Did Codex just lap Claude Code?

A quiet post-launch week, but the usage numbers weren't quiet: OpenAI's Codex claims 7M users and roughly 10x growth in six months, reopening the question of who leads the coding-agent race. Elsewhere, a German consortium shipped a genuinely competitive fully-open 30B model, Apple's OpenAI complaint spilled its wildest details, and open-source agents crossed into venture territory.

Codex claims 7M users and 10x growth — enough to catch Claude Code?

Latent Space flags that GPT-5.6 Codex/Sol reportedly hit ~6M users on July 10-12 and ~7M a day later, per OpenAI figures — roughly 10x growth this year from an estimated 550-700k on Jan 1. The last public Claude Code numbers were ~2M weekly users and $2.5B ARR back in February. OpenAI also shipped Codex/Sol usage fixes: ~10% more usage from inference optimizations, a context rollback from 372k to 272k after billing side effects, and a reversion of experimental reasoning-effort changes.

Why it matters: The harness is now the product surface, and if Codex really is compounding 10x while Anthropic stays silent on numbers, the CLI coding-agent race is far closer than it looked. Treat the counts as self-reported.

Apple's OpenAI complaint: 400 poached staff, an auth bug, and prototypes at interviews

Details from the 41-page filing sharpen the case first reported last week: Apple says 400+ ex-employees now work at OpenAI, that engineer Chang Liu exploited a 'rare' authentication bug to reach Apple's network weeks after leaving ('LOL, I found out I can access the [network storage]'), and that hardware chief Tang Tan had candidates bring CAD files and physical prototypes to interviews. Apple also alleges io used its confidential metal-finishing techniques by misleading a supplier. OpenAI: 'We have no interest in other companies' trade secrets.'

Why it matters: Strip the espionage framing and this is a talent-mobility fight — OpenAI is well-funded enough to ignore the Valley's no-poach norms, and discovery could set precedent for how AI labs recruit from incumbents.

Germany's Soofi S is a fully-open 30B-A3B that tops the open-weight benchmarks

A KI Bundesverband consortium released Soofi S 30B-A3B, a Nemotron-3-Nano-style hybrid (Mamba-2 plus attention) activating 3.2B of 31.6B params, trained on 27T German-weighted tokens on Deutsche Telekom's B200 cloud. It claims the top aggregate scores among fully-open models — over OLMo 3 32B and Apertus 70B — with 73.8% HumanEval and roughly 8x more tokens/sec per GPU than dense 14-24B models at 40k context. Weakness: RULER long-context extraction collapses beyond 32k tokens. Weights, checkpoints, code and a full data inventory ship under OSI's Open Source AI Definition 1.0.

Why it matters: A concrete rebuttal to this week's 'why is no Western lab close to the Chinese open models' hand-wringing — and, with a documented reproducible recipe, more genuinely open than most 'open' releases.

Nous Research raising $75M+ at a $1.5B valuation on its open Hermes agent

TechCrunch reports Nous Research is finalizing a round led by Robot Ventures, with USV participating, at a $1.5B valuation. Its OpenClaw-style local agent Hermes — which ships with built-in skills (web search, coding, image understanding) and auto-learns new ones — has ~214k GitHub stars and ~40k forks, alongside hosted tiers from $20-200/month.

Why it matters: Open-source agents are now venture-scale; Hermes is the self-hostable counterweight to Codex and Claude Code, and the funding signals real demand for agents you can run on your own VPS.

Wan-Dancer breaks the 20-second wall for music-to-dance video

Alibaba's HumanAIGC released Wan-Dancer-14B (weights and inference code), a hierarchical framework that generates 720p/30fps dance videos exceeding a minute directly from music. It decouples global keyframe planning from local refinement and uses time-mapped RoPE embeddings plus an optical-flow loss to fight the temporal drift and identity inconsistency that break diffusion models past ~20 seconds, claiming SOTA across five dance genres.

Why it matters: Minute-scale temporal coherence is the actual hard problem in video generation; shipping open weights means the SOTA claim is testable today rather than a demo reel.

Flint cuts reasoning tokens 2-3x with section-aware trace compression

A solo study trains Qwen3.5-4B and Gemma-4-12B on self-distilled traces where compute and verification spans are kept but narration and transitions are dropped; the models match or beat their originals at ~1.7x fewer reasoning tokens. A sharp finding: flat compression makes greedy decoding loop on 93% of GSM8K at temperature 0, because the model uses computation spans as a termination anchor. Everything is small-scale (322-648 rows per arm, ~1.5 3090-hours) but reproducible, with models, datasets and code released.

Why it matters: A cheap, open recipe to trim inference cost on reasoning models — plus a concrete mechanistic explanation of why compressed models loop, which is useful even if you never train one.

SK Hynix's StreamDQ moves weight dequantization into HBM

An SK Hynix paper proposes StreamDQ, a near-memory architecture that performs on-the-fly weight dequantization inside custom HBM for high-throughput, large-batch LLM inference. It reports up to 7.08x speedup and 90.23% lower energy on mixed-precision GEMM.

Why it matters: If dequantization happens in the memory subsystem rather than the GPU, quantized serving stops paying the bandwidth tax on every weight fetch — potentially a big lever for FP4 and mixed-precision inference at scale.

Also worth a look

Browse previous days →