The whole field pivoted to agents at once. Perplexity, OpenAI, Google, Meta, xAI and DeepSeek all shipped agent-first releases this cycle, and 500K-to-1M-token context windows have quietly become table stakes rather than a headline feature. The contest is moving from which model is smartest to which agent finishes the job.
Perplexity
What happened
Perplexity shipped Brain, an agentic memory layer built as a traceable, self-improving Markdown filesystem, and extended its Computer agent into email: cc computer@perplexity.com and a full task runs and returns in the same thread. Its Search API is now compatible with the Vercel AI SDK, and its Agent API presets auto-reuse stable prompt-cache keys.
What it means for your agentic build
Persistent memory plus inbox-native agents means research and multi-step work can run where employees already operate, without a new interface to adopt. Framework-agnostic API access lowers the cost of trialing Perplexity inside an existing stack, so pilot Computer-in-email on one repeatable workflow and measure hours saved before expanding.
OpenAI
What happened
ChatGPT added native Google Drive support in Library and the composer, letting users browse files, open docs beside chats, and keep content linked to Drive, alongside Codex updates like read-only chat sharing and synced pinned chats. In the background, OpenAI’s S-1 is expected soon and new Ramp data shows it gaining on Anthropic among paying business users.
What it means for your agentic build
Drive grounding turns ChatGPT into a front-end over your corporate documents, which is powerful and risky in equal measure. Audit and tighten folder permissions before enabling access, and read the competitive share shift and a public offering as signals of both more procurement leverage and more roadmap scrutiny ahead.
Anthropic
What happened
Anthropic reported Q2 revenue of $10.9B, up 130%, with its first operating profit, roughly two years ahead of plan. It began EU AI Act machine-readable content marking on all products from August 2, launched Claude Academy, and made computer-use, browser-use, the Skills API and Files API generally available on its platform and via Microsoft Foundry.
What it means for your agentic build
Profitability plus generally available agent tooling and managed cost controls make Claude a lower-risk choice for regulated, budget-governed deployments. Content marking is now a compliance fact rather than a future concern, so add AI-content marking to your checklist for anything customer-facing or published.
Google DeepMind
What happened
Google released Gemini 3.7 Flash just three weeks after 3.6 Flash, with substantial gains in software engineering, coding and agentic workflows, while 3.6 Flash and 3.5 Flash-Lite went generally available with better token efficiency and a low-latency subagent tier. Separately, Google’s A2A agent-interoperability protocol joined the Linux Foundation’s Agentic AI Foundation, the same neutral home as Anthropic’s MCP.
What it means for your agentic build
Cheaper, faster Flash tiers make high-volume agent automation economically viable, so re-run the unit-cost math on any workflow you shelved as too expensive on older pricing. A2A moving under neutral governance alongside MCP de-risks betting on cross-vendor agent interoperability.
Meta AI
What happened
Meta Superintelligence Labs released Muse Code in beta and Muse Spark 1.2, which posts 82.9% on Terminal-Bench with multi-agent coordination for long-horizon coding and a 1M-token context window, and Mark Zuckerberg confirmed Spark 1.2 weights will be open-sourced. Meta also shipped Muse Glimmer, a 29.6B-parameter multimodal model distilled from Muse Spark under Apache 2.0.
What it means for your agentic build
Open-weight, frontier-class coding and multimodal models let you self-host capable agents for data-sensitive work without per-token fees or sending intellectual property to an API. Scope an on-prem pilot for one workload where data residency or latency rules out hosted endpoints.
xAI
What happened
Grok 4.6 launched with a 500K context window and configurable reasoning effort, and is now available on the Google Enterprise Agent Platform. xAI also launched Grok Bot in beta, always-on AI teammates with a cloud computer per account, and rolled Grok Build, which turns a description into a working app or dashboard live, out to every plan.
What it means for your agentic build
Always-on teammates and per-account cloud computers point toward persistent autonomous coworkers, and Grok 4.6’s presence on Google’s enterprise platform eases procurement. Run a tightly scoped pilot with hard guardrails before granting any standing autonomy.
DeepSeek
What happened
DeepSeek-V4-Pro reached general availability, agent-focused with a 1M-token context, up to 384K output tokens, and OpenAI Responses and Codex compatibility out of the box, but peak output pricing jumped to $3.96 per million from $0.87. On August 21 it unveiled experimental V4-Flash-Vision-Exp, a multimodal model it says nears Anthropic’s Opus 4.8 at far lower cost.
What it means for your agentic build
API compatibility makes V4-Pro a near drop-in, low-cost option for agentic and multimodal workloads, but the roughly 4.5x peak price jump shows cheap models can re-price fast. Budget model costs defensively and keep providers swappable so a price hike does not strand a workflow.
Cohere and Aleph Alpha
What happened
Cohere released Command A+ as open source, a mixture-of-experts model roughly twice as fast as its prior models with lower latency and higher accuracy, while expanding its North agent platform in Europe and passing about $240M ARR. Aleph Alpha now operates inside Cohere following the roughly $20B April acquisition, Europe’s largest sovereign-AI deal, backed by the Canadian and German governments.
What it means for your agentic build
An open, fast enterprise model plus a dedicated agent platform gives regulated buyers a deployable, multilingual option with less vendor lock-in, and the combined sovereign stack pairs frontier capability with regional control. If data sovereignty is a hard requirement, trial Command A+ for a private build and put the Cohere and Aleph Alpha stack on your evaluation list alongside Mistral.
This Week’s Structural Trends
Agents are the product. Nearly every lab shipped agent-first releases this cycle, from Perplexity Brain and Grok Bot to DeepSeek V4-Pro and Meta’s multi-agent Muse, and 500K-to-1M-token context has become baseline rather than a differentiator. The competition is shifting from raw model intelligence to how reliably an agent can plan, use tools and complete multi-step work.
Open weights and sovereignty are back as a wedge. Cohere, Meta, Mistral and Aleph Alpha are competing on openness and regional control, while A2A joining MCP under the Agentic AI Foundation signals converging, neutral interoperability standards. The net effect for buyers is more leverage: more deployable, self-hostable options and less lock-in.
The industry is financially maturing and getting regulated. Anthropic’s first operating profit, OpenAI’s imminent S-1, Cohere’s IPO trajectory, DeepSeek’s price hikes and EU AI Act content marking all point in the same direction. The market is moving from land-grab to durable, compliant enterprise businesses, and procurement should plan accordingly.
Sources
https://www.perplexity.ai/hub/blog
https://openai.com/news/
OpenAI is gaining on Anthropic with business users, new data indicates
https://www.anthropic.com/news
https://ai.google.dev/gemini-api/docs/changelog
https://deepmind.google/blog/
https://www.techi.com/meta-ai-breakthrough-superintelligence-labs-2026/
https://releasebot.io/updates/xai
https://www.bloomberg.com/news/articles/2026-08-21/deepseek-unveils-test-model-to-rival-anthropic-s-opus-4-8
https://mistral.ai/news/regional-inference-open-models-new-compute/