Eight gigawatts, a $7.5B model router, and a training run put on hold

WEEK OF AUG 15 – 21, 2026

Eight gigawatts, a $7.5B model router, and a training run put on hold

The money and megawatts moved faster than the models: OpenAI paused its largest frontier RL run, Stripe bought the routing layer, and DeepSeek added vision to its cheap agent stack.

No flagship US frontier model shipped this week. Instead, the infrastructure around models became the story — power, routing, safety overhead and the cost of seeing.

It is the week to remember that frontier AI is now a systems business before it is a model-release business.

FRI · Aug 21, 2026ModelsVision

DeepSeek adds vision to its cheap agent model and ships a free Files API

DeepSeek launched V4-Flash-Vision-Exp with image input, 1M context, JSON output, tool calls, Responses API and an Anthropic-compatible endpoint. Images cost roughly $0.17 per 1,000 at the published tokenization, while peak/off-peak pricing remains in place.

The free Files API lets you upload once and reference an image by file_id across requests. Harness 0.1.1 added support immediately.

WHY IT MATTERS

Screenshot-reading, chart parsing and self-correcting agent loops stop being a separate budget line when vision rides on the same inexpensive model.

DeepSeek API release SiliconANGLE Files API coverage

TUE · Aug 18, 2026SafetyRL

OpenAI pauses its largest frontier RL run over its own cyber evals

OpenAI said it temporarily slowed scaling, including a two-week pause in deployment training, while its largest planned frontier RL run remained on hold. Preliminary evals suggested an upcoming model could meet the Critical cyber threshold; monitoring was extended to all tool-using inference.

Activation classifiers at every sampled token escalate to automated investigators with a 30-minute alert target. OpenAI put monitoring overhead at roughly 20% of observed inference compute.

OpenAI pauses its largest frontier RL run over its own cyber evals — img.helpnetsecurity.com

WHY IT MATTERS

This is the first public price tag attached to restraint: safety monitoring is not free, and the lab stopped its own biggest run because the evals said stop.

OpenAI safety update Help Net Security The Register

TUE · Aug 18, 2026ResearchAgents

Claude designs protein binders against 14 of 15 targets

Claude ran 24–48-hour autonomous campaigns against 16 protein targets: researching targets, choosing epitopes, installing and running open design models, and making every design decision without human input. Two independent labs synthesized every design exactly as delivered.

354 of 1,320 designs bound — a 26.8% hit rate — across 14 of 15 interpretable targets. On RBX1, 28 of 90 designs bound at 3.9 nM versus 9 of 245 human competition entries at 45 nM.

WHY IT MATTERS

An autonomous agent using open tools beat an expert design campaign on a task that used to take specialists weeks. The capability is useful, impressive and explicitly dual-use.

Anthropic research Adaptyv wet-lab validation Dataset

MON · Aug 17, 2026InfrastructureEnergy

OpenAI signs an 8-gigawatt Ohio data center

OpenAI secured roughly 8 GW at the former Portsmouth Gaseous Diffusion Plant in Ohio, with SB Energy building and operating it under a 20-year lease. Nvidia’s filing disclosed credit support capped at $105B for the first 4.25 GW and a $1.5B equity investment in SB Energy.

The site is exclusively Nvidia compute; the first 800 MW is expected in 2028. The deal includes community funds and Codex credits for Ohio students.

OpenAI signs an 8-gigawatt Ohio data center — image.cnbcfm.com

WHY IT MATTERS

This is the new frontier-compute template: the lab takes capacity, the chip vendor takes residual risk, and the contract pays its own grid upgrades.

OpenAI announcement CNBC Data Center Dynamics

MON · Aug 17, 2026AgentsOpen source

Nous Research ships Bot Mode, turning agent profiles into a roster of named bots

Nous Research shipped Bot Mode for Hermes Desktop, replacing the single-agent session list with a roster of named bots. Each bot is a real Hermes profile with its own chat, memory, skills, pinned model and avatar, and they message each other through a persistent Agent Inbox and hand work off by mention. It ships as a desktop plugin rather than a core patch, MIT-licensed, and is now bundled on by default.

The announcement drew real traction — past 900,000 views on the company's own post within a day, and reposted well beyond its usual circle. Hermes Agent is already on this timeline twice this quarter: a harness comparison on 7 August where it posted the lowest cost per success at $0.46, and a $1.5B valuation round reported on 13 July.

Hermes Bot Mode for the Hermes desktop app — Nous Research

WHY IT MATTERS

Most agent products sell you one assistant; this sells a team you assemble once, with separate memory and a separate model per role — which is the shape multi-agent work actually takes in practice. The per-bot persona-and-memory split is a better organising idea than one ever-growing chat.

Bot Mode documentation Hermes-Bot-Mode on GitHub MarkTechPost