There is a version of developer loyalty that runs on API keys. It expires when the next lab drops pricing by 40% and the migration cost is a few days. There is another version that runs on architecture, the kind where you have written 10,000 lines of TypeScript against a platform's event model and the switching cost is a quarter, not an afternoon.
Both bets surfaced this week. OpenAI threw 25 announcements at DevDay on September 29 and the one that mattered most was Dots: always-on agents running on OpenAI's cloud computers, routed through OpenAI's 4,000-plus plugins, persistent and background and tied to your Pro or Business Premium plan. Developer loyalty through infrastructure. Get in the system; the system does not travel.
The day after ChatGPT Finances rolled out to Free users, Anthropic shipped Claude Code Mods. A TypeScript function. A hook that fires on every internal event in the agent loop. Intercept a tool call. Rewrite a prompt before it goes out. Strip a secret before Claude reads it. The whole agent, rewritable, from inside.
Both strategies work. They are not the same strategy.
Claude Code Mods shipped October 1 in v2.1.287, then picked up the "You should know" watcher in v2.1.288 on October 3. The release note reads like a minor feature drop. It is not.
Here is the mechanism. Every time Claude Code fires an internal event, a tool call, a prompt dispatch, an interface element being drawn, it now calls any mod functions registered for that event. A mod is a TypeScript module that ships inside a plugin, installed with /plugin. The mod can observe the event and return nothing. It can rewrite the event and return a modified version. It can block the event entirely. It can replace the behavior with something the original code never anticipated.
Three official examples ship with the release. Token Weather puts a live context forecast above your input prompt. Blast Radius dry-runs dangerous shell commands and shows you what they would delete before Claude executes them. Replay Theater steps through an edit session one change at a time. Each one is a TypeScript module of roughly 30 lines that intercepts a specific event in the loop.
The blast radius is every enterprise team running Claude Code against internal codebases. The immediate use case is compliance: a mod that strips API keys, tokens, or PII from tool output before Claude reads it, before it goes in-context. Until this week, that required a proxy, an extra layer of infrastructure, and a maintenance burden. Now it is a 20-line function installed with one command.
The pattern: Anthropic launched the base Mods system on October 1 and followed four days later with the first built-in mod, "You should know," a side agent that watches your session and flags things Claude might miss or that you would want to know before committing. The extensibility layer is shipping in layers, each one opening a new surface for the ecosystem to build on.
The read: Anthropic is not building a cloud platform for you to trust. It is shipping a runtime you can modify. OpenAI's Agents API puts your workloads on OpenAI's infrastructure and manages the execution for you. Claude Code Mods puts the modification surface in your hands and your repository. One model is a service. The other is a tool. The bet underneath Mods is that engineers who own their agent's behavior stay, not because there is no alternative, but because they have invested in it.
The builder's move: Update Claude Code to v2.1.288 or newer. The official examples are on GitHub under the claude-code-mod topic. Start with the secrets-redaction pattern: fire on tool output, match against your org's secret formats, return the redacted output. You get the compliance behavior without a proxy layer between Claude and your codebase.
OpenAI DevDay ran September 29 in Fort Mason, San Francisco. 25 announcements. The two that matter for builders hitting production this week: Dots and the Agents API public beta.
Dots are always-on agents. Each one runs on GPT-6 Astra with its own cloud computer and browser, 4,000-plus plugins pre-wired, and persistence. It does not stop when you close a tab. You assign a goal; it works in the background, across apps, without re-prompting. Pro and Business Premium subscribers in eligible markets get their first Dot included; the rollout excludes the EEA, Switzerland, and the UK for now.
GPT-6.1 Sol is live in the API as gpt-6.1-sol: near-Astra intelligence at one-fifth the Astra token price. Context window: 1,050,000 tokens. A premium Ultrafast tier generates up to 300 tokens per second in Codex. The Decisions API, powered by a smaller Luna model, is in limited preview for routing and classification workloads.
The cross-lab read: OpenAI is building the operating system. Dots are the running processes. GPT-6.1 Sol is the cost-optimized worker tier. The Agents API is the SDK that lets developers run workloads on that OS without managing infrastructure. This is a coherent architecture. The question is whether developers want to run their agents on someone else's machine.
Anthropic committed $100 million to train 10,000 "Frontier Deployed Engineers" by the end of 2027. First cohorts are running now in San Francisco, New York, and London, drawn from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk, and others. The program structure: in-person graded training on a simulated enterprise deployment, then a 12-week live project inside the engineer's own organization, with Anthropic engineers alongside.
The target is not the developer writing API calls. It is the consultant layer sitting between Anthropic's models and a Fortune 500 deployment. Train that person well and you have seeded your own distribution network in every major client engagement on the planet.
The cross-lab contrast: OpenAI trained developers at DevDay through talks and documentation and sample code. Anthropic is training the implementation layer. OpenAI's DevDay reaches ten thousand developers in a day. Anthropic's Academy graduates two hundred a month for three years. The ROI calculation matters in 2028, not 2026.
ChatGPT Finances expanded to Free and Go users October 2, two weeks after the Business and Enterprise rollout. Users connect financial accounts, get spending analysis and savings suggestions. Opt-in; no account data is stored beyond the session. This is OpenAI's second consumer finance push in a month, alongside the Try-On virtual clothing feature, suggesting a deliberate move into everyday utility rather than just productivity tools.
xAI deprecated grok-voice-transcribe-1.0 on October 2. All requests to that model string now route to grok-voice-transcribe-2.0 at the same price, with higher accuracy. No action required unless you pinned the model string in your config.
Meta Connect 2026 ran September 23-24. The Ray-Ban Meta Audio glasses, camera-free with a 12-hour battery and Muse AI, ship October 13 at $349. The Ray-Ban Meta Gen 3 (12MP 3K camera, six-mic array) is available now at $449. Muse Spark 1.3, the September update, runs across the full glasses line. Watch for a software push that brings calendar and email integrations to the glasses before end of October.
Mistral closed a 3 billion euro Series D in September, led by Samsung, at a 21 billion euro post-money valuation. The largest equity raise in European tech history. The capital goes to training and inference infrastructure. The Magistral reasoning model shipped in September; watch for a Ministral edge deployment push as the compute financing converts to capacity.
Google DeepMind: Demis Hassabis moved to Chair and Chief Scientist of Alphabet in August. Koray Kavukcuoglu is now SVP of Google DeepMind, running Gemini development directly. The first full product cycle under Koray's leadership is visible by Q1 2027.
n:<text> filter to the agents view that matches session names and tasks.claude update). The "You should know" mod activates automatically. The agent view filter is available in the agents panel./diff) now ship as mods and can be turned off or replaced./plugin. Official examples at github.com/topics/claude-code-mod.gpt-6.1-sol. Available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users.grok-voice-transcribe-2.0.Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.