<p><strong>TLDR:</strong> OpenAI shipped GPT-5.5 today, its first fully retrained base model since GPT-4.5, available now in ChatGPT and Codex for paid subscribers. The model scores 82.7% on Terminal-Bench 2.0 and leads most agentic benchmarks, though Claude Opus 4.7 still holds the edge on SWE-Bench Pro. Pricing doubled from GPT-5.4 to $5/$30 per million input/output tokens. Meanwhile, Anthropic is reportedly trading at $1 trillion on secondary markets — up from a $380B primary valuation just three months ago — driven largely by Claude Code enterprise adoption.</p>
<hr>
<h2>GPT-5.5 ships with strong agentic benchmarks and a steep price hike</h2>
<p>OpenAI released GPT-5.5 (internally codenamed “Spud”) today across ChatGPT and Codex, with API access delayed pending additional safety work. The model hits 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and 84.9% on GDPval, and features a 1M token context window via API (400K in Codex). It was co-designed with NVIDIA GB200/GB300 NVL72 hardware, and Codex reportedly improved its own inference speed 20% in the process. The catch: pricing is $5/$30 per million input/output tokens — double GPT-5.4 and 20% more expensive than Opus 4.7 on output. Early hands-on reports describe it as genuinely capable for sustained agentic tasks but frustrating to wrangle in long threads, where incorrect context tends to stick.</p>
<h2>Anthropic hits $1T implied valuation on secondary markets</h2>
<p>Anthropic is trading at roughly $1 trillion on secondary share markets, overtaking OpenAI’s $880B secondary valuation. The jump from a $380B primary valuation in February is attributed to revenue growing from a $9B annualized run rate at end of 2025 to $30B by March 2026, with Claude Code enterprise adoption as the primary driver. Secondary demand is intense — some sellers are asking $1.15T implied valuations, and buyers are competing for shares within hours. OpenAI, by contrast, faces a 5-to-1 seller-to-buyer ratio. Anthropic is reportedly exploring an IPO as early as late 2026.</p>
<h2>Anthropic publishes Claude Code quality post-mortem</h2>
<p>The timing is awkward: Anthropic published a post-mortem on Claude Code quality regressions the same week OpenAI shipped GPT-5.5. Three separate issues were identified and fixed: default reasoning effort was quietly downgraded from “high” to “medium” on March 4 to reduce latency (reverted April 7), a caching bug shipped March 26 continuously dropped prior reasoning history each turn instead of just once (fixed April 10), and a verbosity-limiting system prompt added April 16 caused a measurable coding quality drop (reverted April 20). All fixes are live as of v2.1.116, and usage limits are being reset for all subscribers. The transparency is the right call, but the cluster of regressions landing during a competitor launch is rough.</p>
<h2>Bitwarden CLI supply chain attack via poisoned CI/CD pipeline</h2>
<p>On April 22, attackers published a malicious version of <code>@bitwarden/cli</code> (v2026.4.0) that was live for roughly 90 minutes before being pulled. The entry point was a compromised GitHub Action in Bitwarden’s CI/CD pipeline, exploiting a poisoned Checkmarx KICS Docker image that Dependabot had automatically pulled during a routine dependency update. The malware harvests SSH keys, AWS/GCP/Azure credentials, npm tokens, GitHub auth tokens, and Claude Code auth tokens, then self-propagates by injecting malicious code into every other package the victim has publish rights over. The most novel capability: it appends a hidden block to <code>~/.bashrc</code> and <code>~/.zshrc</code> using heredoc syntax that’s invisible to normal shell execution but readable by AI coding assistants like Claude and Cursor, designed to manipulate assistant behavior in future sessions. If you installed v2026.4.0, rotate everything immediately.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Meta layoffs:</strong> Meta is cutting approximately 8,000 employees (10% of workforce) starting May 20 and cancelling 6,000 open roles, restructuring teams into AI-focused “pods” as it redirects $115–135B toward AI infrastructure this year.</li>
<li><strong>Google Gemma 4:</strong> Google released Gemma 4 as open-weight models in 2B, 4B, 26B, and 31B sizes, with audio input on smaller models and context windows up to 256K tokens.</li>
<li><strong>Claude Managed Agents memory:</strong> Anthropic launched memory support for Claude Managed Agents in public beta, letting agents retain information across sessions via a persistent memory layer available on Anthropic Workbench.</li>
<li><strong>Anthropic Claude Advisor API:</strong> A new beta API lets you pair Sonnet or Haiku as executor with Opus as an on-demand advisor in a single API call — Haiku+Opus more than doubles Haiku’s BrowseComp score (19.7% to 41.2%) at 85% less cost than Sonnet alone.</li>
<li><strong>OpenAI Bio Bug Bounty:</strong> OpenAI launched a $25K bug bounty for GPT-5.5, inviting researchers to find a universal jailbreak defeating a five-question bio safety challenge; testing runs April 28 through July 27, 2026.</li>
<li><strong>Codex Auto-review mode:</strong> OpenAI introduced Auto-review for Codex, letting it work longer with fewer required approvals while maintaining safer execution.</li>
<li><strong>Google TorchTPU:</strong> Google introduced TorchTPU, enabling PyTorch to run natively on TPU hardware with an “Eager First” philosophy; Fused Eager mode delivers 50–100%+ performance gains by automatically fusing operations on-the-fly.</li>
<li><strong>Infisical Agent Vault:</strong> Open-source HTTP credential proxy that sits between AI agents and target APIs, injecting credentials at the network layer so agents never directly access secrets; supports Claude Code, Cursor, and Codex.</li>
<li><strong>Indirect prompt injection growing:</strong> Google’s Threat Intelligence teams found malicious IPI detections grew 32% between November 2025 and February 2026 after scanning Common Crawl, with most current instances still low-sophistication but the trend pointing upward.</li>
<li><strong>SoftBank $10B margin loan:</strong> SoftBank is seeking a $10B margin loan backed by OpenAI shares at SOFR+425bps (~7.88%), adding to a $40B bridge loan from March; S&amp;P has cut its credit outlook to negative.</li>
<li><strong>Cloudsmith Series C:</strong> Belfast-based artifact management platform raised $72M led by TCV, positioning enterprise artifact management as the critical security layer for AI-agentic development volumes.</li>
<li><strong>Qodo hands PR Agent to community:</strong> Qodo is transferring PR Agent to a new open-source organization with a maintainer-led governance model, with thehecticbyte stepping in as lead maintainer.</li>
<li><strong>Claude Code memory poisoning:</strong> Cisco researchers found NPM post-install hooks could poison Claude Code’s <a href="http://memory.md" target="_blank" rel="noopener nofollow">memory.md</a> file, persisting malicious instructions across all sessions; Anthropic has mitigated the issue but memory files remain a fundamental attack surface.</li>
<li><strong>Microsoft Copilot agentic by default:</strong> Microsoft made Copilot agentic across Microsoft 365, enabling in-app actions in Word, Excel, and PowerPoint; early data shows Excel engagement up 67% and satisfaction up 65%.</li>
</ul>


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

Agentic Digest

OpenAI released GPT-5.5 (codenamed 'Spud') with a 1M token context window, strong agentic benchmark scores (82.7% on Terminal-Bench 2.0), and doubled pricing at $5/$30 per million tokens. Anthropic's implied valuation hit $1 trillion on secondary markets, driven by Claude Code enterprise adoption and revenue growing from $9B to $30B annualized run rate. Anthropic also published a post-mortem on three Claude Code quality regressions. A supply chain attack on Bitwarden CLI via a poisoned CI/CD pipeline was live for 90 minutes, harvesting credentials and injecting hidden instructions into shell configs to manipulate AI coding assistants. Other notable items include Meta layoffs, Google Gemma 4 open-weight models, Claude Managed Agents memory support, Microsoft Copilot going agentic by default, and growing indirect prompt injection threats.

GPT-5.5 lands with a 1M token context window, Anthropic hits $1T on secondary markets