Divyam.AI Router Achieves Frontier-Level Accuracy at One-Eighth the Cost
99.5% of deepseek-v4-pro's accuracy at 12.3% of its cost on RouterArena
A single-model deployment pays the same rate for every request: the one-line lookup and the multi-step proof alike. Most production traffic does not need frontier reasoning, yet the queries that genuinely do cannot be identified without first analyzing them. Routing exists to make that judgement per query, which means its value is entirely empirical, resting on how well each individual decision is made.
At Divyam.AI we solve this by right-sizing every request, routing each one to the model that meets the quality the task actually needs, so spend falls without quality following it. Because routing earns its value empirically, it should be shown rather than claimed. To demonstrate what our Model Router does, we evaluate it on open, independent benchmarks built specifically to test LLM routers.
RouterArena is where those decisions can be measured rather than asserted. Evaluated across its full 8,400-query benchmark, Divyam.AI's Model Router places among the top three on the leaderboard, records the second-highest accuracy of any router evaluated, and reduces the cost of frontier-grade reasoning by 87.7% while retaining 99.5% of its quality.
Results at a glance
- Arena Score 75.27: ties for third on the RouterArena leaderboard
- 77.61% accuracy: second-highest of any router on the leaderboard, ahead of both the #2 and #3 ranked entries
- 87.7% cost savings against deepseek-v4-pro, at 99.5% of its quality
- 16.1% cost savings against gemini-3-flash-preview, at 98.2% of its quality
- Cost savings of 80–93% across all nine knowledge domains: the efficiency is broad, not concentrated in a few easy categories
- On the cost–accuracy Pareto frontier: no candidate model is both cheaper and more accurate than the Model Router
- A tunable cost–quality balance, so the operating point can be matched to your workload's economics
Objective
Measure the performance of Divyam.AI's Model Router on the RouterArena benchmark, and quantify the cost and quality trade it achieves relative to strong single-model baselines.
What is the RouterArena benchmark?
RouterArena is an open platform for comprehensive, independent comparison of LLM routers. Its evaluation dataset spans 9 knowledge domains and 44 categories, with easy, medium and hard difficulty tiers within each domain, totalling 8,400 queries. Because the dataset is broad and difficulty-stratified, it rewards routers that can tell an easy question from a hard one, not merely ones that default to a strong model.
RouterArena reports five router-critical metrics, headlined by the Arena Score, which combines accuracy and cost into a single figure so routers with different price points can be compared on equal terms. Cost enters the score logarithmically, meaning cost advantages must be multiplicative factors, not percentages, before they materially move the score.
All results below are produced by RouterArena's own evaluator over the full 8,400-query evaluation set, with code-execution benchmarks executed rather than string-matched.
Where Divyam.AI's Model Router stands on the leaderboard
Divyam.AI's Model Router is among the top three routers on RouterArena with an Arena Score of 75.27, and it achieves the second-highest accuracy of any router on the board, higher than both the second and third ranked entries.
| Accuracy rank | Router | Accuracy | Arena Score |
|---|---|---|---|
| 1 | Cross-Router | 78.14% | 76.12 |
| 2 | Divyam.AI's Model Router | 77.61% | 75.27 |
| 3 | vLLM-SR | 77.18% | 75.30 |
| 4 | Sqwish Router | 76.40% | 75.27 |
| 5 | Weave Router | 76.32% | 72.82 |
| 6 | Nadir-Tumbler | 75.34% | 75.17 |
Read against the field directly:
- More accurate than the #2-ranked router (vLLM-SR) by 0.43 points, while finishing within 0.03 Arena Score points of it, effectively level on the composite metric, ahead on capability.
- More accurate than the #3-ranked router (Sqwish) by 1.21 points, at an identical Arena Score.
- Within 0.53 accuracy points of the top-ranked router.
Accuracy is the harder axis to move: it reflects how well the Model Router understands each query. On that axis Divyam.AI's Model Router is already at the top of the field.
The cost–quality trade
Two baselines frame the result: gemini-3-flash-preview, the most accurate single model available to the Model Router, and deepseek-v4-pro, a powerful frontier reasoning model roughly eight times more expensive.
| Baseline | Quality retained | Cost savings | Cost ratio |
|---|---|---|---|
| deepseek-v4-pro | 99.5% | 87.7% | 0.12× |
| gemini-3-flash-preview | 98.2% | 16.1% | 0.84× |
Against deepseek-v4-pro, the Model Router retains 99.5% of the model's quality for 12.3% of its cost, a 195:1 ratio of cost saved to quality forgone. For a workload of one million queries, that is the difference between a frontier-reasoning bill and roughly an eighth of it, for a fraction of an accuracy point.
Against gemini-3-flash-preview, the Model Router retains 98.2% of quality at 16.1% lower cost, a 9:1 ratio of savings to loss.
Identifying the right model garden
A router is only as good as the models it routes among. With 100+ models available and new releases every few weeks, the harder question is not how to route but what to route between, and most candidates do not earn a place. A model that is expensive without being distinctly better adds cost and no accuracy; a model that is cheap but never the right answer adds nothing at all.
Identifying that shortlist, the model garden worth routing among for a given workload, is part of the work we do when deploying Divyam.AI's Model Router. We profile candidates against representative traffic, keep the ones that win queries no other model wins, and drop the rest.
The RouterArena results show why this matters. deepseek-v4-pro is a strong model by reputation, yet on this benchmark it is not cost-efficient at any operating point: a cheaper model in the garden is both more accurate and 6.8× less expensive (Figure 1). Routing traffic to it would have raised cost without raising quality. Equally, two models retained in the garden proved near-redundant on this traffic and received under 1% of requests between them, a curation signal we act on.
The result is a compact garden where every member is there because it wins something, which is what allows the Model Router to hold near-frontier accuracy at a fraction of frontier cost, and what lets it keep pace as the model landscape shifts.
Where the savings come from
Savings are not the product of uniformly downgrading traffic. They come from spending frontier money only where difficulty justifies it.
By difficulty tier
| Difficulty | Queries | Divyam accuracy | deepseek-v4-pro accuracy | Cost savings |
|---|---|---|---|---|
| Easy | 3,990 | 97.17% | 96.54% | 72.4% |
| Medium | 2,445 | 77.18% | 78.22% | 88.6% |
| Hard | 1,965 | 38.42% | 39.90% | 91.2% |
Two observations stand out:
- On easy queries Divyam.AI's Model Router is more accurate than deepseek-v4-pro: 97.17% versus 96.54%, at 72% lower cost. Frontier reasoning applied to easy questions is spend without return, and the Model Router avoids it.
- On hard queries, where cost matters most, savings are largest at 91.2%. deepseek-v4-pro's cost rises 7.5× from easy to hard as reasoning length grows; Divyam.AI's Model Router's cost rises only 2.4×.
By knowledge domain
| Domain | Queries | Divyam accuracy | Cost savings vs deepseek-v4-pro |
|---|---|---|---|
| Computer science & information | 1,400 | 86.3% | 86.1% |
| Science | 1,400 | 84.9% | 85.3% |
| Technology | 1,400 | 82.6% | 84.1% |
| Philosophy & psychology | 700 | 82.3% | 80.6% |
| Social science | 700 | 75.9% | 86.8% |
| History | 700 | 77.0% | 83.3% |
| Arts & recreation | 700 | 70.9% | 91.3% |
| Language | 700 | 64.9% | 81.2% |
| Literature | 700 | 52.9% | 93.1% |
Savings hold between 80% and 93% in every domain. The Model Router also matches or exceeds deepseek-v4-pro's accuracy in four of the nine domains: science, technology, philosophy & psychology, and social science, while spending 80–87% less in each.
A tunable cost–quality balance
A fixed router forces a single trade-off on every workload. Divyam.AI's Model Router instead exposes the cost–quality balance as a configurable operating point.
Selection weighs each candidate's expected quality against its price, and how much quality is worth paying for is itself a setting; adjusting it traces a continuous frontier of operating points from quality-leaning to cost-leaning.
The configuration reported here is the quality-leaning end of that frontier, chosen deliberately, because RouterArena's Arena Score rewards accuracy strongly. A deployment with different economics is not obliged to inherit that choice:
- Cost-sensitive workloads can move down the frontier, trading measured quality for further savings.
- Quality-critical workloads can hold the current point, or narrow the candidate pool to stronger models.
This is a practical advantage in production, where the right balance is set by the application and its budget, not by a benchmark. We tune this per engagement.
Benchmark setup
- Benchmark: RouterArena, full evaluation set, 8,400 queries across 9 domains and 44 categories, with easy/medium/hard difficulty labels.
- Grading: RouterArena's official evaluator, with code benchmarks (LiveCodeBench) executed, not string-matched.
- Router under test: Divyam.AI's Model Router.
- Pricing: published list prices from RouterArena's cost configuration, applied identically to the Model Router and to every baseline, so all comparisons are like-for-like. Cost per 1,000 queries is the benchmark's reported unit.
- Baselines: single models evaluated on the same 8,400 queries with the same grader, so accuracy and cost differences reflect routing alone.
- Leaderboard positions and competitor figures are as published by RouterArena.
References
- RouterArena benchmark, methodology and live leaderboard: routeworks.github.io · leaderboard
- Who Routes the Routers?, RouterArena's introduction of the benchmark and its metrics: huggingface.co/blog/JerryPotter/who-routes-the-routers
- Divyam.AI, EvalMate and the Model Router: divyam.ai
Interested in how Divyam.AI's Model Router would perform on your traffic, or where on the cost–quality frontier your workload should sit? We are happy to benchmark against your own evaluation set.