AI Coding Assistants

Today's Stories

1. 1

   SpaceX, AT&T, Apple steer into AI's currents—Grok 4.6, open models, DC lobbying

   SpaceX released Grok 4.6, benchmarking against Anthropic's latest models and OpenAI's GPT 5.6; AT&T routes 45 billion daily tokens across open and closed models via a smart router, with open models now powering 25% of use and expected to reach 70–80% over time; Apple hired Nate Gatten, a former American Airlines and JPMorgan government affairs veteran with Republican ties, to lead its global government affairs operation as John Ternus becomes CEO on September 1. Elon's $60 billion acquisition of Cursor, the AI-native coding company, is validated by Grok 4.6's performance; AT&T's shift demonstrates enterprises are building portfolios of both open and closed models rather than locking into a single vendor, with switching to open models already saving 80–90% in certain applications; Apple's government affairs restructuring—with Tim Cook moving to executive chairman while managing White House relationships—signals Big Tech is treating DC lobbying as a first-team priority ahead of tariffs, chip sourcing, and the September Xi-Trump White House meeting.

   How the market assesses Grok 4.6's benchmarks in coming days; AT&T's projection that open models will power 70–80% of its AI workload over time; Apple's Pixel 11 phones reducing RAM from last year's 16 gigabytes to 12 despite price increases of $100 on phones and $50 on wearables, signaling that memory pricing pressure is persisting across even the largest tech companies.

2. 2

   SpaceX closes $60 billion Cursor acquisition

   SpaceX has officially completed its acquisition of AI coding startup Cursor for $60 billion, after announcing the deal in April and moving forward with it in June as SpaceX went public. Cursor gains access to what it describes as the world's largest GPU fleet, housed within SpaceX's computing infrastructure—the same infrastructure SpaceX rents to customers including Anthropic and Google. The deal positions the combined entity to scale AI capabilities using SpaceX's data center capacity.

   SpaceX is simultaneously managing a lawsuit over pollution from its data center gas turbines, even as it expands its computing footprint to support AI development.

3. 3

   Snowflake Observe adds CLI and redesigned MCP for AI agents to access telemetry directly

   Snowflake announced general availability of a redesigned Observe MCP server and a new Observe CLI that provide full parity—every operation in the MCP server is also available as a CLI command. The rollout is happening now across all clusters, with eu-2 and ca-1 regions to follow shortly. The new tools let AI agents and engineers query Observe's telemetry data directly through APIs, the MCP (model context protocol), and command line, rather than only through the user interface. Observability platforms now need to serve both humans and AI agents. The redesigned architecture removes an LLM intermediary layer, cutting latency and operational cost. Engineers are already using these tools to build custom alert-triage agents that automatically investigate alerts and pull production telemetry for debugging, and copilots that assist during incident investigations—capabilities that require direct, programmatic access to observability data.

   Both the CLI and MCP server ship with prebuilt skills for common observability workflows (investigating incidents, tracing failures, validating changes, finding n+1 issues, detecting outliers). The new MCP gives agents direct access to APM service maps, OpenTelemetry data collection setup, and active alert listings. Setup is available now in the Observe UI under "Manage account" → "MCP server details."

4. 4

   Deltix launches AI testing for iOS apps on Mac simulators

   Deltix, an AI-driven testing tool, has entered open beta. It lets developers describe tasks in plain English, then runs them on an iOS Simulator to check whether a real user could complete them. The agent observes the screen like a user would, records screenshots and run logs to your account, and operates locally on your Mac—your source code and builds never leave your device. Developers can now quickly verify whether new flows work for real users without scheduling research sessions or waiting for A/B tests in production. Tasks are saved as Playbooks and replayed deterministically on every build, turning ad-hoc testing into reproducible checks integrated into development workflows. For teams building iOS apps, this can catch usability blockers before they reach users.

   The service is free during open beta with no credit card or invite required. Physical device support (real iPhones attached to your Mac) and Android emulator support are in progress. CLI integration for CI/CD (GitHub Actions, GitLab, CircleCI) and support for React Native and Flutter hybrid apps are planned for later.

5. 5

   Z.ai's GLM-5.3 matches frontier models with 750B parameters

   Z.ai announced GLM-5.3, a model with ~750B parameters that has surpassed Moonshot AI's Kimi K3 and matched or surpassed Claude Fable 5 or GPT-5.6-Sol on many benchmarks. The model is currently available only in the coding plan, coming to the API soon and to Hugging Face (open weights) in two weeks' time. Z.ai built GLM-5.3 by extending post-training on the same base model as GLM-5.2, with no change to the underlying architecture. Z.ai achieves frontier-level agentic coding performance with roughly a third of Kimi K3's parameters, raising questions about how Chinese labs sustain competitive parity despite American companies' commanding resource lead. The article suggests speed of release—days rather than months—is the primary advantage: Chinese labs use the time American labs spend on pre-release testing to continue benchmarking optimization. As model self-improvement loops rely more on user data, faster release cycles could give Chinese models longer practical lifespans before displacement.

   Z.ai is taking a staged approach to GLM-5.3's availability due to its cybersecurity capabilities. Selected security partners will evaluate the model in controlled settings first; broader API access and full model weights will follow once safety evaluations are complete. The article notes that true open-weights release will eventually make staged safety measures less effective, as capability diffusion tends to follow the lowest common denominator across all labs.