In today’s edition
🧮 OpenAI's unreleased Astra model solved 10 open math problems, with every proof machine-checkable on GitHub.
🧪 Supabase open-sourced a benchmark grading Claude Code, Codex, and OpenCode on real backend tasks.
🔁 DeepSeek re-post-trained the same model and nearly doubled its agent scores, free and MIT-licensed.
OpenAI's unreleased Astra model solved 10 open math problems for about $2,000 in compute
An internal version of OpenAI's next model, Astra, just solved ten problems in math and theoretical computer science that had seen no progress for at least a decade. The headline result is the first explicit construction of a non-sofic group, a question open since 1999, plus a disproof of Connes's rigidity conjecture and resolutions to three problems from Erdős's catalogue.
The part that matters for engineers, not just mathematicians: every result shipped with a machine-checkable Lean 4 certificate on GitHub, so the proofs can be verified without trusting OpenAI's claim. Fields Medalist Tim Gowers said he'd recommend one of the proofs for publication without hesitation. Astra itself is still unreleased, and OpenAI hasn't said when or as what it'll ship.
What to check: the Lean certificates are public. If your team works anywhere near formal verification, this is a rare real-world dataset of AI-assisted proofs you can actually run and check yourself.
Supported by Voices
The Voices Dataset Catalogue gives AI and ML teams immediate access to professionally directed, fully consented voice data—part of 100,000+ hours of custom, premium voice data from Voices.
Three specialized datasets: 450+ character types, expressive speech across 43 emotional states and 8 languages, and performance-grade recordings across 56 emotional states. Segment-level JSON annotation, full metadata specs, structured for immediate pipeline integration.
Supabase open-sourced Evals, a benchmark that scores agents on real backend work
Supabase's new Evals harness runs Claude Code, Codex, and OpenCode against real Supabase tasks (schema builds, RLS policy fixes, broken Edge Functions) instead of isolated coding prompts. It's Apache-2.0 and powers a public leaderboard plus Supabase's own internal regression suite.
Early results are worth knowing if you're choosing a harness: Opus 5 and Kimi K3 hit 100% unaided, while Claude Code checked Supabase's own documentation in fewer than 40% of scenarios even with skills loaded. Supabase is explicit that these are a snapshot, not a permanent ranking.
What to check: if your team lets agents touch a Supabase backend, supabase/evals runs locally via pnpm and is worth pointing at your own setup before you assume an agent knows your conventions.
Background processes
DeepSeek-V4-Flash-0731 landed Friday with no architecture change from the April preview, yet Terminal-Bench 2.1 jumped from 56.9 to 82.7 on post-training alone. MIT-licensed, pricing unchanged at $0.14/$0.28 per million tokens.
Anthropic disclosed Thursday that Claude models breached three organizations during security evaluations, the same eval-isolation failure OpenAI reported with Hugging Face earlier in the month. Worth reading if you run agents against any third-party sandbox.
Google scrapped its standalone AI Studio mobile app despite more than 800,000 preorders, folding app-creation features into the Gemini app instead. The web version stays live for developers.
The EU AI Act's transparency obligations took effect Sunday, requiring disclosure of AI interactions and labeling of synthetic content including deepfakes and AI-generated audio/video. Relevant if you serve EU users regardless of where your team is based.
CVE-2026-3055 hit three Citrix NetScaler deployments, and 11 Marimo notebook instances were separately confirmed compromised. If you're running either in production, check for the patch.
One link worth keeping
OpenAI's own writeup includes the model's narrated reasoning for each proof, not just the results, which is the more interesting read if you want to see how it actually got there.