Introducing Governor, a context-driven model router for coding agents Bito Ai
Your coding agent bill, cut in half
Technical design in hours, not days 

Introducing Governor, a context-driven model router for coding agents 

Table of Contents

Agentic coding has taken off, and token use has skyrocketed with it. Much of that spend goes to finding things in your code rather than writing it. In our measurements, 60% to 80% of what a coding agent costs goes to searching and reading its way around your codebase, and nearly all of it is billed at frontier-model prices. 

Today we are launching Governor, a context-driven model router for coding agents. Governor is one layer between your coding agents and your models, and your team keeps using Claude Code, Cursor, Codex, Pi, or GitHub Copilot exactly as they are. 

A router, on its own, cuts the price of a token. Governor also replaces the grepping and searching your agents do, and the two savings stack. It routes each request to the best model and reasoning level, and it grounds every request in your code through Bito’s code context engine. 

For an engineering team, that means the same coding tasks ship for half to a third of the cost, and faster. In one customer’s A/B test on their own engineering tasks, cost per task dropped 48% and task success held at 100%. 

How Governor works 

Coding agents let you change what models you access, so now you just point your base URL at Governor to auto-route across models. Governor speaks the Anthropic and OpenAI APIs, runs in or alongside your existing gateway on your own provider keys, and serves the same code context to every developer server-side. 

Why agent bills grow so fast 

An agent’s bill comes down to two numbers multiplied together: how many tokens it sends and what each token costs. The token count grows with every step the agent takes, and the price depends on which model does the work. Model routing lowers the price, and code context replaces the searching. 

We measured one Claude Code session closely, step by step. The agent’s searches through the codebase returned about 65,000 tokens of results. By the end, carrying those tokens for every turn, the session had processed 6 million. 

That gap explains most of why coding agents cost what they do, and it’s the reason we built Governor’s code context engine. 

A coding agent works in steps. It searches the code, reads a file, makes an edit, runs the tests. Each result gets added to a running transcript, and on the next step the whole transcript goes back to the model, because the model keeps no memory of its own between requests. 

So a file the agent read at step 3 gets paid for again at step 4, and at step 5, and at every step after that. 

Draw it out, and it looks like a staircase. The new work on each step is a thin sliver on top. The grey block underneath, everything being re-sent, gets taller every time, and in a long session it’s nearly the entire bill. Prompt caching makes those re-sends cheaper but doesn’t make them go away. Across our model research, reads from cache were 77% of all the tokens processed. 

How Governor picks the best model for each request 

Before a request reaches a model, Governor decides which model and reasoning level should handle it, from the models your team allows.  It can also add guidance to the prompt. 

Each request is categorised into one of multiple tiers, based on what kind of work it is and how much risk it carries. Tiers run from carrying out a plan the developer already agreed to, up to planning, code review and anything that touches production. For each tier, Governor picks a model and reasoning level from the models your team allows. You decide how aggressive it should be on cost. 

The most expensive model is worth its price on some of the work, but that expense and thinking power is wasted on simpler work. Everything depends on telling the two apart. 

One result from our benchmark made this concrete for us. We gave Sonnet 5.5 and Opus 5.5 the same long coding session. 

When it came time to write the plan, Sonnet scored 91% and Opus 81%. One turn earlier, when the job was working out why something was failing, Sonnet fell to 76% while Opus scored 90%. The same model was the better pick for one step and the worse pick for the step before it. 

Reading each request with its task 

Governor reads each request together with the task it belongs to, because in a coding session the newest message often carries little meaning on its own, as in “looks good, implement this.” Routers that score only the newest message, by its keywords or its size, send that request to the cheapest tier. 

Governor carries the weight of the task forward, so “implement this” gets treated as the build it is. After we switched to sorting by kind of work and risk, the share of our hand-labelled test requests landing in the right tier went from 62% to 100%. 

Built for sessions 

Most routers were designed for chat, with one question in and one answer out. A coding agent turns a single developer request into dozens of model calls. Read a file, run a search, edit, run the tests, read the failure, edit again. 

Judge each of those calls on its own and the model can change halfway through the job, taking the reasoning it had built up with it. Before we built for this, all 6 of the Claude Code turns we checked that started on one model got handed to another partway through. 

Now Governor makes its decision when the developer starts a new request and holds it while the agent works. Across 80 turns since, it hasn’t switched models mid-task once. 

It also keeps track of what’s already cached. Reading a warm Opus cache costs $0.20 per million tokens, while sending that same conversation cold to Gemini Flash costs $0.75. So now and then the expensive model is the cheaper choice for the next request, and Governor moves only when the switch will pay for itself. 

Reasoning level and prompt guidance 

Choosing a model is one of several calls Governor makes. It also sets how hard the model thinks and what guidance it gets. 

Reasoning level moves cost a lot. Turning Sonnet 5.5 up from its default to high reasoning more than doubled what a long session cost and raised its score 11%. Sometimes that’s worth paying for. Turning Opus 5.5’s reasoning below its default cut session time by about a quarter. 

Prompt engineering is another variable Governor uses. It checks whether the developer’s prompt is clear, and whether added instructions would help the model deliver what the developer expects. With better prompts, a less expensive setup matched Opus’s default score at 18% lower cost. Each of these choices saves money on its own, and they stack. 

How Governor powers each request with code context 

A coding agent rediscovers your codebase from scratch on every task. Ask it to change how authentication works, and it starts by grepping for “auth,” reads a result, greps again, opens another file. Every one of those results lands in the transcript and climbs the staircase. 

Governor keeps a live graph of your repositories, down to each symbol and call and how your services depend on each other. When the agent needs to find something, Governor does the research on its own side with a smaller, faster model and hands the agent a short answer, with the file and line each fact came from. 

What code context saves 

In our measurements, 60% to 80% of an agent’s cost goes to searching through code to understand the codebase, nearly all of it at frontier prices. 

We ran one afternoon of real developer requests both ways, tracing an event through Kafka, checking whether a Jira integration already existed, planning a feature, and two code changes. 

Without code context, the agent made 162 tool calls, opened 62 files and took 25 minutes. With Governor it made 14 calls, opened 3 files and finished in 8.5 minutes. It also completed every task, whereas the run without code context finished 90% of them, and the afternoon cost $1.13 instead of $6.81. 

Benchmark and customer results 

We published our AI coding model cost research in September, and the latest run covers ~30 models on about 2,000 real engineering tasks. The clearest lesson was that the most expensive model earns its price on the hardest work and gets overpaid on everything else. Claude Opus 5.5 on every task scored 54.5 out of 60 for $33.22. Keeping Opus 5.5 for planning, code review, and risky or vague requests, with GLM 5.3 Flash on the rest, finished the same 90% for $21.27, 36% less. 

Governor puts that finding to work on live sessions. In our benchmark of long, multi-turn Claude Code sessions, developers on Opus 5.5 kept the same score at 30-70% lower cost, and sessions finished 26% faster. 

In a customer A/B test on their own engineering tasks in Claude Code, Governor cut cost per task by 48%, from $4.12 to $2.14. Task success held at 100% and tasks were completed 38% faster. The agent also took half the reasoning turns and made 53% fewer tool calls. The test used the same model and agent setup with and without Governor, so Governor was the only variable.

What’s next for Governor 

We’re working on smarter and faster routing decisions, on security, and on making Governor easier to deploy and run. We’re also taking it past coding agents into other kinds of AI work. 

On code context, we’re working on adding even more intelligence to bring more context of your whole engineering team to bear on your agentic coding tasks.  

The full method and data are in our AI coding model cost research. To see what Governor would do with your own agent traffic, get access to Governor. 

Picture of Amar Goel

Amar Goel

Bito’s Co-founder and CEO. Dedicated to helping developers innovate to lead the future. A serial entrepreneur, Amar previously founded PubMatic, a leading infrastructure provider for the digital advertising industry, in 2006, serving as the company’s first CEO. PubMatic went public in 2020 (NASDAQ: PUBM). He holds a master’s degree in Computer Science and a bachelor’s degree in Economics from Harvard University.

Picture of Amar Goel

Amar Goel

Amar is the Co-founder and CEO of Bito. With a background in software engineering and economics, Amar is a serial entrepreneur and has founded multiple companies including the publicly traded PubMatic and Komli Media.

Written by developers for developers red heart icon

This article is brought to you by the Bito team.

Latest posts

Best cheap model for coding: Inkling 2 Small vs Claude Haiku 5.5, Gemini 3.8 Flash, DeepSeek and GPT-6.1 Terra

Inkling 2 Small vs Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 for coding

Inkling 2 Small benchmark: the strongest U.S. open-weight coding model we tested, vs DeepSeek, GLM and Kimi

GPT-6.1 Terra vs GPT-6.1 Sol vs GPT-6 Luna for coding: which OpenAI model to use

GPT-6.1 Terra benchmark: how it compares to Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 for coding

Top posts

Best cheap model for coding: Inkling 2 Small vs Claude Haiku 5.5, Gemini 3.8 Flash, DeepSeek and GPT-6.1 Terra

Inkling 2 Small vs Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 for coding

Inkling 2 Small benchmark: the strongest U.S. open-weight coding model we tested, vs DeepSeek, GLM and Kimi

GPT-6.1 Terra vs GPT-6.1 Sol vs GPT-6 Luna for coding: which OpenAI model to use

GPT-6.1 Terra benchmark: how it compares to Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 for coding

From the blog

The latest industry news, interviews, technologies, and resources.

Best cheap model for coding: Inkling 2 Small, Claude Haiku 5.5, Gemini 3.8 Flash, DeepSeek and GPT-6.1 Terra overall scores

Best cheap model for coding: Inkling 2 Small vs Claude Haiku 5.5, Gemini 3.8 Flash, DeepSeek and GPT-6.1 Terra

arrow bito ai
Inkling 2 Small vs Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 coding benchmark overall scores

Inkling 2 Small vs Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 for coding

arrow bito ai
Inkling 2 Small vs DeepSeek, GLM, Kimi and MiniMax coding benchmark overall scores

Inkling 2 Small benchmark: the strongest U.S. open-weight coding model we tested, vs DeepSeek, GLM and Kimi

arrow bito ai