An unreleased Claude raised a proven Riemann-zeta bound to 67.2 percent
The day's other headline AI numbers, OpenAI's cyber benchmark and Meta's open-weight comparisons, were self-reports from the labs that produced them.
Four mathematicians and a Lean checker vetted Claude's Riemann result1
An unreleased research version of Claude lifted the proven lower bound on the fraction of Riemann zeta zeros lying on the critical line from 41.6 to 67.2 percent. One outside report calls it the single largest improvement to that bound in the history of the problem.2 The run took 31 million output tokens across two sessions and about 60 subagents.1 Anthropic's Levent Alpöge and Ralph Furman checked the argument, Brian Conrey and Dan Goldston examined it from outside, and Claude wrote the proof in a form a Lean checker can verify step by step.1 Anthropic says it does not expect the approach to prove the hypothesis itself,1 which would need the fraction to reach 100 percent.2
OpenAI says its new cyber model found two Chrome V8 zero-days3
OpenAI reported that GPT-5.6-Cyber, built on its GPT-5.6 Sol model, found two previously unknown bugs in Chrome's V8 engine that chain together to corrupt memory and escape the V8 heap sandbox, now tracked as CVE-2026-15903.4 On OpenAI's own Advanced Cybersecurity Completion Rate the model answered 95 percent of sensitive security queries, against 57.3 percent for the previous cyber model and 1.5 percent for the general GPT-5.6 Sol.4 That benchmark is OpenAI's own, and the announcement page carrying it would not open to an outside fetch,3 so the figures reach a reader only through OpenAI's account of them. TechCrunch, covering the same launch, confirms the partners OpenAI lined up, among them Accenture, IBM, CrowdStrike, and Cloudflare, but carries none of the vulnerability detail.5
A silicon photonic chip built a 16-qubit GHZ state and ran Grover's search6
A quantum-photonics group at USTC in Hefei reports, in an arXiv preprint not yet peer-reviewed, a room-temperature silicon chip that generates a four-photon state and, by encoding four qubits in each photon, assembles a 16-qubit GHZ state.6 The witnessed 10-qubit entanglement clears eleven standard deviations and is, the authors state, the largest entangled state shown on a photonic chip so far.67 Running Grover's search over four targets on a single-photon cluster state, the chip picks out the marked item with an average probability of 0.987, though its 10-qubit generation rate of 0.154 events per second leaves the throughput far below what a working processor would need.6
Meta released Muse Glimmer, a 30B open-weight agent, under Apache 2.08
Muse Glimmer is a roughly 29.6-billion-parameter dense transformer with a perception encoder, released under Apache 2.0 with a 131,072-token context and quantized builds that fit the model into 32GB of VRAM at a claimed 0.2 percent accuracy loss.8 Every benchmark in its favor is Meta's own, run on a set Meta chose: across 24 tests it reports beating the similarly sized Gemma4-31B and Qwen3.6-27B on about half and trailing on the rest.89 It wins SWE-Bench Pro outright, 51.2 to Qwen's 50.2 and Gemma's 36.9, and loses Qwen on the GDPVal-AA agent benchmark, 953 to 1,141.8
Sources
- Anthropic · Claude raises the proven lower bound for Riemann zeta zeros on the critical line
- Crypto Briefing · Claude improves the Riemann zeta lower bound to 67.2%
- OpenAI · Expanding Daybreak as the cyber-defense window narrows
- The Decoder · OpenAI launches GPT-5.6-Cyber to help defenders find vulnerabilities
- TechCrunch · As AI-led attacks multiply, OpenAI launches a new cyber model
- arXiv · On-chip generation of multi-qubit graph states with high-dimensional encoded single photons (2608.03012)
- The Quantum Insider · Multi-qubit photonic quantum states on a silicon chip
- Meta · Muse Glimmer-30B model card (Hugging Face)
- SiliconAngle · Meta releases open-source Muse Glimmer model with 30B parameters