Agents get real permissions while their costs stay hard to predict

Anthropic paused token-based billing for its Claude Agent SDK as Robinhood opened agentic trading to all customers through an MCP server. Agents are reaching deeper into money, phones and browsers, while their costs and runtime demands remain hard to predict.

Agent costs resist per-token pricing

Anthropic is pausing token-metered billing for the Claude Agent SDK. The shift appears to reflect how hard agent workloads are to forecast: one task can branch into long chains of tool calls and reasoning steps, which makes per-token spend difficult to budget. Teams that planned agent costs around chat-completion assumptions may need to revisit those estimates.

Sunil Pai's post on cutting wasted context, redundant tool calls and bloated prompts treats efficiency as an engineering discipline, not an afterthought. VentureBeat reaches a related conclusion at the organizational level, arguing that enterprises moving agents into production are limited by orchestration, observability, state and governance rather than model choice. If reliability failures are architectural, a stronger model alone is unlikely to fix them.

Agents gain access to money, phones and browsers

Robinhood's MCP server lets AI assistants work directly with brokerage functions for its entire customer base, which takes the protocol out of developer tooling and into a regulated consumer product. On the platform side, Android 17 gives developers deeper system-level hooks for agents that operate across apps instead of staying inside a single app's interface. OpenAI has also added Chrome DevTools Protocol support for browser use in Codex, extending its coding agent into the browser.

Broader access raises the question of who an agent is acting as. Okta is expanding its Google Cloud partnership to secure AI agents and Chrome, built on the idea that agents need identities, permissions and audit trails separate from the human sessions they act for. Integrations like Robinhood's trading server show why that separation is becoming a practical concern rather than a theoretical one.

Open weights, voice models and where AI runs

Z.ai's GLM-5.2 brings a 1-million-token context window, new reasoning controls and explicit support for long-horizon coding, and the company presents it as one of the strongest open models available. The weights are on Hugging Face for teams that want to self-host or fine-tune, which may make it a practical open option for long-running agent work. OpenAI, meanwhile, is reportedly preparing a ChatGPT voice overhaul built on a new model, GPT-Bidi-1, which suggests a move toward real-time bidirectional conversation instead of turn-taking.

On the training side, Nvidia reports that Blackwell posted category-leading results across MLPerf Training 6.0. Closer to everyday software, Microsoft is testing its small Phi Silica model for Windows AI features on Nvidia GPUs instead of only on NPUs, widening the hardware local inference can use. Google Voice takes the cloud route, now offering Gemini-powered note-taking for calls inside a workplace app.

Memory encryption losses and paid crawler access

AMD removed memory encryption features from its consumer CPUs, drawing complaints from users who depended on encrypted memory. For engineers running commodity desktops, it is a reminder that baseline security features can disappear through product segmentation.

In an unrelated change for site operators, AWS WAF now lets them identify AI crawlers and charge them for content access instead of blocking them outright. Publishers gain an option between allowing all scraping and refusing it entirely.