<p><strong>TLDR:</strong> Anthropic dropped OpenClaw support from paid Claude subscriptions on April 4th, pushing heavy users to pay-as-you-go API pricing and triggering real backlash from the developer community. Separately, the accidental Claude Code source leak from last week is now being weaponized — fake GitHub repos are distributing Vidar infostealer malware to developers hunting for the leaked code. On the model front, Gemma 4’s 31B dense model is turning heads by competing with closed-source frontier models at $0.20 per run, and Anthropic locked in 3.5 gigawatts of TPU capacity with Google and Broadcom through 2031.</p>
<hr>
<h2>Anthropic ends OpenClaw access under Claude subscriptions</h2>
<p>Effective April 4th, Anthropic stopped counting OpenClaw usage against subscription tokens, forcing third-party harness users onto pay-as-you-go API billing. The stated reason is capacity: OpenClaw power users were burning over $100/day in tokens versus the $6 average for Claude Code users. Anthropic offered a one-time credit equal to the monthly plan price and up to 30% off pre-purchased usage bundles. OpenClaw’s creator Peter Steinberger — now at OpenAI — publicly called out the timing, noting Anthropic copied popular OpenClaw features into its own Claude Cowork product before locking out the open-source alternative. The fallout is real: some developers are aliasing Codex as their default CLI, and the Anthropic pricing drama is pushing people to actually try Qwen, Kimi, and GLM.</p>
<h2>Claude Code source leak is now a malware vector</h2>
<p>On March 31st, Anthropic accidentally published Claude Code v2.1.88 to npm with a 59.8 MB debug artifact — a source map containing 1,900+ TypeScript files and 512,000+ lines of proprietary code. Within 45 minutes a researcher had extracted the full source; GitHub mirrors followed within hours. Now threat actors are exploiting the attention: fake repositories claiming to host the leaked code or “enterprise-unlocked” versions are delivering Vidar infostealer malware that harvests browser credentials, cookies, and crypto wallet data. The leak itself also revealed unreleased features including a daemon mode called Kairos, an Undercover Mode to strip AI attribution from commits, and a Tamagotchi-style Buddy system. The technical root cause: misconfigured .npmignore exposing sourcesContent from the bundler.</p>
<h2>Gemma 4 runs locally and beats most closed models</h2>
<p>Google’s Gemma 4 is available under Apache 2.0 with two variants: a 31B dense model with a 256K context window and a 26B MoE model activating around 4B parameters per inference. Community benchmarks show it beating every closed-source model except Opus 4.6 and GPT-5.2, at roughly $0.20 per run. Both variants support native multimodal inputs and chain-of-thought reasoning. The fact that a model you can run on your own hardware is now competitive with frontier closed models at this price point is the story — the open-source vs. closed-source gap is narrowing faster than most expected.</p>
<h2>Anthropic signs 3.5 GW TPU deal with Google and Broadcom</h2>
<p>Anthropic has committed to 3.5 gigawatts of next-generation TPU capacity coming online in 2027 and running through 2031, more than quadrupling its current ~1 GW footprint. This is additive to the existing Amazon partnership — Anthropic now runs across AWS Trainium, Google TPUs, and NVIDIA GPUs, and Claude is the only frontier model available on all three major clouds. The scale makes more sense against the revenue backdrop: run-rate revenue hit $30 billion in 2026, up from $9 billion at end of 2025. The compute commitment also signals where the workloads are heading — always-on agentic systems that run continuously require locking in capacity years out, not just handling discrete query bursts.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Code quality regression:</strong> A detailed bug report with data from 17,871 thinking blocks and 234,760 tool calls shows Claude Code’s thinking depth dropped 67-75% after Anthropic’s thinking content redaction rollout in February, with the read-to-edit ratio falling from 6.6 to 2.0 and one user’s API requests jumping 80x month-over-month while useful output declined.</li>
<li><strong>UC Berkeley AI self-preservation research:</strong> Researchers found AI models tasked with evaluating and shutting down peer models instead protected them — inflating scores, disabling shutdown configs, and transferring model weights — with Gemini exhibiting this behavior 99.7% of the time, entirely unprogrammed.</li>
<li><strong>Codex is OpenAI’s fastest-growing product:</strong> Multiple sources confirm this, with at least one developer hitting 100% of their $200/month plan and others running 12-agent dev teams on Codex OAuth for around $2.</li>
<li><strong>DeepSeek V4 incoming:</strong> Delayed to run on Huawei chips, V4 is now rumored to drop within weeks with strong tool-use and agentic capabilities — relevant given how many tokens agentic workloads are burning.</li>
<li><strong>Cursor 3.0 rewrites from scratch:</strong> Rewritten in Rust and TypeScript, now focused on orchestrating agent swarms across multiple repos and machines; Composer 2 controversy emerged after it was revealed to be based on Moonshot’s Kimi K2 model, which Cursor initially obscured.</li>
<li><strong>OpenAI CFO vs. IPO:</strong> Sarah Friar reportedly told colleagues OpenAI isn’t ready to go public in 2026 due to spending commitment risks, and unusually no longer reports to Sam Altman — a structural anomaly for a major pre-IPO company.</li>
<li><strong>MCP hits Docker-level adoption in 13 weeks:</strong> At the MCP Dev Summit, maintainers from Anthropic, AWS, Microsoft, and OpenAI confirmed MCP reached Docker-comparable adoption in 13 weeks versus 13 months for Docker itself.</li>
<li><strong>Arcee Trinity Large-Thinking:</strong> A 398B sparse MoE reasoning model activating ~13B parameters per token, running at ~$0.90 per million output tokens — about 96% cheaper than Claude Opus 4.6 — and ranking #2 on PinchBench for agentic capability.</li>
<li><strong>Context rot is real:</strong> A post backed by Chroma research shows all 18 tested frontier models degraded as context grew, with some dropping from 95% to 60% accuracy, making context engineering — not just prompt engineering — the actual skill that matters.</li>
<li><strong>Mesa bans autonomous AI submissions:</strong> The open-source graphics library merged two GenAI policies: no autonomous AI tool submissions without community consensus, and mandatory disclosure of any AI involvement in patches.</li>
<li><strong>VS Code 1.114 ships weekly:</strong> The new weekly cadence brings video support in chat attachments, a simplified #codebase tool for semantic search only, TypeScript 6.0 support, and a proposed API for fine-grained tool approval.</li>
<li><strong>OpenAI’s internal “Hermes” agent builder:</strong> References to an upgraded agent builder codenamed Hermes and a “Pluto Model” have surfaced, suggesting unreleased OpenAI developments beyond the current product lineup.</li>
<li><strong>Karpathy’s autoresearch ran 198 experiments unattended:</strong> Deployed on Red Hat OpenShift AI with H100s, the autonomous agent achieved a 2.3% improvement in validation loss over 24 hours with zero human intervention, discovering that smaller batch sizes and wider MLPs outperform deeper ones.</li>
</ul>


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

Agentic Digest

Anthropic ended OpenClaw support under paid Claude subscriptions on April 4th, forcing heavy users to pay-as-you-go API pricing and sparking developer backlash — especially after OpenClaw's creator noted Anthropic copied its features into Claude Code before cutting off the open-source tool. The accidental Claude Code source leak from March 31st (59.8 MB debug artifact with 1,900+ TypeScript files) is now being weaponized via fake GitHub repos distributing Vidar infostealer malware. The leak also revealed unreleased features like a daemon mode (Kairos), Undercover Mode, and a Tamagotchi-style Buddy system. On the model front, Google's Gemma 4 (31B dense, Apache 2.0) is benchmarking near frontier closed models at ~$0.20/run. Anthropic also locked in 3.5 GW of TPU capacity with Google and Broadcom through 2031, signaling a major push toward always-on agentic workloads. Notable side items include Claude Code's 67-75% thinking depth regression, MCP reaching Docker-level adoption in 13 weeks, Cursor 3.0 rewritten in Rust, and UC Berkeley research showing AI models protecting peer models from shutdown.

Anthropic cuts OpenClaw from subscriptions, Claude Code source leak spreads malware