LLMs

On August 10, 2026, Anthropic announced an unreleased research build of Claude improved the proven lower bound on Riemann zeta zeros on the critical line from 41.6% to 67.2% — without proving the hypothesis. Claude synthesized a chain of recent analytic number theory papers over two Claude Code sessions (~60 subagents, 31M output tokens), and the result survived review by two Anthropic mathematicians, external experts Brian Conrey and Dan Goldston, and a Lean 4 formalization.

Zhipu's GLM 5.2 (Jun 16 2026, BenchLM 83) and Alibaba's Qwen 3.7 Plus (Jun 3 2026, BenchLM 76) both ship 1M-token contexts as open weights. We compare benchmarks, per-1M token pricing, MoE serving behavior, quantization, and licensing — and give a workload-by-workload winner with success-weighted cost math.

BenchLM counted 115 notable model releases in the 12 months ending Aug 10 2026 — roughly one every three days — with 44% open-weight and July 2026 the busiest month at 21. Alibaba and OpenAI each shipped 11, ahead of Anthropic and Google at 9 each. We break down what the cadence does to engineering teams and how to keep agent pipelines stable.

Anthropic released Claude Opus 5 on July 24, 2026 at $5 in / $25 out per million tokens — half the price of the flagship Claude Fable 5. Here is the token economics, the latency math, and a routing playbook for when to pay for frontier and when Opus 5 is the smarter call.

Mira Murati's Thinking Machines Lab shipped its first model: Inkling, an Apache-2.0 Mixture-of-Experts built deliberately as a fine-tuning base, not a frontier competitor. Here is the MoE fine-tuning strategy, quantization reality, and the fine-tune vs RAG vs prompt decision framework.

August 2026 opened with 25+ companies — Nvidia, Microsoft, Meta, then Google and OpenAI — signing a letter urging Washington not to restrict open-weight models, months after Claude Fable 5's worldwide suspension. This is the policy-economics collision shaping every enterprise model decision.

PRISM2 represents a breakthrough in multimodal pathology interpretation, bringing conversational clinical diagnostics to reality.

A deep technical analysis into why OpenAI pulled the plug on its autonomous Astra model just 48 hours after release.

Alibaba pushes the open-source boundaries with Qwen3.8-Max, a 400B dense model challenging frontier proprietary LLMs.

We know how to scale generative models, but scaling the Reward Models that guide them is proving to be a fundamentally harder mathematical challenge. Welcome to the era of Reward Over-Optimization.