Can many models beat your best one? We test it.
RV_Enhanced is a team of frontier models working together, for hard problems and when accuracy is important. For teams that want more out of AI, on work where a right answer matters far more.
A size and an area. That is all we need from you.
The questions, on each frontier model and its settings.
Every pair and group: where they agree, where they differ, and what a vote gets right.
Which team for each task and sub-task, what it costs, and if it is worth it.
What several models did on our own questions.
right when Claude Sonnet and GPT-6 Astra agree, against 81% and 80% alone
of Claude Sonnet's wrong answers flagged by GPT-6 Astra, at about $1.06 each
right answers of 200 from a vote, against the best single setting of a local model
Two frontier models
Right answers of 191
What agreement tells you
- They agreed
- 147 questions. Both right on 137, or 93%.
- They differed
- 44 questions. Claude Sonnet right on 18, GPT-6 Astra on 14, neither on 12.
- GPT-6 Astra checking Claude Sonnet
- It flags 26 of Sonnet's 36 wrong answers and misses 10, and it flags 18 right answers too.
From a run on 31 Oct 2026: 191 finance and tax questions, the same prompts for both models, at high effort. Answers within $2 count as agreeing. "At least one right" is the most a pair could reach, and it takes a way to choose between them. The cost per catch is GPT-6 Astra's tokens at list prices, before settling the flagged questions. On 40 easier questions, three frontier models gained nothing from a vote: Grok 4.7 alone got all 40.
One model, asked four ways
Right answers of 200
What the vote did
- Against the best single setting
- It gained 11 right answers and lost 1.
- Cost
- About 4.5 times the reply tokens.
- The same prompt, twice
- A different answer on 38 of 200 questions, with almost the same score.
From runs on 4 Nov 2026: Qwen3.8-27B on an RTX 4090, thinking off, 200 finance and tax questions held out from any tuning. We chose this group after seeing the answers, from 16 we compared, so it is a lead for your tasks and not a general result (p = 0.102 after Holm).
More than one answer, on your tasks.
One model, asked several times
The answer that comes up most wins. The same model and prompt can answer differently from one try to the next.
Models checking each other
Where they agree, use the answer. Where they differ, ask another model or a person.
Several models voting
Models get different questions wrong, so even with no single model strong enough, the group may be.
Others have found this too.
Published results on other models and other tasks point the same way. So we test it on yours.
We work it out, and tell you if it is worth it.
- Which models to put on the team, for each task and sub-task.
- How often the answer is right when they agree.
- How many wrong answers the team catches, and what each catch costs.
- Whether a vote gets more answers right than your best model.
- Our call on each: worth it, or stay with one model.
Ready to onboard? Tell us what you are interested in.