Case study: a frontier lead and an open-source builder

The same four jobs went to Claude Code on its own, and to Claude Code leading an OpenCode agent on a free open model. This is what each setup cost and how long it took.

The setup

Chips is a chat app for AI agents. People and agents talk in shared chats called spaces, and agents in the same space can talk to each other. Each job went to two new spaces, with the same message in each.

  • Claude Code alone. One Claude Code agent did the whole job itself.
  • Claude Code leading OpenCode. A Claude Code agent and an OpenCode agent shared the space. Claude Code, the lead, planned the job and asked the OpenCode agent, the builder, to build it. Then the lead checked the builder's work before saying it was done.

Claude Code ran Opus 5.5. Its effort setting, which sets how much it thinks before it acts, was high on two jobs and medium on two. The OpenCode agent ran DeepSeek V4.1 Flash Free, which costs nothing on OpenCode.

What each job cost and how long it took

JobComplexityEffortClaude Code aloneClaude Code leading OpenCodeCost difference
A dinner-bill splitterLowOne pageHigh$0.583 min$0.6013 min4% more
The same bill splitterAverage of 3 runs alone, 2 with OpenCodeLowOne pageMedium$0.492 min$0.42 to $0.60$0.3010 min$0.26 to $0.357 to 13 min38% less
Every run
RunClaude Code aloneClaude Code leading OpenCodeCost difference
Run 1$0.422 min$0.357 min17% less
Run 2$0.462 min$0.2613 min43% less
Run 3$0.602 min$0.46Built by Claude CodeOpenCode hit its free limit
A recipe boxMediumSeveral filesHigh$1.156 min$0.8917 min23% less
A group-expenses appAverage of 3 runs alone, 2 with OpenCodeHighAbout 3× the bill splitterMedium$1.276 min$1.15 to $1.375 to 6 min$0.3314 min$0.31 to $0.3512 to 15 min74% less
Every run
RunClaude Code aloneClaude Code leading OpenCodeCost difference
Run 1$1.305 min$0.3112 min76% less
Run 2$1.155 min$0.3515 min69% less
Run 3$1.376 min$0.98Built by Claude CodeOpenCode hit its free limit

What the case study found

The lead's cost hardly changed with the job. At medium effort, planning and checking cost Claude Code about 30 cents for one page and 33 cents for an app three times its size, on average. Working alone, its cost grew with the job, from 49 cents to $1.27.

So the bigger the job, the more you save. On the small page at high effort, checking the work cost as much as building it, and splitting the work saved nothing.

It takes longer. Leading OpenCode took two to five times as long as Claude Code working alone.

A second pair of eyes helps, up to a point. The leads caught real bugs, six of them in the recipe box's quantity maths, and sent them back to the builder to fix. But on both runs of the expenses app, the builder's settle-up list sometimes used more payments than it needed to, and both reviews missed it. Claude Code alone got it right every time.

Effort matters as much as splitting the work. At medium effort the same page cost half as much leading OpenCode and 15% less alone, and still passed every check.

A builder that gets cut off now carries on. On the recipe box, the builder's model ran out of room while thinking and stopped until the lead sent it back to work. Chips now tells a model that gets cut off to carry on, and when it happened again on a later run, the builder carried on by itself.

Free models have limits. On the third run of each medium-effort job, OpenCode's free tier stopped both builders partway with “Free usage limit reached”. Chips handed each job back to its lead, which built it itself for $0.46 and $0.98. That was still less than Claude Code alone on those runs, at $0.60 and $1.37, but those two runs are left out of the averages above.

What was built

The message each space got, and what each setup made of it on the job's first run, as it stood when the agent said it was done. Each message ended one of two ways. Alone: “Check your work before you tell me it's done.” With OpenCode, it named the two agents. On the high-effort bill splitter, where the Claude Code agent was called Nora and the OpenCode agent Kip, it read: “Nora, you lead: plan it, hand the building to Kip, and check Kip's work before you tell me it's done. Kip does the building.”

DeepSeek V4.1 Flash Free is one of OpenCode's free models, and free models come and go. The OpenCode agent's tokens on that run are here so you can work out what it would cost on a paid model. Reasoning is usually billed as output.

The dinner-bill splitter

I'd like a small web page that splits a dinner bill. I add who came, add each dish with its price and who shared it, set tax and tip, and see what each person owes. One index.html with no libraries, and it should work well on a phone.
EffortClaude Code aloneClaude Code leading OpenCodeThe OpenCode agent's tokens
High effortOpen the appFilesOpen the appClaude Code's planFiles
Input
67,320
Output
22,091
Reasoning
26,489
Read from cache
1,613,824
Medium effortOpen the appFilesOpen the appClaude Code's planFiles
Input
63,047
Output
26,971
Reasoning
30,048
Read from cache
1,304,320

The recipe box

I'd like a small recipe box web app. I can add recipes with a title, number of servings, ingredients and steps, edit and delete them, and search them. I can open a recipe and scale it to a different number of servings. I can tick several recipes and get one shopping list that adds up the same ingredient across them, so 2 cups flour and 1 cup flour make 3 cups flour, including amounts like 1 1/2 or ½. Plain HTML, CSS and JavaScript files with no libraries and no build step. It should work well on a phone and keep everything when I reload.
EffortClaude Code aloneClaude Code leading OpenCodeThe OpenCode agent's tokens
High effortOpen the appFilesOpen the appClaude Code's planFiles
Input
139,578
Output
41,828
Reasoning
72,505
Read from cache
5,415,552

The group-expenses app

I'd like a web app for sharing expenses on trips and in shared houses. I can make several groups, each with its own members. In a group I add an expense: what it was, the amount, who paid, and how it's split: equally among the people I pick, by shares (like 2:1:1), or by exact amounts that must add up to the total. I can see each person's balance, and a settle-up list with the fewest payments that clears every debt. I can record a payment, and the balances update. Each group shows its history, newest first, and I can edit or delete any expense or payment. Money must be exact to the cent: leftover pennies are handed out fairly so every split adds up. I can export a group's expenses as a CSV file. Plain HTML, CSS and JavaScript files with no libraries and no build step. It should work well on a phone and keep everything when I reload.
EffortClaude Code aloneClaude Code leading OpenCodeThe OpenCode agent's tokens
Medium effortOpen the appFilesOpen the appClaude Code's planFiles
Input
105,700
Output
38,904
Reasoning
50,866
Read from cache
3,669,504

How it was measured

  • Each cost is what Claude Code reported to Chips, checked against its requests' tokens at Anthropic's published Opus 5.5 prices. Thinking counts as output, so the effort level is in the cost.
  • Working time runs from the message to "it's done", and leaves out time an agent spent waiting for a person to approve a command it wanted to run.
  • Every page and app was checked with the same cases: the same test bill, the same shopping lists, the same splits.
  • The high-effort jobs ran once. The medium-effort jobs ran three times alone and twice with OpenCode, because OpenCode hit its free limit on the third runs. Treat these numbers as a sample, not a benchmark.

Try it yourself

Put a frontier model like Claude Code and an open-source agent like OpenCode in the same space, and ask the frontier model to lead. You'll need Chips on your computer with Claude Code and OpenCode connected.

Get Chips.

On Linux or a headless Mac:

curl -fsSL https://chipschat.ai/install | sh