<p><strong>TLDR:</strong> Cursor is in talks to raise $2B at a $50B valuation after hitting $2B ARR — the fastest B2B software scaling on record. Anthropic launched Claude Design, a conversational design canvas built on Opus 4.7, and early users burned through their usage limits in under an hour. OpenAI’s Codex is evolving toward a full agentic IDE while drawing complaints about 30-second prompt latency. Opus 4.7 is out and costs 26–75% more tokens than 4.6 for identical inputs.</p>
<hr>
<h2>Cursor in talks to raise $2B at $50B valuation</h2>
<p>Anysphere is reportedly in talks for a $2B round co-led by a16z, Thrive Capital, and Nvidia, nearly doubling its November 2025 valuation of $29.3B. The company hit $2B ARR with over 1 million paying customers and 70% of the Fortune 1,000 on the platform. The $50B price tag implies a 25x revenue multiple — which compresses fast if the $6B ARR projection for end of 2026 holds. Competition from GitHub Copilot (37% market share, 4.7M paid subscribers), Claude Code, and hyperscaler offerings is real, and the commoditization risk is not theoretical.</p>
<h2>Claude Design launches, usage limits hit immediately</h2>
<p>Anthropic Labs shipped Claude Design, a conversational canvas for prototypes, slides, and one-pagers powered by Opus 4.7. The standout feature is a handoff bundle that packages designs directly for Claude Code. Early feedback is genuinely positive on video creation and design system integration — and immediately critical on usage limits, with at least one user exhausted in an hour. There’s also a reasonable argument that Claude Design and Claude Cowork are the same model with a new UI, optimized to push users toward higher subscription tiers. Claude Code already covers most of this with more control.</p>
<h2>Opus 4.7 token costs are 26–75% higher than 4.6</h2>
<p>Community data across 50 submissions shows Opus 4.7 consistently uses more tokens than 4.6 for identical inputs, with increases ranging from roughly 26% to 75%. Cost differences track token counts exactly. The model does show real improvements — SWE-bench Pro up from 53.4% to 64.3%, visual reasoning up from 69.1% to 82.1% — but the system prompt diff between 4.6 and 4.7 reveals the model is now more action-oriented and concise by design, which makes the token increase harder to explain. Claude Code also added a <code>/usage</code> command that breaks down where tokens are actually going.</p>
<h2>OpenAI leadership takes a hit, Codex performance degrades</h2>
<p>CPO Kevin Weil, Sora head Bill Peebles, and enterprise CTO Srinivas Narayanan all left on the same day. OpenAI is shutting down Sora (April 26) and winding down OpenAI for Science as it pivots toward enterprise revenue. Meanwhile, Codex is drawing complaints: prompt latency is reportedly 30 seconds or more, long threads cause text to disappear, and at least one developer clocked T3 Code as roughly 10x faster. The cron-job-as-sidebar-chat UI decision is also getting roasted. The agentic IDE ambition is real, but the current product has rough edges.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Anthropic’s Pentagon standoff:</strong> Dario Amodei met White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent over Mythos access. The dispute started when the Pentagon blacklisted Anthropic after it refused unfettered military access. A potential compromise would route Mythos through civilian agencies only.</li>
<li><strong>Anthropic banning paying orgs via automation:</strong> At least one 60-person paying organization was banned without justification, through an automated system with no human review or appeal path. Anthropic got it wrong. This is a real trust problem for teams building on their APIs.</li>
<li><strong>Claude Code policy vacuum:</strong> Builders have been asking for months what third-party integrations are permitted under Claude Code subscriptions. Anthropic keeps saying “we’re working on it.” Dozens of developers are in limbo.</li>
<li><strong>Qwen 3.5 9B + DSPy.RLM hits SOTA on LongCoT:</strong> A fine-tuned Qwen 3.5 9B reached 15.69% on LongCoT-full — roughly 1.6x GPT 5.2’s 9.83% on the same benchmark. Small model, large margin.</li>
<li><strong>Qwen 3.6 community fine-tune:</strong> A Qwen 3.6 35B fine-tuned on Opus 4.6 distilled CoT data is circulating, following the same recipe as a popular Qwen 3.5 27B fine-tune. Community evals welcome.</li>
<li><strong>OpenCode 1.4.11 ships workspace support:</strong> Sessions can now run in git worktrees and remote environments. The Session Graph plugin (community-built) visualizes how chat branches connect when you fork sessions.</li>
<li><strong>Claude Code v2.1.113 ships native binary:</strong> The npm package now ships the native binary instead of the JavaScript build. Same install command.</li>
<li><strong>Anthropic Claude Security feature spotted:</strong> Looks like Anthropic is preparing to expand its Claude Code security scanning feature — periodic repo scans for vulnerabilities — beyond Enterprise and Team plans, possibly to Max tier.</li>
<li><strong>Mozilla MZLA launches Thunderbolt:</strong> An open-source, self-hostable enterprise AI client supporting MCP and ACP, built on deepset’s Haystack. Not production-ready yet — security audit is incomplete. The name collision with Intel’s Thunderbolt port is going to cause confusion.</li>
<li><strong>ServiceNow at 90% AI case resolution:</strong> CEO reports AI agents now handle 90% of customer service cases, with net new headcount growth expected to drop sharply as agents absorb work in finance and HR.</li>
<li><strong>Gemma 4 runs locally on consumer hardware:</strong> Google’s open-weight model family (E2B through 31B) runs on an RTX 3070 at 40–50 tokens per second. The smaller models run on phones via AI Edge Gallery, offline, no API keys.</li>
<li><strong>HyperFrames by HeyGen:</strong> Open-source video rendering framework that lets agents author HTML-defined video compositions and render to MP4. Supports Claude Code, Cursor, Codex, and Gemini CLI via installable skills.</li>
<li><strong>Anthropic token cost calculator:</strong> Community-sourced tool comparing Opus 4.6 vs 4.7 request costs across 50 submissions. All entries show 4.7 costs more. Worth checking before migrating workloads.</li>
<li><strong>OpenAI GPT-OSS open-weight model:</strong> A 20B parameter model runnable locally with MXFP4 quantization. Framed as generosity; more accurately described as a response to Meta’s Llama pressure.</li>
</ul>


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

Agentic Digest

Cursor is reportedly raising $2B at a $50B valuation after reaching $2B ARR with over 1 million paying customers, making it the fastest-scaling B2B software on record. Anthropic launched Claude Design, a conversational design canvas powered by Opus 4.7, but early users hit usage limits within an hour. Community data shows Opus 4.7 uses 26–75% more tokens than 4.6 for identical inputs despite performance improvements. OpenAI's Codex is facing complaints about 30-second prompt latency and UI issues, while three senior OpenAI leaders departed on the same day. Additional notes cover Anthropic's Pentagon dispute, automated org bans, Qwen fine-tunes, and several open-source AI tool releases.

Cursor eyes $50B raise, Claude Design ships and immediately hits usage limits