Models and pins
Every agent runs DeepSeek V4.1 Flash today, at medium reasoning. Each pin moves only when the benchmark finds a better buy.
Today’s pins
| Agent | Model | Retry model | Cost per review |
|---|---|---|---|
| Code Review | deepseek/deepseek-v4.1-flashmedium reasoning | meta/muse-spark-1.3medium reasoning | $0.102 |
| Security Review | deepseek/deepseek-v4.1-flashmedium reasoning | z-ai/glm-5.3-flashlow reasoning | $0.063 |
| AI Architecture | deepseek/deepseek-v4.1-flashmedium reasoning | z-ai/glm-5.3medium reasoning | $0.050 |
| Log & Errors | deepseek/deepseek-v4.1-flashmedium reasoning | google/gemini-3.8-flashmedium reasoning | $0.039 |
| Docs Review | deepseek/deepseek-v4.1-flashmedium reasoning | google/gemini-3.8-flashmedium reasoning | Not measured yet |
| API Performance | deepseek/deepseek-v4.1-flashmedium reasoning | meta/muse-spark-1.3medium reasoning | $0.043 |
| SOC 2 Review | deepseek/deepseek-v4.1-flashmedium reasoning | z-ai/glm-5.3-flashlow reasoning | $0.038 |
| CI/CD Performance | deepseek/deepseek-v4.1-flashmedium reasoning | deepseek/deepseek-v4-pro-0813medium reasoning | $0.0095 |
Model ids are OpenRouter’s. When a lane’s first attempt fails, the second can run on its retry model. Cost per review is what the pin averaged on IonWarpBench.
Around the agents
| Job | Model |
|---|---|
| The planner, which writes the walkthrough and the repository facts every agent reads | openai/gpt-6-luna-prolow reasoning |
| The checker, which re-reads every finding | deepseek/deepseek-v4.1-flashmedium reasoning |
| Ask IonWarp, the chat in the app | deepseek/deepseek-v4.1-flashmedium reasoning |
How a pin is chosen
IonWarpBench runs each agent’s tasks on candidate models: positive cases with a known defect and negative cases with none. A judge grades every finding, and the bench records what each run cost and how long it took. A model becomes an agent’s pin when it costs least per task while staying above the quality bar.
The results are public: the benchmark, and how we measure.
A different model for your workspace
A workspace or a single project can run another model on any agent, such as a larger one for Code Review. A project’s choice wins over the workspace’s, and the workspace’s over IonWarp’s pin. The model’s cost comes out of your monthly budget.
To set one, write to team@ionwarp.com or use the link on the agent’s page in the app.
{
"code-review": { "model": "meta/muse-spark-1.3", "reasoning_effort": "medium" }
}Keys are the agents’ reviewer ids: code-review, security-review, ai-architecture, log-coverage, docs-freshness, performance-review, soc2-review, cicd-performance.
Where models run
Every call goes through OpenRouter, only to providers that don’t keep or train on what they receive. Security has the details.