Visa built an internal AI “harness” that makes Anthropic’s models cheaper to run for cybersecurity work, according to The Information. The reporting frames this as a practical answer to one of the biggest complaints about frontier models in production: the bill. For a payments company defending against fraud and intrusion at massive scale, cost per query isn’t a footnote. It’s the difference between running AI on a few alerts and running it across everything.
Here’s what matters, up front:
- What happened: Visa developed a software layer, its “harness,” that reduces the cost of using Anthropic’s models for cyber defense tasks, The Information reports.
- Who’s involved: Visa on the deployment side, Anthropic as the model provider.
- Why it matters: It’s a real-world example of a large enterprise engineering around token costs rather than just accepting them, or switching to a cheaper, weaker model.
What a ‘harness’ actually does
Think of a harness as the plumbing around the model, not the model itself. It decides what context to send, how to route requests, when to call the expensive model versus a smaller one, and how to cache or reuse work. Done well, it strips out wasted tokens and redundant calls. The model stays the same. The cost of using it drops.
That distinction is the whole story. Visa isn’t claiming Anthropic got cheaper. Visa built the scaffolding that makes Anthropic cheaper in its specific workflow. Any team running models at scale can copy the idea, even if they can’t copy the exact code.
Why this is a big deal for practitioners
The status quo for most companies has been blunt. If a model costs too much, you either cut how often you use it or you drop to a cheaper tier and eat the quality loss. Visa’s approach points to a third path: keep the strong model, but be smart about how you feed it.
What stands out here is the use case. Cyber defense is high-volume and high-stakes. You’re scanning enormous streams of events, most of them noise, hunting for the few that matter. Running a top-tier model on every event would be ruinously expensive. A good harness lets you apply the expensive intelligence only where it earns its keep, and cheaper logic everywhere else.
That’s a template other regulated, high-throughput industries will recognize. Banks, insurers, telecoms, anyone with a firehose of data and a compliance department watching the spend.
The bigger trend
This fits a shift I’ve been watching across the industry. The early race was about raw model capability. The current race is increasingly about efficiency of deployment: caching, routing, context management, and smaller models handling the easy 90% so the frontier model handles the hard 10%.
Anthropic benefits either way. A customer that finds a cheaper way to run Claude at scale is a customer that runs more of it, not less. Cutting the per-query cost tends to expand usage, not shrink revenue. That’s the quiet incentive behind why providers increasingly ship tools like prompt caching and batch pricing themselves.
What to expect next
- Efficiency becomes a competitive edge. Expect more enterprises to talk publicly about the tooling around their models, not just which model they picked.
- Cost engineering moves in-house. Teams that treat token spend as a fixed tax will fall behind teams that treat it as an optimization problem.
- Security is a proving ground. If AI cyber defense pencils out at scale for a company as large as Visa, other high-volume detection use cases will follow.
The lesson for anyone building with frontier models: the model you choose is only half the decision. How you wrap it, route it, and feed it can move your costs more than swapping vendors ever will.
Full details are in The Information’s report.