Skip to content
DigIntoAI/Model intelligence
Model intelligence feed

Model intelligence.

Follow the changes. Understand what comes next.

Editor’s pickOpenRouter
OpenRouter/OpenRouter; MiniMax

Why AI Router ZDR Labels Are Not a Guarantee: Inside OpenRouter's MiniMax Warning and a 4-Step Hardening Guide

Following Dragos's November 19, 2026 study of two million canary-seeded prompts, OpenRouter emailed affected customers on November 20–21 warning that prompts and outputs sent to MiniMax models from July to October 2026 may have been read by a third party—whether served directly or forwarded via AtlasCloud, Novita, or StreamLake—and removed ZDR labels from affected endpoints pending investigation.

Latest stories126

Newest published
25

ZONOS3-Mini Open Weights: How to Run Text-to-Speech in the Browser and on Local Devices

Zyphra has released ZONOS3-Mini, an open-weight multilingual text-to-speech model supporting 43 languages including Chinese, English, and Japanese. Combining 100M active parameters (800M total) with a 15 Hz dual audio codec, ZONOS3-Mini provides Apache 2.0 weights, MIT-licensed runtimes, and four deployment paths spanning in-browser WebGPU, Apple Silicon MLX, C++/ggml, and NVIDIA CUDA. Here is how the stack works, what hardware it requires, and how its cost structure compares with cloud TTS APIs.

ZONOS3-MiniRead story
26

Gemini 2026 Subscription Tiers, Usage Multipliers & Compute Metering Explained

Since May 2026, Google has replaced fixed daily prompt counts on consumer Gemini plans with compute-based metering governed by a 5-hour rolling refresh and a weekly cap. This guide breaks down the exact multipliers, per-unit cost math, compounding context drain, file upload limits, and cross-client eligibility rules across Free, Google AI Plus ($4.99), Pro ($19.99), and Ultra ($99.99 / $199.99).

GeminiRead story
27

Claude Code Ships v2.1.323 Stable with Experimental Code Mode and a Breaking 1M Context Default

Anthropic published Claude Code v2.1.323 (commit 89eb88e) as the latest stable release on GitHub and its official changelog (dated November 12, 2026 PT / morning of November 13 Beijing time),…

Read story
28

Qwen Code ships v0.27.2 stable and v1.0.0-preview with Qoder integration and standalone installers

On November 12, 2026, the Qwen team pushed two back-to-back releases to the QwenLM/qwen-code GitHub repository: the latest stable build v0.27.2 (commit f42fe0f, bundled with TypeScript SDK…

Qwen CodeRead story
29

A2UI 1.0 gives AI agents interactive interfaces, but platform support still varies

Google's first stable A2UI release lets agents present forms, cards and other controls using an app's own components. Here is what changes in 1.0, how it fits with MCP Apps, and what developers should check before adopting it.

A2UI 1.0Read story
30

Claude Sheets Guide: Plans, Imports, Exports and Beta Limits

Claude Sheets brings shared spreadsheets into Claude on Pro, Max, Ultra, Team and Enterprise. It supports CSV, Google Sheets and Excel imports, XLSX and CSV exports, formulas and charts. Usage counts toward existing plan limits; mobile access is basic read-only.

Claude SheetsRead story
31

Claude Code Says "Prompt Is Too Long"? How to Fix It (November 2026)

Prompt is too long means the conversation plus attached files exceeded the context window, not a quota or 529. Usual line: Context limit reached · /compact or /clear to continue; if auto-compact is off, it names /config. -p: Prompt is too long. Bedrock: Input is too long for requested model. Gateway: capability_rejected: prompt_too_long. Try /compact or /clear (project memory remains). /context shows usage. /mcp disable <name> disables unused MCP; trim large CLAUDE.md. Auto-compact defaults on. A single exchange cannot be compacted; session limits, 529, and Request too large (max 32MB) differ.

Claude CodeRead story
32

How to Set Up Claude IP Allowlisting and Fix 403 permission_error

Claude IP allowlisting (public beta) lets org admins or owners allow only approved public IPv4/IPv6 addresses and CIDR ranges. Claude Platform (Console) and Claude Enterprise (claude.ai) are separate control planes with different entries, roles, and coverage. Outside the list, covered API calls can return 403 permission_error; people may also see that the current network cannot access the organization. Changes can take up to 15 minutes. Setup, formats, lockout prevention, and a 403 checklist follow Anthropic's docs.

Claude IP allowlistRead story
33

How to Use Claude API Message Threads (Beta) for Long Agent Loops

Claude Messages API message threads (beta) let long agent loops send only new messages after the first turn. Create once with the full conversation, system, and tools under beta header message-threads-2026-08-12; continue with new messages plus the previous response id. Threads expire after 5 minutes of idle time and are deleted within an hour. Keep your own history for replay. Pricing and rate limits stay the same.

Claude Messages APIRead story
34

How to Set Up MCP Servers in Claude Code (November 2026)

Start with the Claude Code documentation MCP server, which needs no separate sign-in. Add it from your terminal, check its status, then request a tool call in a session. HTTP suits most hosted services; stdio starts a local process. Local, project, and user scopes determine who gets the configuration and where it loads. Authentication and startup failures need different fixes: an Added message alone does not mean a server connected.

Claude CodeRead story
35

GitHub Copilot vs Cursor—Budget, Limits, and Which Workflow Fits (November 2026)

GitHub Copilot plugs into GitHub and many IDEs and bills chat/agent work in monthly AI Credits; Cursor is a standalone AI editor/agent inside The X Subscription, with a weekly pool for its own models and prepaid third-party usage from Premium up. Personal ladders (USD, ex-tax, Nov 2026): Copilot Free/Student → Pro $10 (1,500 credits) → Pro+ $39 (7,000) → Max $100 (20,000); Cursor Free → Plus $8 → Premium $30 (~4×) → Super $100 (~15×) → Ultra $200 (~50×). Compares budget, metering, workflow, and fit—not a performance bake-off. For reference only.

GitHub Copilot;Cursor(The X Subscription)Read story
36

Mistral Studio Managed Agents: What Stays on Mistral When You Self-Host Workers

Mistral Studio’s Managed Agents and Agents SDK are in public preview. With agents serve, tools and hooks run on a self-hosted worker, but Mistral’s control plane still stores sessions and serves model calls. Studio tracing is on by default and includes prompt and completion content.

Mistral Studio/Managed Agents/Agents SDKRead story