<p><strong>TLDR:</strong> Cursor is taking heat after someone noticed Composer 2’s base model ID literally says <code>kimi-k2p5</code> — a $50B company shipping another lab’s model with no attribution. Anthropic meanwhile pushed Claude Code into Telegram and Discord via MCP, going after the OpenClaw use case directly. Qwen dropped Qwen3-Coder-Next, a 3B active parameter model claiming to beat much larger models on SWE-Bench-Pro. GitHub Copilot quietly pulled Claude Opus and GPT-5.4 from its student plan, and students are not happy.</p>
<hr>
<h2>Cursor’s Composer 2 base model controversy</h2>
<p>Someone checked the model ID behind Cursor’s Composer 2 and found <code>kimi-k2p5</code> — Kimi K2.5 with reinforcement learning, unattributed. A Cursor co-founder acknowledged it: “it was a miss to not mention the Kimi base.” That’s a significant understatement for a company valued at $50B that has positioned itself as an AI-native coding platform. The community reaction is split between “this is fine, they did domain-specific RL” and “this is a UI on top of someone else’s engine with no disclosure.” Composer 2 still took second place on the Next.js evals leaderboard, beating Claude Opus and Gemini, so the underlying capability is real — the transparency problem is separate from the performance question.</p>
<h2>Claude Code gets Telegram and Discord support</h2>
<p>Anthropic shipped Claude Code Channels, letting you connect Claude Code to Telegram and Discord via MCP. The setup requires Claude Code v2.1.80+ and Bun, and it works as a two-way bridge that injects messages into active Claude Code sessions. This is a direct shot at OpenClaw, which built a following largely because it let you run persistent agents you could ping from your phone. Early hands-on reports say the remote control works but feels rough — the agent often goes silent mid-task, leaving you wondering if it’s still running. The Telegram and Discord plugins are open-sourced on GitHub.</p>
<h2>Qwen3-Coder-Next and the Qwen Code CLI</h2>
<p>Qwen dropped Qwen3-Coder-Next, claiming it beats models with 10-20x more parameters on SWE-Bench-Pro at only 3B active parameters. They also launched Qwen Code CLI as an open-source alternative to Claude Code CLI. No independent benchmark verification in the data yet, but if the SWE-Bench-Pro numbers hold up, this is the most interesting open-weights coding model release in a while. Worth watching before drawing conclusions.</p>
<h2>GitHub Copilot cuts student plan models</h2>
<p>GitHub Copilot removed Claude Opus, Claude Sonnet, and GPT-5.4 from its student plan. The announcement has over 5,000 dislikes. Students were saving over $100/month through the plan — losing access to frontier models mid-semester is a real disruption, not just a pricing complaint.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Opus 4.6 found 22 Firefox vulnerabilities</strong> in two weeks, 14 of them high-severity, representing roughly 20% of all critical Firefox bugs fixed in 2025. Anthropic says it generated working exploits for some.</li>
<li><strong>OpenAI’s Responses API now has a container pool</strong> that makes container spin-up for agent workflows roughly 10x faster by reusing warm infrastructure instead of creating new containers each session.</li>
<li><strong>Karpathy’s home automation agent</strong> — he built an agent called “Dobby” that reverse-engineered APIs for Sonos, HVAC, shades, pool, and security system, then created a natural language interface. “Sleepy time” triggers home-wide automations.</li>
<li><strong>T3 Code uses 350.9 MB RAM vs Claude Code CLI’s 635.5 MB</strong> — notable because T3 Code is an Electron app and Claude Code CLI is built with Bun.</li>
<li><strong>OpenAI is giving US and Canadian college students $100 in Codex credits</strong> through a program called Codex for Students.</li>
<li><strong>Appwrite Arena benchmark</strong> tests LLMs across 191 questions on Appwrite’s APIs — GPT-5.4 leads with skills enabled, Claude Opus 4.6 leads without.</li>
<li><strong>OpenClaw 2.0</strong> shipped hybrid BM25 + vector memory search, nested sub-agents with configurable depth, Dashboard v2, iOS/Android/Apple Watch support, and Docker/Kubernetes health endpoints.</li>
<li><strong>Comprehension debt</strong> is the framing getting traction: AI coding shifts the problem from technical debt to code that works but nobody on the team actually understands.</li>
<li><strong>A mechanical engineer with no coding background</strong> built a piping isometric drawing parser with Claude Code in 8 weeks — reduces 10-minute-per-drawing work to 60 seconds, now used daily by his fabrication shop.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

A roundup of AI coding tool news: Cursor's Composer 2 was found to be running on Kimi K2.5 without attribution, sparking community debate about transparency. Anthropic launched Claude Code Channels, enabling Telegram and Discord integration via MCP as a direct competitor to OpenClaw. Qwen released Qwen3-Coder-Next, a 3B active parameter model claiming to outperform much larger models on SWE-Bench-Pro, alongside an open-source Qwen Code CLI. GitHub Copilot removed Claude Opus and GPT-5.4 from its student plan, drawing over 5,000 dislikes. Additional notes cover Claude Opus 4.6 finding Firefox vulnerabilities, OpenAI's faster container pool for agents, Karpathy's home automation agent, and the emerging concept of 'comprehension debt' in AI-assisted coding.

Cursor's Composer 2 runs on Kimi K2.5, Claude Code lands on Telegram and Discord