Fuli Luo said MiMo cut cache-hit input pricing by up to 99% after hierarchical KV-cache optimization for SWA increased cached token capacity by 5x. Alongside DeepSeek's price pressure, the bigger signal is not just a model price war. It is that architecture choices are starting to decide how cheaply long-context and agent workloads can run at scale.

OpenAI Developers announced Secure MCP Tunnel, a way for teams to keep MCP servers inside private networks while ChatGPT, Codex, the Responses API, and AgentKit connect through outbound-only HTTPS. The important part is not simply that OpenAI added another connector. It is the infrastructure layer companies need if agents are going to reach internal data and tools without exposing public endpoints or changing inbound firewall rules.

The Information estimates that Anthropic may be generating at least 35% more revenue than OpenAI, driven by strong growth in coding and enterprise demand. Because the source is paywalled and only partially accessible, this should stay explicitly attributed: it is The Information's estimate, not a publicly confirmed independent figure.

Elon Musk says SpaceX is writing a bare-metal training stack for Grok: Musk said SpaceX is nearly done with V1.0 of an AI training stack written in C, mapped directly to 220k GB300s with 800G NICs, and potentially more than 10x faster than JAX for large training runs. It is still a self-reported claim, but it points to an important direction: at extreme scale, labs may want to verticalize deeper into their own training infrastructure.

DeepSWE pushes coding agents into long-horizon tasks: DeepSWE introduces a long-horizon software-engineering benchmark built from original repository tasks, showing clearer gaps between frontier coding agents than saturated public leaderboards. For builders, the point is not today's winner. It is which benchmarks are starting to look more like real work.

Robinhood lets AI agents trade stocks: Robinhood is beta testing agentic stock trading with separate agent accounts, dedicated balances, notifications, approvals, and fraud controls. This is a notable move from agents-as-advisers toward agents-with-bounded-execution in a highly sensitive domain.

Cognition raises more than $1B for Devin: TechCrunch reports Cognition raised more than $1B at roughly a $25B pre-money valuation, with a $492M ARR run-rate. Coding agents are no longer being judged only by demos or benchmarks. They are starting to be valued like software businesses with visible revenue traction.

Claude Marketplace expands procurement for Claude-powered tools: Claude announced Marketplace partners including Augment Code, Bolt, CodeRabbit, Hebbia, and Legora, letting enterprises apply existing Anthropic spend commitments toward partner tools. Enterprise AI distribution is shifting from buying tools one by one toward marketplace and procurement layers built around model spend.

Sesame launches an iOS preview for personal voice agents: Sesame introduced an iOS preview for lifelike conversation partners with memory across voice and text. Voice agents are moving from "sounds human" demos into mobile distribution, where memory and search can become core product primitives.

OpenAI shared a Tax AI case where Codex helped turn practitioner corrections and production traces into eval targets, then into scoped engineering tasks. The pilot processed 7,000 returns and saved practitioners about a third of preparation time.

Do not use an AI agent only to "do the task."
Use it to tighten the learning loop from real errors.

A practical loop:

Score the loop by three checks:

The useful part is not just that the agent generates output faster. It is that production feedback can become system improvement when teams package corrections into evals and regressions.

Is your growth in-line with your peers in B2B SaaS & AI? 

Benchmark yourself against actual billings data for Maxio’s 2000+ global customers, alongside firsthand company perspectives to understand how growth varied by company size, business model, and strategic focus.

Key takeaways from the report: 

Download the report 

Axios frames a recent news cycle as a snapshot of the entire AI stack: model progress, revenue, infrastructure, markets, and government pressure all heating up at once. It is worth reading if you want to understand why AI news no longer lives in a single lane.

Anthropic explains agent containment through sandboxes, VMs, filesystem boundaries, egress controls, and blast-radius limits. This is the unglamorous but essential part of agent adoption: the more tool access an agent gets, the clearer its boundaries need to be.

TechCrunch uses Aaron Levie's "AI psychosis" framing to argue that some CEOs may be too optimistic about automation because they are far from last-mile review, bugs, contract nuance, and operational bottlenecks. It is a useful counterweight to today's agent optimism.

Forbes frames AI as a sector that is growing up: bigger companies, IPOs getting closer, and public investors preparing to scrutinize revenue, margins, infrastructure costs, and concentration risk. It is a useful lens for Cognition, OpenAI/Anthropic revenue, and the capital-market story around AI.

Axios cites Gallup/Walton Family Foundation findings that many K-12 teachers have little or no formal guidance on AI use, even as AI starts reshaping classroom work and student expectations. It is the education version of the same pattern: adoption is outrunning governance.