OpenAI says the changes to its model training will increase compute overhead by 20% of observed inference workload; the increase will not be handed to customers
THE SO WHAT
A 20% compute overhead hit for expanded chain-of-thought monitoring—absorbed by OpenAI rather than passed through—shows safety and abuse detection are now core COGS, not optional features. If you’re self-hosting or fine-tuning, budget similar overhead for monitoring and guardrails instead of treating them as afterthoughts.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIZ.ai prices GLM-5.3 API access at $1.40 per million input tokens and $4.40 per million output tokens, unchanged from GLM-5.2
Holding GLM-5.3 pricing flat at $1.40 / $4.40 per million tokens despite added cyber capabilities is a clear shot in the frontier-model price war. For teams building on APIs, budget less around list price changes and more around switching and multi-vendor orchestration costs.
Applied AIMotif Technologies CEO on Business Strategy, AI
A Korean startup openly aiming to build models that compete with frontier players shows how national ecosystems are backing local champions rather than defaulting to imports. If you deploy AI in Korea, plan for a mixed stack—global APIs plus local models tuned to language, regulation, and data residency.
Applied AISorry, Samsung — Google says its latest AI smartphone camera tool can measure bodies better than the Galaxy Watch's bioelectric sensors
If camera-based body metrics start to rival dedicated bioelectric sensors, the smartphone becomes the primary health sensor and wearables become accessories to its data layer. Health product teams should assume camera-first diagnostics as a near-term constraint and design around lighting, pose guidance, and verification, not just watch-based inputs.
Applied AISources: Beijing allowed ByteDance and Tencent to each receive about 10,000 Nvidia H200 chips at their mainland China facilities in recent weeks
Even “small batches” of ~10,000 H200s per company give ByteDance and Tencent enough headroom to train and serve competitive frontier-scale models domestically. Anyone assuming a hard compute ceiling on Chinese consumer AI over the next 12–24 months should update their threat models and partnership assumptions.