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 

Composer for the easy turns, Opus for the hard ones: what switching models mid-session really costs

Table of Contents

A customer asked us a fair question. Cursor’s Composer 2.5 is cheap and fast. Opus 5.5 is strong and expensive. Why not use Composer for the routine parts of a coding session and switch to Opus when the work gets hard?

So we tried it. We took a 20-turn coding session from real work on our own services and ran it in Cursor three ways: Composer on every turn, Composer on 12 routine turns with Opus 5.5 on the 8 hard ones, and Composer on the routine turns with Cursor Auto on the hard ones. We picked the hard turns the way a careful engineer would: the investigation at the start, the production incident in the middle and the final audit.

Mixing helped, but much less than you’d expect, and it cost more than people assume.

60%64%68%72%76%80%84%88%92%$0$5$10$15$20$25$30$35$40$45Cost per 20-turn session (cheaper →)Score (better ↑)Opus 5.5, every turn90% · $43.37 · 112 minSonnet 5.5, every turn87% · $39.22 · 110 minGovernor Quality89% · $30.73 · 102 minGovernor Balanced87% · $24.70 · 76 minGovernor Cost83% · $8.80 · 65 minComposer on easy turns + Opus on hard75% · $18.51 · 46 minComposer on easy turns + Auto on hard71% · $17.51 · 42 minComposer only63% · $4.90 · 38 min
One 20-turn agentic coding session on two real services. Score as a percentage; cost and minutes per session. Cheaper is to the right and better is up.
SetupScoreCost per sessionSavings vs Opus 5.5Minutes
ModelOpus 5.5, every turn90%$43.37baseline112
GovernorQuality mode89%$30.7329%102
ModelSonnet 5.5, every turn87%$39.2210%110
GovernorBalanced mode87%$24.7043%76
GovernorCost mode83%$8.8080%65
CursorComposer 2.5 on easy turns + Opus 5.5 on hard turns75%$18.5157%46
CursorComposer 2.5 on easy turns + Cursor Auto on hard turns71%$17.5160%42
CursorComposer 2.5, every turn63%$4.9089%38

Composer plus Opus got to 75%. Governor’s Cost mode scored 83% for less than half the money. Balanced mode matched Sonnet at 87% for $24.70, under what Sonnet costs for the same session.

The easy turns weren’t easy

Here’s what surprised us. The hard turns went fine. When Opus handled the investigation, it scored the same as Opus running the whole session, about 90%. The points were lost on the turns everyone would call routine.

Routine turn, handled by Composer in the mixComposer in the mix vs Opus every turn
Write the implementation plan
70%89%
Run the test suites, record a baseline
62%88%
Implement the fix
60%89%
Handle a requirement change from product
36%86%
Respond to a code reviewer
73%93%
Re-run the tests, compare with the baseline
56%90%
Write the pull request descriptions
55%91%
Write the on-call runbook
80%93%

Composer, mixed sessionOpus 5.5, every turn

What these turns have in common is that each depends on something established earlier in the session. The model has to remember a baseline from eight turns back, or carry a race condition found in the investigation into the code it writes.

That’s where Composer slipped, and the judge’s notes are specific about it. When asked to re-run the tests and compare against the baseline, Composer compared against the wrong snapshot. It called a failing check “new” when that check had been failing since the start, and it reported the bot’s tests as unchanged when 13 had been added. On the requirement change it said one of the two services needed no change at all, missing a status update that could overwrite the user’s answer. On the implementation turn it said a five-second limit was met while an awaited update still ran on a retrying client that could hold the answer far longer.

None of that was hidden. The baseline it should have used was sitting earlier in the same conversation, and the code was right there in the repo. The answers read fine, which makes this worse: you’d ship them.

So “use the cheap model for the easy stuff” falls apart in agentic work, because late-session work is rarely easy in the way it looks. Writing a PR description is simple. Writing a correct one for a change made across two services over twenty turns isn’t.

Every switch back to Opus re-buys the conversation

The second problem is cost, and it’s easy to miss.

Model providers cache the conversation so each new request only pays full price for what’s new. That cache belongs to one model. When the session moved from Composer back to Opus, Opus had to write the whole conversation into its cache again. When Opus followed Opus in these sessions, it usually wrote 25,000 to 60,000 new tokens to its cache. On the turns right after a Composer stretch it wrote 140,000 to 350,000.

97KTurn 126KTurn 244KTurn 346KTurn 4225KTurn 1387KTurn 14267KTurn 17196KTurn 20
Opus cache writes per Opus turn, in tokens, mean of two sessions. Pink bars are the turns right after Composer handled the session.

In dollars, those rewrites came to about $5 to $5.50 of each $17 to $20 session, roughly 30% of the bill. Composer’s side of the switch is close to free. The expense lands every time you switch back to the expensive model, which is exactly when you wanted it.

Cursor Auto made this a little worse. On the hard turns it moved between Opus 5.5, GPT-5.6 Terra, Sonnet 5 and Grok 4.7, so there were more switches, and it scored lower than handing those turns to Opus.

What works instead

The Governor settings in the table don’t swap models mid-session. Balanced mode keeps Opus for the whole session and lowers how hard it thinks, with written guidance on how to work: read the code before changing it, cover every part of the request, check instead of assuming. Cost mode runs Haiku 5.5 the same way. The cache stays warm, nothing gets re-bought, and every turn is handled by a model that can carry the whole session.

That’s why Cost mode beats a Composer and Opus mix by 8 points at half the price. One capable model with good guidance does better than a cheap model that drops context with an expensive one standing by.

How we tested

The session is 20 turns on two real services: investigate a production slowdown, prove the cause by experiment, plan, implement, test, handle a requirement change and code review, fix a follow-up incident, then write PR descriptions, a runbook and a final audit. Every answer was graded against an answer key checked against the code, by a judge scoring five dimensions. The judge is Claude Opus 5.5, and the answer keys tie it to facts rather than taste. We ran two sessions per setup. Session-to-session spread was 1 to 2 points.

Costs are what each session cost, as charged by Cursor or the model provider.

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

Google’s biggest jump yet: Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.1 Pro for coding

Which non-Claude model to pick for coding: Gemini 4 Carbon vs GPT-6.1 Sol and GPT-6.1 Terra

Google is back in the race for coding: Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2

Muse Ember vs DeepSeek, MiniMax, Kimi, GLM and Inkling: the cheapest models for coding agents

Muse Ember benchmark: how Meta’s cheap model compares to Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and GPT-6.1 Sol for coding

Top posts

Google’s biggest jump yet: Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.1 Pro for coding

Which non-Claude model to pick for coding: Gemini 4 Carbon vs GPT-6.1 Sol and GPT-6.1 Terra

Google is back in the race for coding: Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2

Muse Ember vs DeepSeek, MiniMax, Kimi, GLM and Inkling: the cheapest models for coding agents

Muse Ember benchmark: how Meta’s cheap model compares to Claude Haiku 5.5, Sonnet 5.5, Opus 5.5 and GPT-6.1 Sol for coding

From the blog

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

Gemini 4 Carbon vs Gemini 4 Argon and Gemini 3.1 Pro coding benchmark overall scores

Google’s biggest jump yet: Gemini 4 Carbon vs Gemini 4 Argon vs Gemini 3.1 Pro for coding

arrow bito ai
Gemini 4 Carbon vs GPT-6.1 Sol and GPT-6.1 Terra coding benchmark overall scores

Which non-Claude model to pick for coding: Gemini 4 Carbon vs GPT-6.1 Sol and GPT-6.1 Terra

arrow bito ai
Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2 coding benchmark overall scores

Google is back in the race for coding: Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2

arrow bito ai