Hemant Bhatt

May 16, 2026

Practical Model evaluation for OpenCode and OpenRouter

For agentic coding, intelligence is not just about writing a neat patch. Most real software work falls into three connected buckets: writing code, deploying it, and maintaining it inside environments that rarely behave as cleanly as the docs promised.

That is why we will use the following benchmark mix. SWE-bench will help us with understanding the coding capabilities of the model. Terminal-Bench with whether the model can deal with the environment around that patch: commands, tests, setup, failures, and all the small frictions that make agentic coding real. There scores to tell us exactly where a model stands in intelligence.

I am weighting SWE-bench Pro the highest because harder repo work is the best proxy for serious software engineering. SWE-bench Verified still helps as a familiar comparison point. Terminal-Bench earns a large share too because an agent that cannot operate the shell is not really an agentic coding partner. It is just a very confident autocomplete box with extra steps.

Input tokens and output tokens are usually priced differently for each model, so comparing only one column gives you a crooked picture. A coding agent also reads much more than it writes, which makes input cost especially important. To keep the comparison practical, I use a blended cost: 80% input tokens and 20% output tokens.

Cost alone only tells you how painful a model is to run. Intelligence tells you how often it can finish the work without needing rescue. The useful comparison is what happens when you hold both ideas together.

The cost multiple makes the table easier to judge because it gives you a ruler. You can immediately see how expensive the models are relative to each other. Kimi is roughly 7 times cheaper than Opus, while Tencent Hy3 is nearly 86 times cheaper than Opus. Tencent Hy3 is given 1X because it is the cheapest model in the list.

The composite score is calculated from the three benchmarks with this weighting: 45% SWE-bench Pro, 30% SWE-bench Verified, and 25% Terminal-Bench 2.0. SWE-bench Pro gets the largest weight because harder repo work is the closest signal for real software engineering. Terminal-Bench still gets a serious share because agentic coding also means running commands, reading failures, and working inside an environment.

A normalized intelligence score makes this easier to judge. Instead of staring at three benchmark columns separately, the composite gives you one ruler for raw coding-agent ability. It is not perfect, but it is much better than comparing models by whichever benchmark happens to look nicest.

The simplest way to pick is to ask what failure costs. Not all mistakes have the same price.

A bad answer on a README summary is annoying. A bad answer on a database migration is a calendar event. The same model can be a good choice in one situation and a bad default in another because the stakes changed.

Use this picker when you are sitting in OpenCode and deciding what to run before the next chunk of work. It is intentionally practical: start with the task, then pick the model.