Frontier models keep getting smarter, and agent bills keep growing faster than token prices fall. Here is where the money actually goes, and the layer we are building to manage it.
That layer is Governor. It works on the two forces that set your AI bill at once, the number of tokens each task consumes and the price you pay for every one. Cut both, and the same engineering output costs a fraction of what it did before, with no change to how your developers work.
The cost curve is bending the wrong way
Every model generation lands dramatically more capable, and teams respond rationally by handing agents bigger tasks, longer runs, and more autonomy. The result is a curve that surprises most engineering leaders the first time they see it on an invoice. Price per token falls, while tokens per task grow faster.
The mechanics stay simple and brutal. On every step of a run, an agent reads back its entire transcript so far. A 50 KB file it opened on step 8 still gets processed again on step 80. Cost climbs faster than linearly as the run gets longer. In the runs we instrumented, an unmanaged agent averaged around 75 reasoning steps per task, and most of those steps searched the codebase rather than fixing it.
Engineering leaders have noticed. Gergely Orosz recently reported heads of engineering wishing for an intelligent router that picks the right model for the right task. A first wave of tooling has emerged, OpenRouter’s auto router being one example, and without knowledge of your codebase those tools leave a lot to be desired.
Cost = tokens × price. Routing only touches one factor.
A model router sends easy requests to cheaper models and reserves frontier models for the hard ones. That represents real money. A capable model and a state of the art one can carry a 5 to 10 times price gap per token, and routers typically claim 20 to 30% savings. We expect every serious team to run one.
Routing optimizes price per token and leaves token count untouched, and for coding agents, count is the bigger lever. Most of what an agent consumes goes to the search for where the fix belongs rather than the fix itself. Directory listings, greps, and file dumps ride along in context and get read again on every later step. When we decomposed the savings in our benchmark, the breakdown looked like this.
- 66%, reading back less context on every step
- 22%, writing less new content into context
- 11%, generating fewer tokens
Two thirds of the waste comes from the compounding effect of a bloated transcript. No router touches that, because a router changes the price of the tokens rather than the count.
Governor manages both factors in one layer
Governor is a drop in endpoint that sits between your harness, whether Claude Code, Cursor, or Codex, and your models or your existing gateway. You point your base URL at it, and nothing else about your workflow changes.
AI Architect, fewer tokens. A continuously updated map of your codebases, covering architecture, components, and dependencies, served to the agent so it skips the discovery phase entirely. Controlled benchmarks show 47% lower token cost per task in aggregate, and up to 68% on individual tasks.
Model Router, cheaper tokens. Each request goes to the model that can actually handle it. Frontier models step in where they pay for themselves, and capable, cheaper models cover everywhere else.
Coming next, prompt compression, KV cache management, and more, with each one adding another multiplier on the same bill.
The two factors multiply. Halve the tokens and route the remainder intelligently, and the same engineering output costs a fraction of what it did before, with no change to how your developers work.
What the numbers look like on real customer data
We ran a controlled A/B on a customer’s own tasks, using the same set of real engineering tasks and the same agent, Claude Code, with and without Governor’s AI Architect connected. Every arm completed every task, so the comparison measures efficiency rather than whether the work gets done.
| Metric | Claude Code alone | With Governor |
|---|---|---|
| Avg cost per task | $4.12 | $2.14 (−48%) |
| Avg time per task | 6m 34s | 4m 14s (−36%) |
| Avg agent turns | 47.6 | 23.6 (−50%) |
| Avg tool calls | 46.6 | 21.8 (−53%) |
| Task success rate | 100% | 100% |
Controlled A/B benchmark on customer tasks, 2026. Same tasks, same model, same harness, and the only variable is Governor.
The fix the agent ships stays identical in both arms. What collapses is everything it does before it can start, the hunt for where the change belongs, and the flailing when it cannot find it.
Why this layer, why now
Both ends of your AI stack stay in motion. Models leapfrog each other every few months, and harnesses ship weekly. The one stable place to manage cost, quality, and policy is the layer between them, and most organizations lack that layer today.
Routing will likely become a standard feature of that layer. The deeper savings live in the token count itself, and that requires something harder to build, an understanding of your code rather than your traffic alone. Governor is that smart layer between your harness and your models, the one that makes every token count, starting with the two biggest levers and adding more from there.
See it on your own data
If your agent bill grows faster than your team’s output, we will run this same A/B benchmark on your tasks and your codebase, same agent, same work, with and without Governor.
