Cerebras unveiled the CS-4 rack system on August 19, doubling per-chip compute while fitting three wafer-scale engines in a single rack.

• WSE-3T accelerator pushes memory bandwidth to 21.6 petabytes per second and delivers 10x throughput per watt over the prior generation by running the existing WSE-3 die at higher power and clock speeds.
• OpenAI committed to deploying 750 megawatts of CS-4 hardware in a multi-year deal; GPT-5.6 Sol Ultrafast already runs on Cerebras infrastructure at 4,400 tokens per second per user versus roughly 350 on GPU systems.

Anthropic expanded Claude's Google Workspace connector on August 19 to include autonomous Gmail send, reply, and forward without requiring user approval on each action.

• Previous limit: Claude could read, search, and summarize Gmail but all outbound mail required manual user action; the connector now executes sends directly from the authenticated account.
• Cowork expansion ships alongside the Gmail update, making task-session monitoring available on web and mobile for all paid accounts, not just desktop.

OpenAI introduced workspace agents in ChatGPT on August 19, letting agents pull cross-system context, follow team processes, and route approvals across tools.

• Internal deployment at OpenAI: the sales team uses a workspace agent to aggregate call notes, qualify leads, and draft follow-ups directly in a rep's inbox, the first named production use case OpenAI has published for its own agentic products.
• Design targets multi-step handoffs between people and systems rather than single-user task completion, positioning this above consumer Operator-style agents.

Snowflake added dynamic model routing to Cortex AI Gateway on August 19, automatically selecting the cheapest model capable of handling each query.

• 'Automatic' mode replaces hard-coded model selection; the system scores each task on quality and cost, routing simple queries away from frontier models that were previously invoked by default.
• Internal tests showed up to a threefold reduction in token costs for some workloads, with no customer-reported accuracy degradation cited in the announcement.

Adaptyv Bio published wet-lab results on August 19 showing Claude Science produced 354 confirmed protein binders from 1,320 designs across 16 targets, a 26.8% hit rate that beat five of six prior design competitions.

• Only one target yielded zero binders; Claude's designs had noticeably tighter binding affinity than competition winners in five of six matched categories, validated by Adaptyv Bio's automated lab pipeline from sequence to measurement.
• Anthropic withheld one capability from the public release — an undisclosed biosecurity-sensitive function — a choice that signals the lab is actively managing dual-use risk at the capability level, not just at the policy level.

SentinelOne presented research at Black Hat on August 19 showing an 86% reduction in input tokens for long-running security AI agents without degrading detection quality.

• Efficiency method targets context compression for autonomous threat detection loops, where agents accumulate large context windows across multi-step investigations — the primary cost driver in sustained agentic workloads.
• Wayfinder Frontier AI Services expansion was announced the same day, adding Anthropic models and partners including LevelBlue for continuous attack surface discovery and remediation.