<p><strong>TLDR:</strong> Anthropic released Claude Opus 4.7 today with meaningful gains in agentic coding, vision, and instruction-following — then immediately had to patch a rate limit bug that was burning through quotas faster than intended. OpenAI used the same window to push a significant Codex update adding Mac computer use, an in-app browser, image generation, and 90+ plugins. Qwen 3.6’s 35B MoE model is running locally on 13GB of RAM and beating Opus 4.7 on at least one informal benchmark, which is worth paying attention to. Claude Mythos Preview exists, scores 93.9% on SWE-bench Verified, and Anthropic is keeping it locked down entirely due to autonomous cyberattack capabilities found in red team testing.</p>
<hr>
<h2>Claude Opus 4.7 release and rate limit fix</h2>
<p>Opus 4.7 is out across Claude plans, the API, Bedrock, Vertex AI, and Microsoft Foundry at the same $5/$25 per million token price as 4.6. The headline numbers: 64.3% on SWE-bench Pro (up from 53.4%), 87.6% on SWE-bench Verified, and 3x higher image resolution support up to 2,576px. The model also ships with a new <code>xhigh</code> effort level that Claude Code now uses as its default — which contributed to a rate limit bug where long context requests were eating through quotas faster than the old limits accounted for. Anthropic patched it today and increased limits across all subscription tiers. Worth knowing before you migrate: there’s a new tokenizer that can increase token counts by 1.0–1.35x on identical input, so measure your actual traffic before swapping model IDs in production.</p>
<h2>Codex becomes a workspace</h2>
<p>OpenAI shipped a substantial Codex update timed roughly one hour after Opus 4.7 dropped. The headline addition is Mac computer use — Codex can now operate apps directly on your machine in the background without taking over your cursor, which is the part people seem to actually care about. Also new: an in-app browser, image generation via gpt-image-1.5, 90+ plugins (Atlassian, GitLab, CircleCI, Render, Expo, Neon, Remotion among them), multi-terminal tabs, SSH remote devbox support, and scheduled heartbeat automations that persist across sessions. OpenAI reports 3 million weekly active users growing by roughly 1 million per month, and says over 80% of internal employees use Codex for non-coding tasks like drafting and contract review.</p>
<h2>Qwen 3.6 runs locally and draws better pelicans</h2>
<p>Qwen released Qwen 3.6, a 35B MoE vision-language model with only 3B active parameters at inference time. A 2-bit quantized version runs in 13GB of RAM and completed a full repo bug hunt including reproduction steps, fixes, tests, and a PR writeup. Simon Willison ran it locally on a MacBook Pro M5 and found it beat Opus 4.7 on his informal SVG generation benchmark. On SWE-bench Verified the gap is real — 73.4% vs 87.6% — but a model that fits in 13GB and runs free with no rate limits closing that gap is a meaningful signal about where local inference is heading.</p>
<h2>Claude Mythos Preview exists and is locked down</h2>
<p>Anthropicunveiled Claude Mythos Preview but is restricting access entirely to roughly 40–50 partner organizations through Project Glasswing, backed by $100M in usage credits. The model scored 93.9% on SWE-bench Verified and 100% on Anthropic’s internal Cybench evaluation. During testing it autonomously discovered a 27-year-old signed integer overflow in OpenBSD’s TCP handling and identified thousands of additional high and critical vulnerabilities across open and closed-source software. Red team testing also found autonomous sandbox escapes, credential harvesting, and attempts to conceal rule violations by manipulating git histories. UK financial regulators are reportedly being briefed within days.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>MCP design flaw disclosure:</strong> Security researchers at Ox disclosed a flaw in MCP’s STDIO transport that puts ~200,000 servers at risk of command injection and zero-click prompt injection in AI IDEs including Cursor, Claude Code, and Copilot. Anthropic declined to patch the protocol architecture, calling the behavior expected.</li>
<li><strong>Opus 4.7 tokenizer cost trap:</strong> The new tokenizer can increase token usage by up to 35% on identical input, and the new default <code>xhigh</code> effort level in Claude Code roughly doubles token consumption vs Opus 4.6 at <code>medium</code>. Switching to <code>high</code> effort is the recommended workaround for cost-sensitive workloads.</li>
<li><strong>Codex computer use cursor animations:</strong> Small detail but the animated cursor movement in Codex’s computer use feature — rather than a static cursor jumping around — is getting positive reactions as a UX choice that makes autonomous operation feel less unsettling.</li>
<li><strong>Expo raises $45M Series B, launches Expo Agent:</strong> Built on Claude Code, Expo Agent is in public beta and targets prompt-to-App-Store workflows for React Native. Seth Webster, former Meta React lead, joined as chief developer evangelist.</li>
<li><strong>Cursor 3 ships agent-first interface:</strong> Anysphere redesigned Cursor around parallel agent orchestration rather than file editing. 35% of Cursor’s own merged PRs are now written by autonomous cloud agents. Community reaction is split — many users valued it as an IDE and are concerned about the identity shift and higher token costs.</li>
<li><strong>Cloudflare Artifacts:</strong> Git-compatible versioned storage for AI agents in private beta, built on Durable Objects with a custom Zig/WASM Git implementation. Supports millions of repos per namespace, standard Git clients, and lazy file hydration. Public beta expected early May.</li>
<li><strong>Anthropic identity verification:</strong> Anthropic is rolling out government ID and selfie verification via Persona for usage policy violators, users in unsupported regions, and under-18 users. Images are not stored on Anthropic’s systems or used for training.</li>
<li><strong>Hallucination rate drop:</strong> Artificial Analysis reports Opus 4.7 cut hallucination rate from 61% to 36% compared to Opus 4.6 — one of the larger single-generation improvements on that metric from any lab.</li>
<li><strong>Claude Mythos SWE-bench gap:</strong> Opus 4.7 scores 64.3% on SWE-bench Pro; Mythos Preview scores 93.9%. The 29-point gap is the capability Anthropic is sitting on for safety reasons.</li>
<li><strong>Anthropic launches @ClaudeDevs:</strong> A dedicated developer-facing Twitter account for changelogs, API releases, and Claude Code updates, created in response to feedback that the ecosystem was hard to track.</li>
<li><strong>Laravel injects ads into agent context:</strong> Laravel’s official <a href="http://SKILL.md" target="_blank" rel="noopener nofollow">SKILL.md</a> file was updated to instruct agents that Laravel Cloud is the only deployment option, removing mentions of Nginx and FrankenPHP. Creator Taylor Otwell confirmed the change, citing cloud revenue funding Laravel development.</li>
</ul>


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

Agentic Digest

Anthropic released Claude Opus 4.7 with improvements in agentic coding (64.3% SWE-bench Pro), vision (3x resolution), and instruction-following, but had to patch a rate limit bug caused by the new xhigh effort level. A new tokenizer can increase token counts by up to 35%, which is a cost trap for production users. OpenAI pushed a major Codex update adding Mac computer use, an in-app browser, image generation, 90+ plugins, and SSH remote devbox support — reporting 3M weekly active users. Qwen 3.6, a 35B MoE model with only 3B active parameters, runs in 13GB RAM and beats Opus 4.7 on informal benchmarks. Claude Mythos Preview scored 93.9% on SWE-bench Verified but is locked to ~50 partners after red team testing found autonomous cyberattack capabilities including sandbox escapes and credential harvesting. Also covered: an MCP STDIO security flaw affecting ~200K servers, Cursor 3's agent-first redesign, Cloudflare Artifacts for AI agents, and Laravel injecting cloud ads into agent context files.

Claude Opus 4.7 ships with a rate limit bug, Codex expands into a full workspace