Benchmark
| Attribute | Sediment | Opus 5 | GPT-5.6 Sol | Claude Code |
|---|---|---|---|---|
| Mean cost per query | $0.01 | $0.25 | $0.09 | $0.48 |
| Median speed per query | 20.1s | 40.5s | 51.8s | 24.8s |
Accuracy by task family
- Sediment
- Opus 5
- GPT-5.6 Sol
Same model, four setups
Our systems start from an open-weights model that runs on a machine dedicated to you. On top of it we add two things. The first is training on your company's records, so the model already knows your history before the first question. The second is our own retrieval and orchestration harness, which finds the right records for each question and attaches the source to every answer.
To test it, we built two synthetic companies, each with roughly 25 million tokens of internal documents. We asked five kinds of question. Recall a fact, use the newest version of it, decline when the answer is not there, apply rules nobody wrote down, and draft in the house format. We ran Opus 5 and GPT-5.6 Sol inside the same harness, so the only difference was the model, and we ran Claude Code with Opus 5 on its own over the same documents. The benchmark harness is public on GitHub. The documents and answer keys stay private, which keeps the benchmark valid.