# How We Beat Claude Opus with a Smaller Model.

By Tommy · Published 2026-05-20 · Updated 2026-05-20
Source: https://pre.dev/blog/how-we-beat-claude-opus-with-a-smaller-model/

**And cut our per-task token bill in the process.**

TL;DR: Orchestration beats brute reasoning, a better harness yields higher accuracy for less money.

## The AI Token Trap

For the last two years, the standard way to improve a coding agent has been simple:

**Pay more for a bigger model.**

Buying more "reasoning" is a brute-force solution. We chose a different path.

We didn't buy a bigger model. We built an architecture-first model harness optimized for production-grade systems

Can a better harness beat a frontier model on the industry's most brutal benchmark?

## The Arena: Terminal-Bench 2.0

Terminal-Bench is the industry-standard evaluation framework for AI agents.

-   **The Tasks**: 89 real-world scenarios.
-   **The Test**: Diverse, multi-step problems.
-   **The Result**: Programmatic verification. You either solve it, or you don't.

### Wait, What is a Harness?

The AI Model: Just ingests and outputs text.

The Harness: The environment around theAI. It handles memory, multi-step workflows, tool execution, and verification loops.

The Model is the engine; the Harness is the entire car.

predev and Claude Code are harnesses; Sonnet and Opus are models.

## The Competitors

predev + Sonnet 4.6 vs Claude Code + Opus 4.5

We ran predev + Sonnet 4.6 on the harbor reference harness against the Terminal-Bench 2.0 task set.

Configuration: n=89, pass@1, identical task IDs.

The Claude Code numbers were taken directly from their public submissions on tbench.

![TerminalBench2.0 Results pre.dev (56.2%) vs Claude Code (53.9%)](/blog/content/images/2026/05/results-terminalbench.png)

TerminalBench2.0 Results predev vs Claude Code

## The Verdict

-   predev + Sonnet: 56.2%
-   Claude Code + Opus: 53.9%

**We dropped an entire model tier and still finished ahead.**

### Efficiency > Brute Reasoning

The agent's architecture matters more than its weights.

Accuracy went up, while the per-task model bill went down.

## How We Did It

The Core Architecture:

-   **Plan Before Code**: Extracts a structured blueprint before touching a file.
-   **Dynamic Execution**: Uses ToDo dependency graphs to solve non-linear problems.
-   **Strict Verification**: A blind verifier re-runs acceptance criteria and freely disagrees.

## Who is This For?

We built predev for real-life scenarios where ROI matters.

Our customers include:

-   Agencies building for their clients.
-   Vendors building PoCs for their prospects.
-   Enterprise teams building internal data pipelines.
-   Star tups getting their MVP out on time and on budget.

## The Takeaway

If ROI matters to your business, stop burning tokens on brute force.

You don't need a frontier lab to beat one.

You just need a system engineered for the actual work.

If this applies to you, please reach out to discuss optimizing your token spend.

**The Data:**

If you want to explore the whole benchmark in detail, head over to Adam's [full break-down](https://pre.dev/benchmark).

The Open-Source Trajectories & Harbor Results are [available on GitHub.](https://github.com/predotdev/predev-terminal-bench-2.0)

---

About pre.dev: pre.dev is the Long Horizon Coding Agent: it plans, builds and verifies software for hours, in the browser or the terminal. pre.dev Labs also sells RL Tasks to frontier AI labs for model post-training: hard, verified coding tasks built from licensed private production code, each shipped as a runnable Harbor container with a graded verifier, a hidden reference solution and solver trajectories, at 1,000+ new tasks a month. https://pre.dev