TLDR: OpenAI posted $5.7B in Q1 2026 revenue but burned through $6.95B, running a -122% adjusted operating margin. Anthropic confirmed Mythos-class models are coming to general release once stronger safeguards are in place. DeepSeek permanently cut V4 Pro pricing by 75% and declared AGI as its primary goal heading into a $10B funding round. The model wars are accelerating while the unit economics remain deeply broken.
<hr>
<h2>OpenAI’s Q1 2026 financials: $5.7B revenue, $6.95B in losses</h2>
OpenAI is on track for $30B in 2026 revenue and over $36B in losses at current margins. ChatGPT weekly active users stalled around 905 million, well short of the 1 billion target, with a free-to-paid conversion rate of roughly 6%. The numbers are being released ahead of a rumored IPO filing, and Anthropic is doing its own financial messaging in parallel — neither company’s core economics look fundamentally different from a year ago.
<h2>Anthropic’s Mythos models are coming, no date given</h2>
A Project Glasswing update confirmed that Mythos-class models will reach general availability once Anthropic develops what it calls “far stronger safeguards.” No timeline beyond “near future” was given. Meanwhile, Anthropic is raising Claude Code subscription prices starting June 15th — users will get a monthly credit allowance, and heavy usage will spill over into API-rate billing. Microsoft reportedly canceled its internal Claude Code licenses over the resulting cost. Anthropic has still surpassed OpenAI in business adoption, with 34.4% of businesses using Anthropic versus OpenAI’s 32.3%.
<h2>DeepSeek cuts V4 Pro pricing 75%, targets AGI</h2>
DeepSeek permanently dropped V4 Pro to $0.435 per million input tokens and $0.87 per million output tokens, with cached input nearly free at $0.003625. The lab is also raising external capital for the first time, targeting a $10B valuation with at least $300M in this round — founder Liang Wenfeng told investors the goal is AGI, not short-term commercialization. Models will stay open-source under permissive licensing.
<h2>Google I/O: Gemini 3.5 Flash, Android agents, and a $916 OS claim under scrutiny</h2>
Google shipped Gemini 3.5 Flash as the new default in AI Studio, launched Gemma 4 with 256K context and Apache 2.0 licensing, and showed Android Studio generating full apps from a single prompt. The headline claim — that Gemini agents built an operating system for $916 — is getting picked apart: the “single prompt” was thousands of lines long, human intervention standards were undefined, and Google hasn’t released the prompt, code, or logs for independent verification. Antigravity CLI (formerly Gemini CLI) also shipped as a terminal interface for Gemini agents.
<hr>
<h2>Also notable</h2>
<ul>
<li>GitHub breach via VS Code extension: Hacker group Team PCP compromised roughly 3,800 internal GitHub repositories through a malicious VS Code extension; repositories are reportedly up for auction.</li>
<li>Nvidia driver vulnerability: Nvidia is urging GPU owners to patch immediately for 14 vulnerabilities, including a critical 8.8/10 memory flaw.</li>
<li>Codex ships to ChatGPT mobile: OpenAI brought Codex to iOS and Android across all plans, with Remote SSH going GA, Hooks for policy enforcement, and HIPAA-compliant local use for enterprise. Appshots on macOS lets you attach your current app window — screenshot plus underlying text — directly to a Codex thread.</li>
<li>Cursor SDK released: Cursor launched a Python and TypeScript SDK for building custom agents on Composer 2.5, with 90% off Composer usage over the long weekend.</li>
<li>GPT-5.5 vs. Opus 4.7: DHH and others are reporting GPT-5.5 has closed the gap significantly on complex agent work, with some finding it now competitive with or ahead of Claude Opus 4.7 for long-running tasks.</li>
<li>Grok Build rolling out: xAI is preparing Grok Build for SuperGrok subscribers; currently CLI-only, desktop app coming later. xAI also launched Grok Skills, a persistent workflow system with native Office and PDF file generation.</li>
<li>NVIDIA Nemotron diffusion models: NVIDIA released a family of diffusion language models (3B to 14B) that generate tokens in parallel rather than sequentially, hitting ~865 tokens/second on B200 hardware — roughly 4x the autoregressive baseline.</li>
<li>Zaya1-8B architecture: Zyphra’s 8B model introduces Compressed Convolutional Attention claiming 4x FLOP reduction and Markovian RSA for bounded-context reasoning chains; on one hard math problem it outperformed GPT-5.5 and Claude Opus, though benchmarks are self-reported.</li>
<li>turbopuffer crosses $100M ARR: The object-storage-based vector database used by Cursor, Anthropic, Notion, and Linear hit $100M ARR with less than $1M in venture capital raised.</li>
<li>Huxe shuts down: The AI podcast generation startup founded by former NotebookLM developers raised $4.6M then shut down as Spotify, Adobe, and Google commoditized the feature.</li>
<li>Sora shutdown strands Critterz: OpenAI shut down Sora in March, leaving the AI-produced animated feature without its core production tooling mid-production; Cannes 2027 is now the likely next window.</li>
<li>Salesforce Agentforce gap: 29,000 deals and $800M ARR, but flagship customer demos from Williams-Sonoma and others turned out to be works in progress; stock is down 30% in 2026.</li>
<li>Web dev AI survey (7,258 devs): 63% now use AI for more than half their code, up from a minority in early 2025. Claude leads in paid subscriptions. Top complaints: hallucinations (64%), poor code quality (53%), lack of context (38%).</li>
<li>AI-generated open source spike: Socket data shows npm package creation spiked exponentially since January 2026, with maintainers increasingly disabling contributions due to low-quality AI-generated PRs.</li>
<li>Microsoft open-sources RAMPART and Clarity: RAMPART is a Pytest-native framework for adversarial safety testing of AI agents; Clarity is a design-phase sounding board tool for clarifying assumptions before coding starts.</li>
<li>Pulumi <code>do</code> command: New CLI command for direct CRUD operations on any cloud resource without a full IaC project or state file, shipping as a research preview in v3.242.0.</li>
<li>KanBots: Open-source kanban desktop app that dispatches Claude Code and Codex agents in parallel, each in its own git worktree, with live cost tracking.</li>
<li>Cloudflare Browser Run rebuild: Rebuilt on Cloudflare’s own Containers platform, achieving 4x higher concurrency (120 simultaneous browsers vs. 30) and 50% faster response times, completing a six-layer managed agent infrastructure stack.</li>
<li>Andrej Karpathy joins Anthropic for frontier LLM R&amp;D, per Dev Weekly #112.</li>
</ul>

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

Agentic Digest

OpenAI reported $5.7B in Q1 2026 revenue against $6.95B in losses, running a -122% adjusted operating margin, with ChatGPT stalling at 905M weekly active users. Anthropic confirmed Mythos-class models are coming once stronger safeguards are ready, while raising Claude Code subscription prices — reportedly causing Microsoft to cancel its internal licenses. DeepSeek cut V4 Pro pricing by 75% and announced a $10B funding round with AGI as its stated goal. Google I/O showcased Gemini 3.5 Flash and a disputed claim of building an OS for $916. Additional items include a GitHub breach via malicious VS Code extension, Nvidia GPU driver vulnerabilities, Codex on mobile, Cursor SDK launch, NVIDIA Nemotron diffusion models, turbopuffer hitting $100M ARR, and a survey showing 63% of developers now use AI for more than half their code.

OpenAI's -122% operating margin lands, Anthropic's Mythos models get a timeline