Claude Fable 5 in Cogny: The Case for a Multi-Tier AI Model Strategy

Today we're shipping something we've been genuinely excited about: Claude Fable 5, Anthropic's newest frontier model, is now live in Cogny for all Cloud customers. No migration, no settings to change — your deepest analysis work is already running on it.

Back in March, when news first leaked that Anthropic was testing a step-change model, we wrote that the teams who benefit most from each new model generation are the ones whose data is structured and ready for it. That moment is here, and the early results on real customer workloads are exactly what we hoped for: deeper cross-channel reasoning, longer analysis chains that don't lose the thread, and noticeably sharper recommendations.

But the more interesting story isn't "we added a new model." It's how we added it — and why we think the architecture behind it is the one every serious AI product will converge on.

One Model for Everything Is the Wrong Architecture

There's a tempting simplicity to picking "the best model" and using it for everything. It's also wrong, for the same reason you don't send every task in your company to your most senior strategist.

AI workloads in a product like Cogny span an enormous range:

A weekly cross-channel audit that joins paid media, organic search, and revenue data, reasons over a quarter of history, and produces a prioritized action plan. This is high-stakes work. It can take minutes. Nobody is staring at a spinner — it lands in your inbox.
A quick lookup in chat — "what did we spend on Meta last week?" — where the user is waiting and the right answer is a number, fast.

These are different jobs with different success criteria. The audit should be best effort: take the time, burn the tokens, get it right, because a single good budget reallocation pays for a year of the product. The lookup should be instant: a frontier model would give the same answer, just slower and at many times the cost.

So Cogny doesn't run on one model. It runs on a tiered model strategy:

Frontier tier — now Claude Fable 5. The most important, most complex work: scheduled deep-dive audits, cross-channel performance analysis, budget forecasting, the reasoning-heavy reports our customers make decisions on. These tasks are allowed to be slow because the output quality is the entire point.
Fast tiers — smaller, quicker Claude models. Interactive chat, simple retrievals, summarization, classification, and the high-volume glue work that happens thousands of times a day. These tasks are allowed to be simpler because latency is the user experience.

Every task gets routed to the cheapest model that can do it excellently — and the hardest tasks get routed to the best model that exists.

Why This Optimizes Cost and Performance (Not a Trade-Off)

The usual framing is that you trade quality against cost. A tiered strategy breaks that trade-off, because the two curves optimize different things:

Performance where it's felt. Users feel latency in interactive flows and quality in analytical output. By putting fast models where people wait and frontier models where people decide, both experiences improve simultaneously. Your chat got snappier. Your Monday-morning audit got smarter. Same product, same week.

Cost where it compounds. Frontier models are priced like frontier models. If you run one on every trivial call, you either eat the margin or pass the cost to customers. By reserving frontier capacity for the work that actually needs it, we can afford to be generous with it — letting Fable 5 run long, multi-step analyses without cutting corners — precisely because we're not wasting it on tasks a faster model handles just as well.

Upgrades become free wins. This is the part we love most. Because tasks are routed by tier rather than hardwired to a model, swapping in Fable 5 was a routing decision, not a rewrite. When the next model generation lands, the same is true. Your accumulated data, context, and experiment history — the stuff Cogny has been helping you structure all along — instantly gets a smarter brain on top of it.

What Fable 5 Changes in Practice

Frontier-tier work in Cogny is exactly where Fable 5's strengths land:

Longer reasoning chains that hold together. Multi-phase audits — pull the data, find the anomalies, explain the causes, propose the experiments — stay coherent across more steps. Fewer "lost the plot halfway through" moments, more analyses that read like a senior analyst wrote them.
Better tool use over real data. Cogny's agents write and run real queries against your warehouse and your marketing platforms via MCP. Fable 5 is measurably better at getting complex queries right and building on intermediate results.
Sharper synthesis across channels. The non-obvious connections — paid search cannibalizing organic, a creative fatigue pattern hiding inside a healthy ROAS — are exactly the kind of cross-domain reasoning that improves most with each frontier generation.

If you're a Cloud customer, you don't need to do anything. Your scheduled reports and deep analyses are already running on Fable 5 as of today.

The Takeaway for Anyone Building (or Buying) AI Products

When you evaluate an AI product, don't just ask which model it uses. Ask how it decides which model to use. A product that answers "we use [frontier model] for everything" is either slow and expensive, or quietly cutting the depth of its analysis to keep response times tolerable. A product with a deliberate tier strategy can give you frontier-quality thinking on the decisions that matter and instant responses everywhere else.

That's the bet we made when we built Cogny on Claude — a bet we've written about before — and days like today are why. The models keep getting better. Our job is to make sure that every time they do, the improvement shows up in your reports the same week, with zero effort on your side.

Fable 5 is live. The vibes, as the team keeps saying in Slack, are very good.

Want to see what a frontier model finds in your marketing data? Get started with Cogny — connect your channels, and let it run.