With generic prompts, frontier models averaged 47.2% accuracy filtering financial documents for Bridgewater’s analysts. A coin flip. With the best prompts an expert could write, that accuracy climbed to 78.2%. The threshold Bridgewater’s investors required before trusting the filter inside a real workflow was 80%. No frontier model, with any prompt, crossed it. And GPT-5.4 cost 43% more than GPT-5.2 without meaningfully improving accuracy.

Bridgewater didn’t wait for a smarter model to ship. Mira Murati announced it plainly: “Bridgewater used their unique financial knowledge and partnered with us on @tinkerapi to fine-tune a model that helps their analysts focus on what’s important. Experts improving AI that empowers experts.” They fine-tuned a Qwen3-235B model, Alibaba’s open-weight model, and reached 84.7% accuracy: 29.8% fewer mistakes than the best frontier model, at 13.8x lower inference cost.

How Bridgewater beat the most expensive models on the market

Thinking Machines’ technical writeup breaks down three pieces that stacked into the improvement: interleaved batching (+12.1 points), a CISPO loss function with asymmetric clipping (+10.1 points), and on-policy distillation with teacher models that get dynamically promoted during training (+3.1 points), all running on GRPO as the reinforcement-learning backbone. None of those three pieces is a smarter model. They’re adjustments to how training happens, not to how big the model is.

The part that actually generalizes to any company sitting on its own data isn’t the training technique. It’s how they handled the fact that their own labeled data was noisy. They trained a first model on the labels they already had, then ran that same model back over its own training set. Every time the model disagreed with a label, that example got routed to a senior investor for review, because either the example was genuinely hard or the original label was wrong. The model’s own confusion became the tool for catching bad labels, instead of paying an expert to manually re-check ten thousand examples one by one.

Anthropic had the exact same problem with its own product

The same week, Anthropic published how it uses Claude internally so its own team can answer data analytics questions without pulling in a human analyst every time. The starting point was 21% accuracy: one in five attempts, Claude got it right. After building a layer of business definitions (what “weekly active user” means exactly at this company, which table is the source of truth for each metric, which analytical workflows are already encoded as reusable steps), accuracy climbed past 95%, and to 99% in some specific domains. InfoQ covered the detail that matters most: Anthropic tested letting a model generate that definition layer automatically from raw tables and logs, and the result was worse than having no layer at all. The definitions had to come from someone who understands the business, not from a model reading a database schema.

It’s the same conclusion Bridgewater reached, by a completely different path. Bridgewater changed the weights of a smaller model using its own examples. Anthropic left the model alone and built a layer of its own context on top. Neither company solved the problem by asking the frontier model to guess better. Aaron Levie put it well this same week, in a line that applies to both cases: the battle in AI is a battle for context, not for which lab has the biggest model.

The threshold decides, not the preference

Neither company made its call out of ideology. Bridgewater had volume: thousands of financial documents its own analysts had already classified over the years, enough to fine-tune a model on real examples, even partially mislabeled ones. Anthropic had something else: an internal business understood well enough to write down, precisely, what each metric means and where each piece of data lives. Neither recipe works without the input that feeds it.

What the two stories share is the starting diagnosis: calling the frontier model’s API with the smartest prompt possible has a ceiling, and that ceiling is measurable. 78.2% for Bridgewater. 21% for Anthropic with no context layer. Both companies measured the ceiling before deciding what to build on top of it, instead of assuming the next model version would solve it on its own.

What we do at IQ Source

Most companies we talk to never measure that ceiling. They try a prompt, decide it “sort of works,” and ship it to production without knowing whether they’re sitting at 47% or 84%. In the discovery phase of AI Maestro, we measure that number for every candidate process before recommending anything: how well a frontier model with the best possible prompt actually handles the task, against the accuracy threshold the business genuinely needs to trust the output without constant human review.

When the frontier model already clears that threshold, there’s nothing to build. Using it directly is the right call, and the cheapest one. When it doesn’t, the question stops being which model is smarter and becomes which of two paths applies: if there’s volume of proprietary examples, fine-tune a smaller model the way Bridgewater did; if what’s missing is well-defined business context, build that layer the way Anthropic did. I’ve argued before that charging for the outcome forces the change to actually happen: that same Go/No-Go gate is what decides whether a process is ready to automate, or needs one of these two paths first.

No lab is going to tell you what percentage your process sits at today. That measurement is yours to take, and it’s the first number you should know before signing any AI contract.

Measure your process’s real ceiling before you automate it