<p><strong>TLDR:</strong> Kimi K2.6 dropped today as a 1T-parameter MoE open-source model with benchmark numbers that rival Claude Opus 4.6 and GPT-5.4, and it’s already shipping in Fireworks, OpenCode, and HuggingChat. GitHub Copilot paused new signups for Pro, Pro+, and Student plans after agentic workflows blew past the cost assumptions baked into flat-rate pricing. Anthropic quietly cut Claude Code’s prompt cache TTL from 1 hour to 5 minutes in early April, and some developers are only now noticing the token cost spike. The dev tools market continues its rapid reshuffle, with Claude Code and Cursor pulling ahead of Copilot in mindshare even as Copilot still holds the broadest install base.</p>
<hr>
<h2>Kimi K2.6 ships as a serious open-source coding agent</h2>
<p>Moonshot AI released Kimi K2.6, a 1T-parameter MoE model (32B activated) under a Modified MIT License. Benchmark numbers are strong: 80.2% on SWE-Bench Verified, 58.6% on SWE-Bench Pro, 89.6% on LiveCodeBench v6. The model supports a 256K context window, native INT4 quantization, and claims to orchestrate up to 300 parallel sub-agents. Moonshot also shipped Kimi 2.6 Code, a Claude Code-style terminal agent built specifically for K2.6. Real-world impressions are mixed — some developers find it genuinely competitive for agentic tasks, others say it falls apart on anything complex. Ethan Mollick’s take is worth noting: open-weight models tend to over-index on benchmarks relative to actual use, and K2.6 is probably no exception, though it’s still a good model.</p>
<h2>GitHub Copilot pauses new signups, tightens limits</h2>
<p>GitHub is pausing new signups for Copilot Pro, Pro+, and Student plans. The stated reason is service quality for existing users, but the underlying driver is straightforward: agentic workflows consume far more compute than the original flat-rate plans anticipated. Opus models are being removed from the cheaper Pro tier, with Opus 4.7 staying on Pro+ only — and Opus 4.7 carries a 7.5x request multiplier over Opus 4.6, making it effectively 250% more expensive. Existing subscribers unhappy with the changes can cancel and request an April refund by contacting support before May 20. Copilot Free remains open. This is the innovator’s dilemma playing out in real time: Copilot built its user base on subsidized flat-rate pricing, and now the economics of agentic usage are forcing a reckoning.</p>
<h2>Anthropic silently cut Claude Code’s cache TTL</h2>
<p>In early April, Anthropic changed Claude Code’s default prompt cache TTL from 1 hour to 5 minutes without any public announcement. One developer’s analysis of 1,140 sessions showed cache busts jumping from 39 to 199 per day, pushing daily costs from $6.28 to $15.54 — a projected $277 monthly increase. Anthropic’s defense is that shorter caches are cheaper for one-shot queries, which is true but irrelevant for sustained agentic sessions. You can check your own JSONL logs under <code>~/.claude/projects/</code> to see the impact. Setting <code>CLAUDE_CODE_DISABLE_1M_CONTEXT=1</code> caps context at 200K tokens as a partial workaround.</p>
<h2>Claude Desktop installs a browser bridge without asking</h2>
<p>A privacy consultant documented that Claude Desktop for macOS silently installs a Native Messaging manifest that pre-authorizes browser extensions across multiple Chromium-based browsers — including browsers not yet installed on the machine — without disclosure or user consent. The bridge grants access to authenticated sessions, DOM state, form fields, and screen capture, running outside the browser sandbox at user privilege level. Security researchers confirmed the technical claims are reproducible. Anthropic hasn’t responded publicly, and a formal regulatory complaint is being prepared under EU ePrivacy Directive Article 5(3).</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Cursor 3</strong> shifts the primary interaction model from file editing to orchestrating parallel agents across multiple repos, with a plugin marketplace for MCPs and subagents; 35% of Cursor’s own merged PRs are now written by autonomous cloud agents.</li>
<li><strong>Anthropic + Amazon</strong> expanded their deal: Amazon invested an additional $5B (total $13B), and Anthropic committed to over $100B in AWS spend over 10 years, securing up to 5 GW of compute capacity built around Trainium2 through Trainium4.</li>
<li><strong>Claude Opus 4.7 vs 4.6 benchmark</strong> from one developer: Opus 4.7 took 11m 10s and 200K tokens on a coding task vs Opus 4.6’s 6m 11s and 80K tokens — slower and more expensive, but caught more edge cases and bugs.</li>
<li><strong>MCP STDIO RCE</strong> — OX Security disclosed that MCP’s STDIO transport allows arbitrary OS command execution across ~200,000 exposed servers; Anthropic’s position is that this is expected behavior and client developers should implement their own filtering.</li>
<li><strong>Cloudflare Project Think</strong> introduces Fibers, a durable execution primitive that checkpoints its own instruction pointer rather than replaying from snapshots, alongside sub-agents, session trees, and sandboxed Dynamic Workers for agent-generated code.</li>
<li><strong>Google Antigravity bans</strong> — multiple reports of paying Antigravity customers getting banned for using third-party usage tracking tools, with bans cascading to Gemini CLI; affected developers are switching to Claude Code and Codex.</li>
<li><strong>Expo raised $45M Series B</strong> and launched Expo Agent in public beta, an AI tool built on Claude Code that takes a prompt and produces a production-ready iOS and Android app.</li>
<li><strong>OpenAI Codex Chronicle</strong> feature in testing: continuous screen capture to build memories, with noted risks including prompt injection from visited websites and rapid rate limit consumption.</li>
<li><strong>Lovable clarified</strong> that its public project visibility issue was a documentation failure, not a breach — chat messages on public projects were visible and have since been fixed, but code remains intentionally public.</li>
<li><strong>GitGuardian’s 2026 Secrets Sprawl report</strong> found 28.6M new secrets leaked in public GitHub repos last year (+34% YoY), with Claude Code co-authored commits leaking secrets at roughly 2x the baseline rate.</li>
<li><strong>OpenAI lost three senior executives in one day</strong> — CPO Kevin Weil, Sora creator Bill Peebles, and enterprise CTO Srinivas Narayanan — coinciding with Sora being shut down at roughly $1M/day in compute costs.</li>
<li><strong>Anthropic’s Claude Code cache TTL workaround</strong>: set <code>CLAUDE_CODE_DISABLE_1M_CONTEXT=1</code> to cap context at 200K tokens and reduce unnecessary cache busting in long agentic sessions.</li>
</ul>


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

Agentic Digest

Kimi K2.6, a 1T-parameter MoE open-source model from Moonshot AI, launched with strong benchmark scores (80.2% SWE-Bench Verified) and ships in Fireworks, OpenCode, and HuggingChat. GitHub Copilot paused new signups for Pro, Pro+, and Student plans as agentic workflows blew past flat-rate cost assumptions, with Opus models being tiered up. Anthropic silently cut Claude Code's prompt cache TTL from 1 hour to 5 minutes in April, causing significant cost spikes for developers in sustained agentic sessions. Claude Desktop was found to silently install a browser bridge granting broad access without user consent. Other notable items include Cursor 3's shift to parallel agent orchestration, a $5B Amazon-Anthropic deal expansion, an MCP STDIO RCE disclosure affecting ~200K servers, and OpenAI losing three senior executives in one day.

Kimi K2.6 lands as a real open-source coding agent, GitHub Copilot closes signups