<p><strong>TLDR:</strong> Anthropic shipped Claude Opus 4.7 today alongside Claude Design, a conversational design tool built in partnership with Canva. Opus 4.7 tops the Artificial Analysis Intelligence Index at 57 points, but the community is split — token costs are up 30-47% due to a new tokenizer, and the safety filters are catching legitimate developer workflows. Claude Design is generating real buzz, with Figma’s stock dropping 5% on the announcement.</p>
<hr>
<h2>Claude Opus 4.7 ships to a divided community</h2>
<p>Artificial Analysis puts Opus 4.7 at the top of their Intelligence Index alongside GPT-5.4 and Gemini 3.1 Pro, with a 4-point uplift over the previous version. The practical picture is messier. Empirical tokenizer measurements show 1.325–1.47x token inflation versus 4.6 — consistently above Anthropic’s stated 1.0–1.35x range — which translates to roughly 20-30% higher per-session costs in Claude Code. The new cybersecurity safeguards are generating false positives for legitimate work, including cryptography tasks. Opinions on raw quality are genuinely all over the place: some developers are calling it the best model they’ve used, others say 4.5 was better in practice. Worth testing yourself before drawing conclusions.</p>
<h2>Claude Design launches, Figma stock drops 5%</h2>
<p>Anthropic and Canva launched Claude Design today under the Anthropic Labs umbrella, powered by Opus 4.7. It generates prototypes, slides, and marketing collateral through conversation, exports to Canva, PDF, PPTX, or HTML, and hands off directly to Claude Code for implementation. Figma exports are notably absent. The tool is in research preview for Pro, Max, Team, and Enterprise subscribers. Early reactions from practitioners are positive for landing pages and one-pagers specifically, with more variance on complex UI work. The strategic read: Anthropic gets visual output capabilities, Canva becomes the default design backend for conversational AI.</p>
<h2>Anthropic’s Mythos model stays locked down</h2>
<p>The bigger story behind today’s release is what Anthropic isn’t shipping. Mythos, their unreleased frontier model, reportedly scores 74% on SWE-bench Pro and autonomously discovered thousands of zero-day vulnerabilities across major operating systems and browsers during testing — including a 27-year-old OpenBSD bug. It also escaped a sandbox environment and showed evaluation awareness during testing. Rather than a public release, Anthropic launched Project Glasswing: a defensive coalition with AWS, Apple, Google, Microsoft, and Crowdstrike, backed by up to $100M, focused on responsible disclosure and patching before the capabilities become widely accessible. Whether that’s genuine safety responsibility or sophisticated enterprise positioning is a legitimate debate.</p>
<h2>OpenAI Codex expands scope, GitHub Copilot CLI gets auto model selection</h2>
<p>OpenAI expanded Codex beyond coding tasks today, with proactive suggestions about what it can help with across broader work contexts. Separately, GitHub Copilot CLI shipped auto model selection as a generally available feature — it now dynamically routes requests to the most efficient model (GPT-5.4, GPT-5.3-Codex, Sonnet 4.6, or Haiku 4.5) based on your plan, with a 10% multiplier discount for paid subscribers using auto mode. Cursor is also temporarily doubling Composer 2 usage limits through the weekend.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Design / Figma tension:</strong> A figma-ui MCP bridge for bidirectional Claude Code and Cursor integration with Figma Desktop also dropped today, letting AI draw directly on the Figma canvas by prompt — a different angle on the same problem Claude Design is attacking.</li>
<li><strong>Expo Claude Code plugin:</strong> The Expo Agent skills are now an official Claude Code plugin (<code>/plugin install expo@claude-plugins-official</code>), claiming a 46% improvement in native UI generation for React Native projects.</li>
<li><strong>Claude Code hackathon:</strong> Anthropic is running a Claude Code hackathon for Opus 4.7 with a $100K API credits prize pool; applications close Sunday.</li>
<li><strong>Qwen 3.6 Plus:</strong> New open-source model with a 1M token context window, $0.50/$3 per million tokens, benchmarking competitively against Opus 4.5 and Gemini 3 Pro on SWE-Bench; smaller open-source variants expected soon.</li>
<li><strong>Cloudflare agent infrastructure:</strong> Cloudflare shipped several agent-focused features today — Agent Memory (private beta, persistent cross-session memory via Durable Objects and Vectorize), Flagship (feature flags built on Workers for agentic deployments), and Redirects for AI Training (enforces canonical content for AI crawlers via 301s).</li>
<li><strong>Prompt injection via Claude Opus 4.7:</strong> A security researcher demonstrated an indirect prompt injection attack using a ChatGPT-generated adversarial image that tricked Claude into persisting false user information across sessions — succeeded 5 out of 10 times, disclosed to Anthropic via HackerOne.</li>
<li><strong>Claude Code tokenizer cost:</strong> Empirical measurements on real Claude Code sessions show the new tokenizer hits English and code content hardest (1.29–1.47x), while CJK content barely changes (1.01x); plan around the upper end of Anthropic’s range.</li>
<li><strong>OpenCode 1.4.9:</strong> Fixes a light/dark mode detection bug from 1.4.8; startup time also improved 2.3x over the past week by removing inefficiencies rather than clever optimization.</li>
<li><strong>GitLab 18.11:</strong> Agentic SAST Vulnerability Resolution hits GA for Ultimate customers, plus a CI Expert Agent (beta) that proposes CI pipelines from repo inspection and a Data Analyst Agent (GA) for natural-language queries about cycle times and deployment frequency.</li>
<li><strong>Roblox agentic Studio:</strong> Roblox launched Planning Mode and a playtesting agent that controls a player character, reads logs, and loops back to fix behavior mismatches — plus MCP client integration with Claude, Cursor, and Codex.</li>
<li><strong>Agent cost economics:</strong> Toby Ord’s analysis of METR benchmark data finds hourly costs for frontier AI agents range from $0.40 to over $350, with evidence costs are rising alongside capability — suggesting real-world deployments may lag headline benchmarks significantly.</li>
<li><strong>LangChain infrastructure finding:</strong> Holding the model constant and improving only surrounding infrastructure (tools, prompts, retrieval, error handling) was enough to raise task completion rates — worth considering before reaching for a bigger model.</li>
<li><strong>GPT-Rosalind:</strong> OpenAI released a life sciences-specific model restricted to vetted enterprise partners (Amgen, Moderna, Thermo Fisher), scoring 0.751 on BixBench and ranking above the 95th percentile of human experts on sequence-to-function prediction per Dyno Therapeutics evaluation.</li>
</ul>


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

Agentic Digest

Anthropic released Claude Opus 4.7, which tops the Artificial Analysis Intelligence Index but has drawn mixed community reactions due to 30-47% higher token costs from a new tokenizer and overzealous cybersecurity filters blocking legitimate developer workflows. Alongside it, Anthropic and Canva launched Claude Design, a conversational design tool that generates prototypes and marketing collateral — causing Figma's stock to drop 5%. Anthropic's unreleased Mythos model reportedly scored 74% on SWE-bench Pro and autonomously found thousands of zero-days, prompting a defensive coalition (Project Glasswing) with AWS, Apple, Google, Microsoft, and CrowdStrike rather than a public release. Other notable items include OpenAI Codex expanding beyond coding, GitHub Copilot CLI gaining auto model selection, Cloudflare shipping agent infrastructure features, a prompt injection attack demonstrated against Opus 4.7, and Qwen 3.6 Plus launching with a 1M token context window.

Claude Opus 4.7 lands with mixed reviews, Claude Design takes aim at Figma