A Fracturing Market, a Talking Music App, and an Agent That Never Closes

Three stories from the past two weeks aren't about a single new model or law — they're about where the chatbot layer is drifting: away from one dominant vendor, away from the open web, and away from the chat window itself. Here's what happened, and what each one is worth checking against your own stack.

1. The Single-Vendor Bet Stopped Paying Off

A year ago, ChatGPT held roughly 76–79% of worldwide web visits to consumer generative-AI chatbots. As of this month it's down to around 53–54%, with Google's Gemini climbing from single digits to 27.9% and Anthropic's Claude nearly tripling in a single quarter to 9.2%. None of this is one company stumbling — it's three labs shipping genuinely competitive models on overlapping timelines, plus enterprises that spent 2025 learning the cost of being locked to one provider's pricing, rate limits, and outage windows.

The practical shift is that "pick a model" is no longer a one-time decision. Teams that hard-coded a single vendor's SDK into their product are now rewriting request/response glue they thought was a solved problem, while teams that abstracted the model call behind a stable interface are swapping providers in an afternoon. If you haven't already, that abstraction is worth building before you need it, not after a price increase or an outage forces the question:

interface ChatProvider {
  stream(messages: ChatMessage[], opts: ChatOptions): AsyncIterable<ChatChunk>;
}

const providers: Record<string, ChatProvider> = {
  gemini: new GeminiProvider(),
  claude: new ClaudeProvider(),
  openai: new OpenAiProvider(),
};

function getProvider(name = process.env.DEFAULT_CHAT_PROVIDER ?? 'gemini'): ChatProvider {
  return providers[name] ?? providers.gemini;
}

A market this contested rewards whoever can act on the next price cut or benchmark win fastest — and that's an architecture property, not a model-selection skill.

2. Conversational AI Is Becoming a Feature, Not a Destination

Spotify began rolling out "Talk to Spotify" in beta on July 14 — a text-and-voice assistant, currently limited to Premium subscribers 18 and over in the US, Ireland, and Sweden, that holds context across follow-up questions to recommend music, build playlists, queue tracks, and answer questions about a listener's own history or a song's backstory. It joined Google's AI Mode adding grocery ordering and design generation inside search results, and a wider pattern of general apps bolting on a chat surface scoped tightly to what that app already does.

What's notable isn't the chat UI — every app has one now — it's the scope. Talk to Spotify doesn't browse the web or touch anything outside Spotify's own catalog and a user's own library; the assistant is a narrow front end onto a fixed, pre-existing action set, not a general-purpose agent wearing a Spotify skin. That's a much smaller trust and safety surface than a general assistant, and it's achievable specifically because the product scopes the tool list down to what the domain actually needs:

const SPOTIFY_ASSISTANT_TOOLS = [
  'searchCatalog',
  'createPlaylist',
  'queueTrack',
  'getListeningHistory',
  'getTrackMetadata',
] as const;
// No fetchUrl, no sendEmail, no generic code execution — the tool list
// is the entire attack surface, and it maps one-to-one to product features.

If your product is adding a conversational layer to something that already has a defined feature set, resist reaching for a general agent framework by default — a narrow, explicit tool allowlist is both safer and usually a better product fit than an open-ended one.

3. The Agent Is Leaving the Chat Window Entirely

Google's Gemini Spark — the always-on assistant built on Gemini 3.5 and its Antigravity agent harness, first shown at I/O in May — expanded from a browser-based feature to a standalone Mac app on July 1. Spark runs on dedicated cloud VMs rather than a device, so it keeps working on multi-step tasks (summarizing long email threads, watching a Gmail inbox for deadlines, flagging hidden fees in a recurring bill) after a user closes the laptop, and it takes new assignments by email rather than requiring the user to open an app at all. Voice agents are following the same pattern from a different direction: production voice-agent deployments grew 340% year over year, funding for the category hit $2.1 billion, and voice agents now field an estimated 40% of enterprise calls — none of which happens inside a chat window either.

Both point to the same underlying shift: the text box was never the essential part of a chatbot, it was just the first interface that happened to fit the technology. As agents get reliable enough to run unsupervised for longer stretches, the interaction model is moving toward whatever channel a task already lives in — email, a phone line, a background job — with the chat window becoming one entry point among several rather than the product itself.

What This Means for Builders

None of these three shifts changes what a chat completion API does — they change what surrounds it. Put a provider-agnostic interface between your product and whichever model you call today, because the vendor that wins this quarter's benchmark won't necessarily win next quarter's price war. If you're adding conversational AI to an existing product, scope its tool access to that product's own feature set instead of reaching for a general-purpose agent framework by default. And if your roadmap only has a chat window as an entry point, look at whether the task your users actually want done would be better served meeting them in email, voice, or a recurring background job — the box on the page is a UI choice, not the product.

— Maya