#ai-development

ChatGPT plugins can now add extensions such as side panels and file viewers, and support proposed MCP Events. What changed for developers building plugins.

OpenAI's Agents API now runs computer use in a hosted browser, with website approvals and app-handled sign-in. How it works and what you can build with it.

Tests check a few inputs; a proof covers all of them. Bend, Verus, Dafny and Lean compared on what you write, what they prove, and how an agent proves first.

Anthropic says an AI agent made 30+ scientific models about 4x faster in four weeks, overseen by two staff new to kernel work. How to run it yourself.

Assign responsibility for an AI-built app after handover. Define who fixes failures, updates dependencies, pays ongoing costs and proves recovery works.

Record the services, access owners, scheduled jobs and recovery steps an AI-built app needs. Use a practical reference before taking over its operation.

Specify the files an AI-generated 3D asset must include. Match editable scenes, meshes, textures and animation data to the changes your team needs to make.

What nine AI coding CLIs do before you choose anything: trust prompts, telemetry defaults, and mid-task keystrokes, pinned to sources read on August 24, 2026.

Agent-written tables can state a correct formula and still print cells that break it. Why fact-checking misses this, and a recompute pass that catches it.

Identical dollar-per-million pricing does not mean identical cost: tokenizers split the same text into different counts. A vocabulary reference table.

Seventeen coding agents, six questions each, vendor documentation only. What the published terms say on retention and training, and which cells stayed open.

Grok 4.6 reached Cursor the day it launched. The argument that switching cost has moved to the harness, set out with its counter-evidence alongside it.

Every vendor names its reasoning-effort settings differently and the rungs do not line up. What the ladders look like, and where thinking cannot be disabled.

Cross-model agreement is correlated evidence, not independent confirmation. Models share training data and share errors. A reconciliation method.

Tokens fix theming. Screenshots catch what tokens cannot express: spacing rhythm, hierarchy, density, alignment drift. Where vision models help.

Frontier models cluster on most benchmarks, so the brief does more work than the model pick. The anatomy of a spec that raises output quality.

DeepSeek V4 Flash exits preview into public beta as the 0731 checkpoint: same 284B architecture, vendor-stated agent benchmarks, no weights posted yet.

GitHub made Copilot code review's agent skills and MCP connections generally available on July 29, one day after the MCP spec shipped. A rollout checklist.

ARC Prize independently administered Claude Opus 5's 30.16% ARC-AGI-3 result. Almost every other launch-day benchmark number is vendor-run on a vendor harness.

DeepSeek retires its deepseek-chat and deepseek-reasoner API aliases on July 24, 2026 at 15:59 UTC. The confirmed migration steps, plus the claims to watch.

SpaceXAI open-sourced Grok Build's Rust harness under Apache 2.0 days after a privacy scandal. Why a public repo is not the same as a security audit.

OpenAI's GPT-5.6 caching overhaul, plus Anthropic and DeepSeek tiers, reshapes agent cost math. How to design cache-first, model-homogeneous agents.

A same-day LLM eval harness needs 20-50 real tasks, automated grading, and a baseline model to diff against — not hundreds of labels. Qualify a new model fast.

OpenAI retires 13 model snapshots on July 23, 2026, including every pre-5.3 Codex variant. Replacements aren't 1:1, so pinned configs need testing first.

Auto mode's classifier-gated permissions are now default on AWS Bedrock, Vertex AI, and Foundry. What enterprise teams should review before rollout.

Vercel now auto-detects Lovable's TanStack Start apps for zero-config deploys. A practical pipeline for taking vibe-coded prototypes to production.

ChatGPT Work and Codex share one GPT-5.6 usage pool, per OpenAI's own docs. Access splits by plan and surface - plus the CLI update you need.

GPT-5.6 GA, Grok 4.5, and Fable 5 all moved this week. A dated marketing routing matrix mapping model to task and cost-per-finished-task, not brand loyalty.

OpenAI's GPT-5.6 goes GA July 9 across ChatGPT, Codex and the API: official model IDs, ultra multi-agent mode, Programmatic Tool Calling, and access by plan.

Wire Google's official GA4 MCP server and a community Search Console server into Claude Code: service-account auth, read-only scopes and least-privilege grants.

A dual-model content review pairs one model as drafter and a second frontier model as adversarial critic, with human sign-off — grounded in 2026 bias research.

OpenAI's GPT-Live launched July 8 with full-duplex voice that listens while it talks and delegates hard questions to GPT-5.5. The developer API is signup-only.

Anthropic extended free Fable 5 access on paid plans to July 12, then July 19 — resolved July 18: Max and Team Premium keep access for good from July 20.

Tencent's Hy3 shipped Apache 2.0 at a sub-300GB FP8 footprint. A routing guide: send agentic tool-use to Hy3 and repo-scale coding to a frontier model.

Meta shipped Muse Image (Arena #2) and previewed Muse Video, its first in-house generation models, free in Meta AI and coming to Advantage+ creative.

Microsoft's Dataverse coding-agent plugin now runs in Claude, Cursor and GitHub Copilot, with one router enforcing existing RBAC across 15 auditable MCP tools.

xAI is now SpaceXAI: on July 6, 2026 Grok, X and Colossus folded into SpaceX under one ticker. For AI buyers it is an integration story, not a benchmark one.

Tencent shipped Hunyuan Hy3, a 295B open-weight MoE, under Apache 2.0 on July 6, 2026. It trails GLM-5.2 on coding but wins on agentic tasks at half the memory.

xAI shipped a no-code Voice Agent Builder on July 1, 2026: a plain-language prompt to a working phone agent in about two minutes, on one $0.05/min meter.

AI image generation statistics for 2026: Adobe Firefly nears $300M ARR, Midjourney tops $200M, 22B+ Firefly assets, plus the Getty and Andersen copyright cases.

AI music statistics for 2026: 44% of Deezer's daily uploads are AI, only 1-3% of streams, 85% flagged as fraud, plus Suno's $300M ARR and the licensing fights.

AI video generation statistics for 2026: Kling's $500M run-rate, 50.7% human detection accuracy, the real per-second API pricing floor, and market size decoded.

Agentic AI collapsed custom-build costs, so more firms win by owning tools than renting SaaS. A 2026 build vs buy AI decision matrix and when buying still wins.

91% of popular CLAUDE.md and AGENTS.md files carry config smells. Lint leakage, context bloat, and how to fix the files that make agents ignore your rules.

Anthropic's Claude Code Artifacts turn a coding session into a private, live, shareable HTML page. What they are, the 16 MiB CSP limits, and team uses.

Epoch AI found errors in 42% of FrontierMath problems and shipped v2 on June 12, 2026. Scores jumped, rankings held — here is what that means for model choice.

Route each request to the cheapest model that can handle it. Model routing cuts real LLM bills 40-85% with no visible quality loss. The 2026 engineering guide.

Meta, Google, and TikTok now ship official ads MCP servers. The capability differences, the paused-by-default guardrail, and a staged rollout playbook.

Snowflake CoCo executes SQL, dbt models, and pipelines end-to-end in the warehouse. The ADE-Bench results, the human-on-the-loop model, and the perimeter limit.

HITL escalation is the gap layer for production agents: evals and observability are covered, the handoff layer is not. The calibration math and async patterns.

GitHub's new desktop app turns Copilot into a control center for parallel agent sessions, each in an isolated worktree. Plus the GA Copilot SDK explained.

Technical guide to AI function calling across OpenAI, Anthropic, and Google APIs. Syntax comparison, implementation patterns, and best practices.

Xiaomi's MiMo-V2-Pro has 1T+ parameters with 42B active, ranking #8 worldwide. $1/$3 per million tokens. Complete model guide and deployment.

Perplexity launches Comet, an AI-native browser for iPhone with built-in research assistant and autonomous task completion. Setup and features.

Meta's Manus AI agent launches as a desktop app for local device integration. Direct AI agent access without cloud dependency. Full setup guide.

Visual ecosystem map of the AI agent protocol landscape: MCP (97M downloads), A2A (50+ partners), ACP, and UCP. How they connect and overlap.

GPT-5.4 Mini launches for free-tier ChatGPT users with 54.38% SWE-Bench Pro performance, only 3 points behind full GPT-5.4. 2x faster guide.

GPT-5.4 Nano at $0.20 per million input tokens is OpenAI's cheapest model. API-only, designed for classification, extraction, and subagents.

Mistral Forge enables enterprises to build custom AI models on proprietary data. Pre-training, post-training, and RL for agentic performance.

Anthropic launches Claude Dispatch for phone-to-desktop remote control. Pair mobile with Mac via QR code, send tasks from iPhone to desktop.

NVIDIA releases Dynamo 1.0, an open-source distributed inference OS achieving 7x performance boost for Blackwell GPUs. Adopted by major clouds.

GTC 2026 complete recap: Vera Rubin platform, NemoClaw enterprise agents, Nemotron Coalition, Dynamo 1.0, and the $1 trillion order pipeline.

Twelve AI models launched in a single week of March 2026 from OpenAI, Google, Mistral, xAI, and more. Developer guide to capabilities, pricing, and selection.

OpenAI's Codex Subagents reaches GA with manager agents coordinating specialized workers across repositories. Setup guide for multi-agent coding workflows.

OpenAI teased GPT-5.4 the same day as GPT-5.3 Instant launch. Rumored 2M context window, enhanced reasoning, and what it means for the AI model roadmap.

Build retrieval-augmented generation systems grounded in your company data. Vector databases, chunking strategies, evaluation, and deployment patterns.

Build production-ready streaming AI chat with Vercel AI SDK 6 and Next.js. Server Actions, structured output, tool calling, and multi-provider setup guide.

OpenAI retires GPT-4o, GPT-4.1, and o4-mini from ChatGPT on February 13, 2026. Complete migration guide to GPT-5.2 for businesses and developers.

Analyzing the Claude Sonnet 5 'Fennec' Vertex AI leak. What the leaked model ID, benchmarks, and pricing suggest about Anthropic's next major release.

MCP (Linux Foundation standard), LangChain, and CrewAI compared for AI agents. Architecture, integration patterns, and production readiness guide.

GPT-5.2-Codex achieves 56.4% SWE-Bench Pro and 64% Terminal-Bench. Learn Codex CLI setup, 400K context workflows, and cybersecurity use cases.

Gemini 3 Flash delivers 78% SWE-bench at $0.50/1M tokens—1/4 the cost of Pro. Developer guide covering 1M context, thinking levels, and API integration.

Master vibe coding with natural language programming. 82% of developers now use AI tools. Complete guide with best practices.

Moonshot AI's Kimi K2 Thinking achieves SOTA with 1T parameters, INT4 training, 200-300 tool calls. First open model competitive with GPT-5/Claude.

Build autonomous AI marketing teams with multi-agent orchestration. Learn MCP, ACP & A2A protocols. Complete 2025 guide to save 60% of your time.

GitHub Agent HQ unifies AI coding agents from Anthropic, OpenAI, Google, and more. Mission Control, enterprise governance, and MCP integration explained.

MiniMax M2 achieves 69.4 on SWE-bench at 8% of Claude's cost. Complete guide to China's open-source AI model for agents, coding & multimodal apps.

Compare v0, Lovable & Bolt AI app builders: $20-30/mo pricing, 12-min MVPs, full-stack capabilities. Balanced 3-way comparison to choose the right platform.

Review Claude Sonnet 4.5 launch-period features: Claude Code 2.0 checkpoints, subagents, Agent SDK, and context-management tools.

Qwen3 model-family guide refreshed with 2026 caveats for Qwen3.6-Max-Preview, Coder, Thinking models, and deployment strategies.

Historical 2025 comparison of Kimi K2, Qwen 3 Coder, and GLM 4.5 refreshed with April 2026 caveats for newer Chinese and frontier AI model selection.

Compare v0, Lovable, Bolt, and Replit with current pricing caveats, export paths, model-routing changes, security, and team-fit advice.

Claude Code guide updated for current Anthropic coding models, terminal workflows, enterprise controls, cost planning, and safe AI development.