The Chatbot Stack Is Splitting: Cheaper Models, Deeper Connectors, AI That Interviews Your Customers

Three stories broke within the same week, and none of them is about a chatbot getting a better personality. Anthropic cut the cost of running its best model by up to 45% and gave it memory that survives across sessions. Semrush and Adobe each wired their own proprietary data and tools directly into chat through the Model Context Protocol. And a consumer-research startup raised $50 million because its AI agent can conduct a customer interview as well as a human moderator can. Read together, they trace where the money and the engineering effort are actually going: not into a smarter reply, but into a cheaper model, a deeper connector, and a chatbot that does a job no chatbot used to do.

1. The Model Layer Gets Cheaper and Starts Remembering

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1 — the same underlying model shipped at two safeguard levels, with Mythos reserved for vetted cybersecurity and life-sciences organizations that need capabilities the general-release Fable keeps constrained. Both support a million-token context window, but the number that matters for anyone running these models in production is the cache pricing: a cache hit on Fable 5.1 dropped to $0.25 per million input tokens, down from $1.00 on Fable 5, cutting a typical workload's cost by roughly 25% and a heavily agentic one — the kind that re-reads the same context dozens of times per task — by closer to 45%. Alongside the price cut, Claude picked up memory that now works across both chat and Claude Cowork, with editable topics and a sensitive-topics setting so a user can see and prune what the model has retained rather than treating memory as an opaque black box. Effort can also be adjusted mid-conversation instead of forcing a restart. None of this is a benchmark win; it's Anthropic making the model cheaper to run at scale and more useful to run continuously — the two prerequisites for treating a chatbot as infrastructure rather than a novelty.

2. MCP Turns Vertical SaaS Into a Chat Connector

The second story is about who gets to plug into that infrastructure. Semrush launched a native Claude connector this week giving eligible subscribers conversational access to 28.8 billion keywords, 43 trillion backlinks, and 317 million logged LLM prompts — SEO and search-visibility data that previously lived behind Semrush's own dashboard, now queryable from inside a chat thread. Adobe shipped something structurally similar aimed at a different workflow: Adobe for Slack, a Slackbot wired through MCP that takes a plain-language request in a thread and routes it to the right Adobe application, turning whatever material is in the channel into a PDF, an image, or a video without anyone opening a separate app. Neither is a chatbot with new knowledge baked into its weights — both are the same pattern Anthropic's own Salesforce integration established weeks earlier, now spreading to specialist vendors: expose your product's data and actions as an MCP server, and let a general-purpose assistant become the front door to it. For a company sitting on data too narrow to justify training into a frontier model, that's a faster and cheaper path to an AI-native product than building your own.

3. A Chatbot That Interviews People Instead of Answering Them

The third story runs in the opposite direction from the first two — not cheaper infrastructure or a new connector, but a new job description. Conveo, which uses AI agents to conduct in-depth customer interviews in 15 languages and mine the transcripts for market-research insight, raised a $50 million Series A led by DST Global. Its "StoryLines" product runs recurring AI-moderated video interviews on a schedule a brand sets — weekly or monthly checks on brand health, ad creative, or a new product launch — and more than 400 enterprises, including several Fortune 500 names, are already paying for it. The interesting part isn't the funding size; it's that the chatbot here isn't answering a support ticket or drafting an email. It's asking the questions, following up on a vague answer the way a trained moderator would, and turning that conversation into structured signal. That's a genuinely different use of conversational AI than the support-desk and copilot framing most of the industry still defaults to.

What Ties These Together

A commoditizing model layer, connectors that let any vertical plug in without training its own model, and a chatbot doing qualitative research instead of customer support are three ends of the same rope: as the base model gets cheap and standardized, the differentiation moves to what it's connected to and what task it's been pointed at.

// The shape every connector-based integration shares: the assistant
// doesn't own the data or the interview, it orchestrates access to both.
interface McpConnector {
  name: string;
  query: (question: string) => Promise<Record<string, unknown>>;
}

async function answerFromConnector(connector: McpConnector, question: string) {
  const data = await connector.query(question);
  return assistant.synthesize(question, data); // reasoning stays generic, data stays domain-specific
}

If you're deciding where to spend engineering time this quarter, the honest answer these three stories point to is: probably not on the model call itself. It's on the connector that gets your own data into someone else's chat surface, and on finding the task in your product that a conversation is actually better suited for than a form.

Suggested visuals: a pricing chart comparing Fable 5 vs. Fable 5.1 cache-read cost per million tokens; a simple diagram showing an MCP connector as a bridge between a vertical SaaS product's data and a general-purpose chat assistant; and a short annotated transcript excerpt showing a Conveo-style AI follow-up question next to a vague consumer answer.

— Maya