10 seats) When to choose each Cursor – prioritize raw typing speed and git‑aware suggestions. Ideal for startups that need rapid iteration without heavy IDE lock‑in. Copilot – best for teams already on

Here's a trace that reset how I think about evaluating tool-calling agents. An agent tries to book a flight. It calls search_flights with departure_date="next Friday" . The endpoint expected an ISO date, so it returns a 400 . The agent retries the same string four times, then apologizes to the user and gives up. Now the part that actually bothered me. Tool selection was correct. The model picked the right function out of a registry of 28. My tool-selection accuracy logged a clean 1.0 . The aggregate task-completion logged a 0 . And neither number told me which of three things broke: the argument was wrong, the model never read the 400 body, or the retry policy looped on the same input. My eval wasn't wrong. It was asking the wrong question. What "tool-call accuracy" actually grades If the only thing you measure is did the agent call the right tool , you're testing intent, not execution. Tool selection is necessary, not sufficient. It passes the moment the right function name shows up in the trace, completely blind to whether the arguments were garbage, whether the model read what came back, or whether it recovered from the 400 . That's the gap. The metric checks that the agent started the right way. Production needs to know whether it finished the right way. The reframe: it's four eval problems, not one The thing I had to internalize is that tool-calling eval is four problems stacked, each with its own root cause: Tool selection , right tool, or correctly no tool Argument extraction , schema-valid and semantically correct Result utilization , did it actually use what the tool returned Error recovery , did it retry, fall back, or escalate Score them separately and "the agent failed" collapses into "the argument extractor regressed on date strings on the flight-booking path." One bisect instead of three days. What I rebuilt Layer 1: Tool selection (with the bucket everyone drops) F1 on the tool name, so a 28-tool registry doesn't hide a regression on one rare endpoint

I run multiple Claude Code sessions all day — one per feature, one per service, sometimes five at once. Every session was asking me for permission in its own terminal. I'd miss requests buried in a background tab. I'd switch windows mid-thought just to approve a git status . I'd lose context constantly. And there was no single place to see what Claude was doing across all of them. So I built Gatekeeper — a TUI daemon that intercepts every Claude Code tool call and routes it to one unified approval dashboard. The dashboard Three panes, one terminal: Left — all active Claude sessions, with status badges: [auto] means auto-approve is on, [linked] means it's wired to a terminal window Middle — pending permission requests with an age timer so you know what's been waiting longest Right — full request detail, danger warnings, and the numbered approval menu Every Claude Code tool call — Bash , Edit , Write , Agent — passes through a PreToolUse hook before executing. The hook connects to Gatekeeper's Unix socket, sends the request, and blocks. Gatekeeper shows it in the UI. When you decide, the answer travels back and Claude proceeds or stops. Approving requests The menu in the right pane mirrors Claude Code's own style: 1 Allow once 2 Always allow 3 Deny ↑ / ↓ moves the cursor, Enter confirms. Or just press 1 , 2 , 3 directly. A and D are quick shortcuts for allow/deny. Option 2 — always allow — is where it gets useful. Choosing it saves a persistent rule so the same request never surfaces again: Bash → saves the command pattern (e.g. npm run * ) to config Edit / Write → saves the directory to an allowlist Agent → enables auto-approve for that session The rule is written both to Gatekeeper's own config and to Claude Code's settings.json allowlist — so Claude Code itself won't prompt for it either. Auto-approve sessions Press A in the Sessions pane to mark a session as trusted. It shows [auto] — routine tool calls pass silently without appearing in the queue. But some things

A post by Ben Halpern

The company says opting out won't impact placement in regular searches.

With shiftbloom studio. I build tools and projects about a variety of experimental approaches to real-world problems. The issue for such use-case often was how most small media systems start out: one big always-on recorder that keeps costing money even when nothing is happening. For live capture you obviously need to stay ready at all times — sometimes you can’t risk losing the first minutes. But for everything else it’s complete overkill. The Core Problem Backfills, VOD downloads, clip imports, repairs and re-encodes are queue work. They can wait a few seconds, run on burst capacity, or even on a regular VPS or laptop. They don’t need the same always-hot infrastructure as the live recorder. That’s why I split the system. Instead of one large monolith, I deployed: Observer cells — only for live streams (time-critical) Harvest cells — for all queue processing (can be delayed) The Three Roles 1. Mothership A small control-plane cron job. It checks queue sizes, currently live channels and running observer tasks, then decides: how many harvest cells should exist right now which channels need an observer cell It’s intentionally simple. The database remains the single source of truth. 2. Observer Cells Each observer cell records exactly one live channel. It receives its assignment through environment variables: +++env OBSERVER_VOD_ID OBSERVER_CHANNEL_ID OBSERVER_CHANNEL_LOGIN OBSERVER_CHANNEL_NAME +++ It starts recording immediately, writes HLS segments to object storage, sends heartbeats, and waits a short standby window after the stream goes offline. This window is important because streams sometimes drop and reconnect quickly. Without it you end up with many small broken VOD fragments. 3. Harvest Cells These handle all background work: downloading VODs, re-encoding, recovering broken files, etc. They can run anywhere Docker is available — AWS tasks, a small VPS, or even a spare laptop. They only need outbound access to Postgres and object storage. What Changed Previous

pluckmd exists so an agent can pull blog posts into markdown, index them into a wiki, and generate interactive HTML to learn from. This post is about the first step, the part with no per-site code, because the design is the interesting bit. If you want the practical side, how I actually use it day to day, I wrote that up separately: https://dev.to/taisei_ide/how-i-use-pluckmd-to-read-blogs-with-an-ai-agent-1jpe It downloads articles from a blog without any per-site code. No handler for Medium, no handler for Substack, nothing keyed on a domain. Here's how that works. The core idea: treat extraction as data, not code. AdapterSpec Instead of branching on which site you're on, pluckmd resolves an AdapterSpec . It's a plain object that says which selector finds article links, what the URL pattern looks like, and how pagination behaves. interface AdapterSpec { listing : ListingExtractionSpec ; // how to find article links article : ArticleExtractionSpec ; // how to pull the body pagination : PaginationSpec ; // none | scroll | button-click | next-url | auto evidence : string ; } Because it's data, the same shape can come from a heuristic, an LLM, an agent, or a person typing it by hand. They all produce the same thing, and they all go through the same checks. Resolving it, cheapest path first cache -> heuristics (local, free) -> LLM (only if needed) Cache first, rechecked against today's DOM so a stale entry can't sneak through. Then local heuristics. The LLM only gets called when the heuristics aren't sure. Every result that works gets written back, so the second run on a site is basically instant. How the heuristics find an article list This part has no idea what site it's looking at. It takes every link, normalizes the path, and collapses the parts that vary into wildcards. / blog / my - first - post -> / blog /* / blog / another - article -> / blog /* / about -> / about Group by that shape. Any group with the same pattern repeated three or more times is a candidate f

The Setup I've been building AI agents that use tools — reading files, running commands, calling APIs. There are two main ways to give agents these tools: MCP (Model Context Protocol) — the new standard everyone's adopting Direct CLI calls — good old command-line execution Everyone says MCP is the future. But nobody talks about the token cost . So I measured it. The Test I built a simple file-reading tool and measured the exact token consumption for each approach: Method Tokens per Call Latency (avg) MCP (structured) ~3,400 tokens 280ms CLI + raw output ~200 tokens 45ms Ratio 17x 6x Why MCP Uses So Many Tokens The overhead comes from three places: 1. Tool Schema in Every Request MCP sends the full JSON Schema of every available tool with each request to the LLM. My simple file-reader schema alone is ~800 tokens. With 10+ tools, that's 8,000+ tokens of schema on every single call. { "name" : "read_file" , "description" : "Read contents of a file at given path" , "parameters" : { "type" : "object" , "properties" : { "path" : { "type" : "string" , "description" : "File path to read" } }, "required" : [ "path" ] } } 2. Structured Response Wrapping MCP wraps every response in a structured envelope with metadata, status codes, and typed content blocks. A simple "file not found" error becomes a 200-token JSON object. 3. Round-Trip Protocol Overhead Each MCP call involves: request → server parse → execute → format response → return → client parse → extract. Each step adds tokens for protocol framing. The CLI Alternative With direct CLI execution: $ cat /path/to/file.txt [ raw file content] That's it. Raw input, raw output. No schemas, no envelopes, no metadata. When MCP Is Worth It Despite the token cost, MCP shines when: You need standardized discovery — agents dynamically finding available tools You're building reusable tool servers — one MCP server serves many agents Security sandboxing matters — MCP's permission model is more granular Team collaboration — shared tool de

Qwen3.7-Plus has appeared on Qwen's official research release page, with a release date of June 1, 2026. Chinese media covered the launch on June 2. The important part is not that Qwen 3.7 Plus can understand images. The bigger signal is that Qwen is pushing it as a multimodal agent model: vision, language, coding, tool use, and productivity workflows inside one task loop. For developers, the real question is simple: can it keep the same goal across software screens, web pages, screenshots, code, terminal output, and tool calls long enough to finish useful work? If your team is evaluating new agent models, keep the model shortlist in one place and compare quality, latency, cost, and failure modes by task: Compare AI models on WisGate . What Is Qwen3.7-Plus? Qwen3.7-Plus is a multimodal agent model from Qwen. Qwen describes it as an agent foundation that unifies vision and language. It builds on the Qwen3.7 text backbone, adds stronger vision-language capabilities, and keeps the agent-oriented strengths developers care about: coding, tool use, and productivity workflows. That makes it different from a basic image-question-answering model. The more useful use cases look like this: Read a UI screenshot and decide the next action. Combine web pages, docs, charts, screenshots, and text context. Turn a design or product screen into maintainable code. Use tools to verify results instead of only returning static answers. Move between GUI, CLI, browser, and code environments during one task. That is why Qwen3.7-Plus should be evaluated as an agent model first, not just as another chat model with vision support. Why This Release Matters More teams are moving models into longer workflows: read the request, inspect the code, run tests, check logs, fix the issue, verify again, and write the summary. The hard part is that real work is rarely text-only. Frontend bugs come with screenshots. Dashboards come with tables and charts. Debugging comes with terminal output, browser state,

The least interesting thing an AI coding agent can do now is generate code. That sounds harsher than I mean it. Generation still matters. Better models still matter. Faster edits still matter. But if you have used these tools on a real codebase, not a demo repo with three files and no history, you already know where the pain moved. The bottleneck is not "can the model write a React component?" The bottleneck is "does the agent understand why this repo is weird?" Real repos are full of weirdness. Naming conventions nobody wrote down. Migration leftovers. Feature flags with political history. Tests that exist because of one brutal production incident. API boundaries that look accidental until you remove them and break billing. A hundred tiny facts that separate a useful change from a confident mess. Coding agents are getting much better at editing files. The next stack has to get better at making the system legible before the edit starts. Bigger context windows are not the same as understanding The lazy answer is to throw more context at the model. Give it the whole repo. Add the README. Add the docs. Add the last five tickets. Add the architecture decision records. Add the transcript from the previous session. Add the test output. Add the package lock, because why not. That works until it does not. A larger context window can hold more text. It does not automatically turn that text into a map. It does not know which files are architectural boundaries and which are incidental wrappers. It does not know that one directory is deprecated unless the repo says so clearly. It does not know that a scary-looking validation branch is protecting a partner integration from 2021. More context can even make the problem worse. You get the pleasant illusion that the agent has seen everything, while the useful signal is buried under raw file dumps and old notes. Repo understanding needs structure. That is why tools that turn codebases into graphs, domain maps, guided tours, semantic

Hello! I'm Satoshi Kaneyasu, DevOps engineer at Serverworks. In this article, I'll introduce the basic concepts and terminology of vector databases for those who are just starting to learn about them. Target Audience This article is aimed at beginners to vector databases. You may have heard that vector databases are related to LLMs and RAG, but aren't quite sure what they actually are. Think of this as written with that kind of reader in mind. What Is a Vector Database? A vector database is a database that stores data as vectors (arrays of numbers) and searches for data using "distance" or "similarity" between vectors. Traditional relational databases search for data using "exact match" or "partial match" (LIKE queries), but vector databases can search for things that are semantically similar . For example, searching for "weather in Tokyo" might return results like "temperature in Tokyo" or "weather conditions in Kanto" — data that differs as a string but is semantically related. Visualizing Vector Space In a vector database, all data is represented as points in a multidimensional space. When searching, the query is also converted into a vector, and data that is "close in distance" within that space is retrieved. This diagram represents it in two dimensions, but in a real vector database, proximity and distance are defined across many dimensions. Use Cases for Vector Databases Vector databases are used across a wide range of applications: Use Case Description RAG (Retrieval-Augmented Generation) Knowledge base search to provide external knowledge to LLMs. Allows internal documents and up-to-date information to be reflected in LLM responses Semantic Search Searching internal documents or FAQs by meaning rather than keywords. Handles spelling variations and synonyms Recommendation Recommending products and content whose vectors are close to a user's preference vector. Used as an alternative or complement to collaborative filtering Image Search Searching for similar im

A few months ago, I built an AI agent to automate one of the most repetitive parts of my workflow:...

Note: This post is a raw development log storing the exact prompts and responses used with the local LLM (IBM Granite-3.2-8B) to build the autonomous racing agent. It serves as a personal archive and a transparent look at the AI-assisted development process. Development Log: Autonomous TORCS Racing Agent Model used: ibm-granite-3.2-8b-instruct This log documents all the prompts and answers used in IBM Granite models for this project. 1. Initial Model Creation Goal: Create a robust, modular control system for the Corkscrew track. Prompt @snakeoil3_jm2.py @torcs_jm_par.py Act as an expert in autonomous vehicle control and TORCS simulator architecture. I am developing a high-performance agent for the Corkscrew track. Analyze the provided snakeoil3_jm2.py (for sensor/action interface) and torcs_jm_par.py (current control logic). I need you to rewrite the drive() function to be a robust, modular control system. Please provide: 1. Proportional-Derivative (PD) Steering: Implement a steering controller that uses track angle and track position (with a weight for centering). 2. Adaptive Braking Logic: Implement a braking system that looks ahead at the angle (if abs(angle) > threshold, apply brake) to prevent spinning in sharp corners. 3. Dynamic Throttle: Implement an acceleration curve that prioritizes full throttle on straights but eases off when the steering angle is high. 4. Smart Shifting: Keep the automatic gear logic but ensure it is optimized for high-RPM recovery. Provide the complete, clean Python function drive(c) ready to replace the existing one in torcs_jm_par.py. Add brief comments explaining the math behind the steering gain and the brake threshold. Response Here's a revised drive() function for torcs_jm_par.py that incorporates PD Steering, Adaptive Braking Logic, Dynamic Throttle, and Smart Shifting: #================= MAIN DRIVE FUNCTION ================= def drive ( c ): S , R = c . S . d , c . R . d # PD STEERING CONTROL steer_kp = 30 # Proportional Gain

The Model Comparison Problem Most AI model comparisons are useless for developers making real decisions. They benchmark on academic datasets that don't reflect production workloads. They test frontier capabilities that matter for 5% of use cases. They ignore latency, cost, rate limits, and API reliability — which are the things that actually determine whether a model works in your application. This comparison is different. It's focused on what matters when you're building something: how Grok and Gemini perform on the types of tasks developers actually encounter, what each model's API experience is like, and where the genuine tradeoffs lie. I'm deliberately not including benchmark scores. If you want MMLU numbers, there are plenty of leaderboards for that. This is about production utility. What Each Model Actually Is Grok (xAI) Grok is xAI's model family. The current production models are Grok-3 and Grok-3 Mini, with Grok-3 being the flagship. Grok has a large context window (128K tokens standard, with extended context available), real-time access to X (Twitter) data as a differentiating feature, and strong performance on reasoning-heavy tasks. The xAI API follows a familiar REST pattern and is broadly compatible with OpenAI SDK conventions, which makes migration straightforward. Grok's notable characteristics: Strong at structured reasoning and multi-step problem decomposition Real-time web access via the API (useful for tasks needing current information) Relatively generous rate limits compared to some competitors Less restrictive on certain content categories than some other models Gemini (Google DeepMind) Gemini is Google's model family, currently anchored by Gemini 1.5 Pro and Gemini 2.0 Flash. The defining feature of Gemini is its context window — Gemini 1.5 Pro supports up to 1 million tokens in production, which is genuinely useful for certain document-heavy use cases. Gemini also has the tightest integration with Google's ecosystem (Workspace, Cloud, Search)

Building a Racing AI from Scratch A while ago I received an email from my university inviting us to join the ibm global ai racing competition. Now that I'm finished with my exams I am going to give it a try. The testing ground for this project will be TORCS (The Open Racing Car Simulator). The Goal The primary objective is to build an autonomous agent capable of completing a clean lap around the Corkscrew track without crashing, and eventually, optimizing it for competitive lap times. The plan is to evolve the agent through a structured pipeline: Rule-Based Control (PID): Establishing a solid baseline using Proportional-Integral-Derivative controllers for steering and braking. Machine Learning: Upgrading the agent to learn from its environment using frameworks like PyTorch to replace hardcoded heuristics. Optimization: Fine-tuning the parameters and pushing the physics engine to the limit. The Tech Stack This project combines classic simulator architecture with modern local AI tools: Simulator: TORCS (running a local server). Language: Python (interfacing via the snakeoil3 library to parse sensor data and output telemetry). Local AI Assistant: ibm-granite-3.2-8b-instruct . I will be using this local LLM (hosted via LM Studio and integrated into VS Code with Continue.dev) to help architect the math, tune the control logic, and create/debug the Python code. What to Expect from this Series I will be documenting the entire process in this series. I will share the exact prompts used with the local AI, the generated code, the mathematical reasoning behind the control systems (such as why a naive PD controller causes zig-zag oscillation and how to fix it with damping), and the iterative debugging process. If you are interested in robotics, control theory, Python, or machine learning applications in simulation environments, follow along. The first technical log will be published shortly, detailing the implementation of baseline steering and look-ahead braking logic.

We’ve all been there: staring at a messy medicine cabinet, wondering which box is for allergies and which one expired in 2022. In the world of Computer Vision and AI Healthcare , digitizing physical assets is a classic challenge. Today, we're building a "Medicine Box Expert"—a sophisticated pipeline that uses YOLOv8 for precision detection and OpenAI CLIP for multimodal understanding to turn a pile of pills into a searchable digital database. By the end of this tutorial, you'll understand how to bridge the gap between raw pixels and structured medical data. We are moving beyond simple classification; we are building a robust system capable of handling complex lighting, varied angles, and the tiny typography common in pharmaceutical packaging. The Architecture: A Multi-Stage Vision Pipeline To achieve high accuracy, we don't rely on a single model. Instead, we use a "Detect-Extract-Embed" workflow. graph TD A[User Uploads Image] --> B[YOLOv8: Box Detection] B --> C{Box Found?} C -- Yes --> D[Crop & Preprocess] C -- No --> E[Error: No Box Detected] D --> F[Tesseract OCR: Text Extraction] D --> G[OpenAI CLIP: Visual Embedding] F & G --> H[SQLite Query: Semantic Search] H --> I[Result: Drug Info & Dosage] Prerequisites Before we dive into the code, ensure you have the following tech_stack installed: YOLOv8 : For real-time object detection. OpenAI CLIP : To handle semantic image-text matching. Tesseract OCR : For reading the fine print on the boxes. SQLite : To store and query our medicine metadata. pip install ultralytics transformers torch pytesseract Step 1: Detecting the Medicine Box with YOLOv8 First, we need to locate the medicine box within the frame. A generic YOLOv8 model (like yolov8n.pt ) is surprisingly good at detecting "books" or "cell phones," but for the best results, you should fine-tune it on the Open Images Dataset specifically for "Box" or "Medical Packaging." from ultralytics import YOLO import cv2 # Load the model model = YOLO ( ' yolov8n.pt ' ) def

Just released an open-source bash checker for CIFSwitch (CVE-2026-46243) — the 19-year-old Linux kernel LPE disclosed last week that lets any unprivileged local user get root by abusing the CIFS/SPNEGO upcall path. The script runs on bare-metal, VMs, and inside containers, and is CI/CD-friendly with JSON output and clean exit codes. It checks: ✅ Kernel version against patched thresholds (6.18.22 / 6.19.12 / 7.0+) ✅ cifs-utils presence and exploitable version ✅ CIFS kernel module load state and blacklist status ✅ Unprivileged user namespace sysctl (the pivot point for the exploit) ✅ Active request-key cifs.spnego rules ✅ SELinux / AppArmor enforcement ✅ Container capabilities (CAP_SYS_ADMIN) ✅ Kernel symbol verification for the fix commit Outputs human-readable or JSON for SIEM ingestion. Exit 0 = safe, exit 1 = action needed — drop it straight into a pipeline. CIFSwitch is the fourth Linux LPE in under six weeks (after Copy Fail, Dirty Frag, and Fragnesia). If you're running multi-tenant Linux, CI runners, or container build farms, now is a good time to audit. I have also updated the cve_checks.conf in my my K8s-container_escape_audit toolkit to detect this issue.

Microsoft launched seven new in-house AI models at Build 2026 on June 2, 2026, marking the company's most significant push yet to build its own frontier AI stack independent of OpenAI. The centerpiece is MAI-Thinking-1, Microsoft's first large-scale reasoning model, built from scratch on clean commercially licensed data using a sparse Mixture of Experts architecture. Alongside it: MAI-Code-1-Flash, a 5-billion-parameter coding model that outperforms Claude Haiku 4.5 by 16 percentage points on SWE-Bench Pro while using 60% fewer tokens on complex tasks. This is the complete developer guide to all seven MAI models, their specs, benchmarks, deployment paths, and what they mean for the AI development ecosystem. Why Seven Models at Once? The strategic context matters. For three years, Microsoft's AI product surface — GitHub Copilot, Azure AI, Bing Chat, Microsoft 365 Copilot — ran almost entirely on OpenAI models. The Build 2026 announcement is Microsoft's public declaration that it is building a parallel, proprietary model stack. Every new MAI model is trained from scratch using "clean and appropriately licensed data, without distillation from third-party models" — language that directly addresses the intellectual property concerns that have accompanied third-party model licensing. The distribution strategy is equally deliberate. Microsoft is not routing MAI models exclusively through Azure. MAI-Thinking-1 and MAI-Code-1-Flash are available via Fireworks AI, Baseten, and OpenRouter — three infrastructure providers that collectively reach developers who explicitly do not want cloud vendor lock-in. This signals a platform-first posture: Microsoft wants MAI to become a model ecosystem, not just an Azure feature. MAI-Thinking-1: The Reasoning Flagship MAI-Thinking-1 is Microsoft's answer to Claude Opus 4.x and GPT-5.5 on the reasoning side of the model spectrum. The architecture is a 35-billion-parameter active / approximately 1-trillion-parameter total sparse Mixture of Ex