<p><strong>TLDR:</strong> Anthropic’s Claude Mythos is real and restricted — 93.9% on SWE-bench Verified, access limited to AWS, Apple, Microsoft, and Google. DeepSeek V4 specs leaked ahead of a late April release, running on Huawei Ascend chips with a 1T parameter MoE architecture. OpenAI is quietly making Codex its main product, not a side project. Meanwhile, Claude Code’s source map leak exposed Kairos, an always-on autonomous agent that nobody officially announced.</p>
<hr>
<h2>Claude Mythos: restricted access, serious numbers</h2>
<p>Anthropic confirmed Claude Mythos (Project Glass Wing internally), and the numbers are hard to dismiss: 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, and autonomous discovery of a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw. Access is restricted to a small group of major cloud providers, with Anthropic briefing US officials and committing $4M to open source security. Pricing is $25/million input tokens and $125/million output tokens. The “too dangerous to release” framing is getting skeptical reads from parts of the community, but the benchmark jumps are real enough that dismissing it entirely seems premature.</p>
<h2>Claude Code source map leak reveals Kairos</h2>
<p>A 60MB JavaScript source map shipped publicly in a Claude Code update, exposing ~512,000 lines of TypeScript. The most notable find is Kairos — an always-on autonomous agent designed to run 24/7, take proactive actions without user prompts, and monitor GitHub pull requests. Two other features from the leak have since shipped: Ultraplan (offloads complex planning to a cloud Opus 4.6 session) and Buddy (a Tamagotchi easter egg, now removed). Kairos hasn’t shipped yet, but it’s clearly in active development.</p>
<h2>DeepSeek V4 specs leak ahead of late April release</h2>
<p>Chinese sources are reporting DeepSeek V4 specs: 1T parameter MoE with ~37B active, 1M context window, sparse attention via mHC and DSA, Engram memory architecture, and Huawei Ascend chip support. That last point matters — it would make V4 the first frontier model not dependent on NVIDIA hardware, which is the real story here. No official release date, but signals are loud enough that late April looks credible.</p>
<h2>OpenAI is reorganizing around Codex</h2>
<p>The $100/month plan, the Scratchpad feature (multiple parallel Codex sessions), Codex CLI updates 0.119.0 and 0.120.0, and the acquisition of Cirrus Labs (Tart, Cirrus CI) for the Agent Infrastructure team — OpenAI is clearly treating agentic coding as the core product now, not a feature. Cirrus CI shuts down June 1. The rate limit cuts for Plus users after the $100 plan launched are annoying, but the direction is obvious.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Qwen 3.6 Plus benchmark claims:</strong> Community posts are circulating numbers showing Qwen 3.6 Plus at 61.6 on Terminal-Bench and 56.6 on SWE-bench Pro, claiming first place on multiple agentic coding evals — take with appropriate skepticism until independently verified.</li>
<li><strong>Claude Opus 4.6 degradation:</strong> Multiple engineers and at least one data point (“thinking 67% less than at launch”) are reporting quality regression. Anthropic hasn’t published a transparent performance report, and the community is frustrated that this debate has been running for three years without one.</li>
<li><strong>Linux kernel now allows AI-written code:</strong> Human contributors bear full legal responsibility, AI tools cannot add Signed-off-by tags, and all AI-generated code must be personally reviewed before submission.</li>
<li><strong>Benchmark integrity problem:</strong> Berkeley researchers built an automated scanning agent that achieved near-perfect scores on SWE-bench, WebArena, OSWorld, and others without solving a single task — exploiting seven recurring vulnerability patterns including answers shipped with tests and broken scoring logic.</li>
<li><strong>Azure MCP Server 2.0:</strong> Microsoft shipped a stable release with 276 tools across 57 Azure services, Managed Identity auth, sovereign cloud support, and Claude Code integration.</li>
<li><strong>Meta Neural Computer paper:</strong> Meta published research proposing to replace conventional computer architecture with a single neural network trained on screen recordings. Current prototypes hit 54% character accuracy on terminal simulation and 4% arithmetic accuracy. Direction is interesting; gap from usable is enormous.</li>
<li><strong>Cirrus Labs joins OpenAI:</strong> The team behind Tart (Apple Silicon virtualization) and Cirrus CI is joining OpenAI’s Agent Infrastructure team. Cirrus CI shuts down June 1, all source-available tools are being relicensed with no fees.</li>
<li><strong>Vercel Sandbox claims top microVM speed:</strong> Vercel is positioning its sandbox as the fastest microVM-based environment for coding agents and parallel compute workloads.</li>
<li><strong>opencode supports 900+ models:</strong> The model picker breadth is impressive enough that it crashed at least one downstream tool (RepoPrompt) when a user tried to load the full list.</li>
<li><strong>Mac Mini eGPU support:</strong> Apple officially approved TinyGPU drivers for AMD and Nvidia GPUs on macOS via Thunderbolt/USB4, expanding local model capacity for Mac Mini-based agent setups.</li>
<li><strong>GLM-5.1 open-sourced under MIT:</strong> Claims to beat GPT-5.4 on SWE-bench Pro and can run autonomously for 8 hours — the open source lag narrative is getting harder to sustain.</li>
</ul>


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

Agentic Digest

A roundup of major AI coding developments: Anthropic's Claude Mythos achieves 93.9% on SWE-bench Verified but is restricted to major cloud providers. A Claude Code source map leak exposed Kairos, an unreleased always-on autonomous agent. DeepSeek V4 specs leaked showing a 1T parameter MoE model running on Huawei Ascend chips — potentially the first frontier model independent of NVIDIA. OpenAI is reorganizing around Codex as its core product, acquiring Cirrus Labs for agent infrastructure. Additional items cover benchmark integrity concerns, Linux kernel AI code policy, Azure MCP Server 2.0, and several open-source model releases.

Claude Mythos locked behind closed doors, DeepSeek V4 specs leak with Huawei chips