Today the frontier AI race quietly relocated – off the cloud and onto the desk. Perplexity and Nvidia shipped a fully local AI agent the same week OpenAI made its case for owning the entire compute stack, while Europe’s sovereignty bloc hardened around the EU AI Act. The through-line for B2B buyers: where your agents run is becoming as strategic as how smart they are.
Perplexity
What happened
On August 25, Perplexity launched Portable Computer, an on-device AI agent built in close partnership with Nvidia that runs serious agent workloads locally at zero token cost. Separately, Nvidia is reportedly weighing a fresh investment that would value Perplexity above $30 billion, a jump of more than 50% from its $20 billion mark a year ago.
What it means for your agentic build
Local agents collapse per-token economics and keep data on-premises, which is compelling wherever cloud latency, cost, or data residency are blockers. Evaluate Portable Computer for regulated or high-volume workflows before your next cloud-inference renewal.
OpenAI
What happened
CEO Sarah Friar published ‘The full stack behind abundant intelligence,’ framing OpenAI as an integrated system spanning chips, data centers, frontier models, and devices. The company also touted Jalapeno inference efficiency, shipped an Admin plugin for ChatGPT Work and Codex, and reinstated a 5-hour usage cap for Plus subscribers.
What it means for your agentic build
OpenAI is signaling that capacity, not just capability, is the moat, and that enterprise governance is now table stakes. Factor the reinstated usage caps and the new admin controls into any Codex or ChatGPT Work rollout plan.
Anthropic
What happened
Anthropic pushed Claude deeper into agentic enterprise work, adding Claude Mythos 5 to Claude Security for Enterprise with codebase scans, vulnerability findings, and suggested patches, and launching a Defender Advantage Fund. It also began embedding EU AI Act-compliant watermarks in all Claude outputs worldwide.
What it means for your agentic build
Claude is positioning as a security-native agent platform with built-in provenance, not just a chat model. If cyber defense or content compliance matter to you, Anthropic now addresses both directly in one stack.
Google DeepMind
What happened
Google reshuffled its AI leadership: Demis Hassabis becomes chairman and Alphabet chief scientist while Koray Kavukcuoglu takes over DeepMind. Meanwhile Gemini 3.6 Flash and 3.5 Flash-Lite went stable and production-ready, even as the flagship Gemini 3.5 Pro slipped past multiple release deadlines.
What it means for your agentic build
Leadership turbulence and a delayed flagship raise near-term delivery risk, but the cheaper, faster Flash tiers are usable today. Lean on Flash for cost-sensitive agentic subtasks and keep a fallback for any roadmap that depends on Pro.
Meta AI
What happened
Meta Superintelligence Labs released Muse Spark 1.2, which scored 82.9% on Terminal-Bench with a 1M-token context, under a modified Llama open-weight license. It also shipped Muse Glimmer, a 29.6B-parameter Apache-2.0 model built to run continuously as a local agent on consumer GPUs.
What it means for your agentic build
Meta is arming the open-weight ecosystem with genuinely capable coding and always-on local agents. For teams that want control without a per-token bill, these are credible self-hosted foundations, so budget for the GPUs and MLOps to operate them.
xAI
What happened
xAI shipped Grok Bot, its entry into the ‘AI teammates you can give real work to’ category, running on persistent cloud computers with connectors and approvals. Its flagship Grok 4.6, with a 500k-token context, is now also available inside Google’s Enterprise Agent Platform and Model Garden.
What it means for your agentic build
Grok is entering the durable-agent category and, via Model Garden, becoming procurable inside existing Google Cloud contracts. It is worth piloting where you already run on GCP and want multi-model agent options.
DeepSeek
What happened
DeepSeek made V4-Pro generally available, a 1M-context agentic model with OpenAI Responses API compatibility and Codex support out of the box, plus low, high, and max thinking-effort levels. It also introduced peak and off-peak billing and raised V4 output rates.
What it means for your agentic build
Drop-in Responses API compatibility slashes switching cost, and off-peak billing rewards batch and async workloads, but the price increase erodes the cheap-frontier narrative. Model your real duty cycle and peak-hour mix before committing.
Europe’s Sovereign Stack: Mistral, Cohere and Aleph Alpha
What happened
With the EU AI Act’s Article 50 transparency rules now in force, Mistral launched European Compute Units, expanded Microsoft-backed EU capacity, and released Leanstral 1.5 plus an Agentic Search retrieval layer. Cohere shipped the open-source Command A+ mixture-of-experts model and continued integrating Aleph Alpha’s sovereign infrastructure, now backed by Schwarz’s STACKIT cloud.
What it means for your agentic build
Vendor domicile, data residency, and self-hosting are now hard procurement criteria in Europe, not talking points. If you sell into or operate in the EU, shortlist a sovereign-capable provider now to de-risk compliance.
This Week’s Structural Trends
Agents are moving on-device. Perplexity’s Nvidia-built Portable Computer and Meta’s Muse Glimmer both target always-on local agents, collapsing token costs and keeping data local — a direct answer to cloud cost and privacy objections.
The unit of AI is shifting from model to teammate. Grok Bot, DeepSeek V4-Pro, Anthropic, and Perplexity all frame their releases around persistent, tool-using agents that complete multi-step work, so buyers should weigh durability, connectors, and approvals rather than benchmarks alone.
Sovereignty is now a purchasing spec. The EU AI Act’s Article 50 taking effect has turned data residency and provenance — Anthropic’s watermarking, the Mistral, Cohere, and Aleph Alpha EU stacks — into contract-level requirements across the whole field.
Sources
siliconangle.com/2026/08/25/perplexity-ai-launches-portable-computer-on-device-ai-agent/ · openai.com/index/the-full-stack-behind-abundant-intelligence/ · anthropic.com/news · deepmind.google/blog/ · en.wikipedia.org/wiki/Meta_Superintelligence_Labs · docs.x.ai/developers/release-notes · tech.yahoo.com/ai/articles/deepseek-officially-launches-v4-pro · mistral.ai/news/regional-inference-open-models-new-compute/ · betakit.com/cohere-releases-its-most-powerful-ai-model-as-open-source/ · futurumgroup.com/insights/cohere-acquires-aleph-alpha