Prioritize the benchmark result.
Use the performance selection in the supplied snapshot for the relevant workload. Review task coverage and benchmark comparability alongside the score.
Recommendation methodology
Know what goes into
the recommendation.
The scores, prices and calculations behind each performance-to-cost comparison.
Supplied recommendation snapshot
Read the source status with the numbers.
Arena scores and reference prices were supplied by the project owner. Image and video prices are approximate. The leaderboard snapshot date and public API IDs remain unconfirmed; supplied data has not been independently verified.
Raw Arena scores and reference prices were supplied by the project owner. The differences below are calculated from those values. Independent verification of the leaderboard source, snapshot date and billing specifications remains pending.
01 / Performance data
Scores are signals.
Your workload decides the fit.
The current snapshot references Arena and includes a raw score for each recommendation. These scores are supplied data; they have not been independently verified against a dated leaderboard.
Benchmark results are attributed to their named source, rather than produced by Roten. Coverage and usefulness vary by task. Evaluate the model on your own application to decide whether the tradeoff works for you.
- Benchmark
- Arena
- Benchmark snapshot date
- Not supplied
- Leaderboard source
- Source URL not supplied
- Data received
- 2026-10-12Receipt date, not the benchmark or pricing snapshot date.
02 / Selection
Two useful answers to one decision.
Smarter performance-to-cost balance.
Consider the score gap and the supplied API price together. The right balance depends on the workload and the amount of inference you run.
Selections are curated from active Roten models. The website follows the supplied recommendation pairs and does not automatically rerank the full upstream catalog.
When the same model wins both categories, both columns identify that model. The comparison does not introduce a weaker alternative to fill a second position.
03 / Cost comparisons
Compare the price for the unit you use.
The latest supplied pricing snapshot is the source of the displayed rates. Those prices are used directly; no additional provider discount is applied.
Input and output savings are separate.
Text models charge different rates for input and output tokens. The main comparisons show the reduction on each side separately, using rates quoted per one million tokens.
Token price comparison formulas
Input cost reduction (%) =
(performance input rate − value input rate)
÷ performance input rate × 100
Output cost reduction (%) =
(performance output rate − value output rate)
÷ performance output rate × 100A mixed workload needs an explicit ratio.
When a combined reference cost is used, the mix is 75% input / 25% output, normalized to one million combined tokens: 750,000 input and 250,000 output tokens.
Optional mixed-token reference formula
Mixed reference cost =
(0.75 × input price) + (0.25 × output price)
Reference cost reduction (%) =
(performance reference cost − value reference cost)
÷ performance reference cost × 100Use your own ratio for production estimates.
A mixed cost reduction applies to that reference workload. Your savings can change with the amount of input and output your application uses.
Media comparisons use supplied reference prices.
Image costs are compared per image, and video costs per second. Approximate supplied prices produce approximate cost differences. These figures compare the supplied references; final costs depend on the billing specification for the request.
Image and video reference cost formula
Reference cost reduction (%) =
(performance unit price − value unit price)
÷ performance unit price × 100
Use the same unit: one image or one second.
Mark the result approximate when the supplied price is approximate.Image size, quality, duration and other billable settings have not been supplied as a matched specification. Confirm those settings before using the reference comparison as a production quote.
When one video model holds both recommendation positions, the model, score and price are the same on each side. Its cost difference is zero.
04 / Arena differences
Show the points and the percentage.
The point difference subtracts the Best Performance score from the Best Value score. The percentage uses the Best Performance score as the denominator. A negative result means the value selection has a lower Arena score.
Arena difference formulas
Arena point difference = value score − performance score
Arena percentage difference =
(value score − performance score)
÷ performance score × 100
A negative difference means a lower score than Best Performance.Displayed percentages are rounded for readability; the raw scores stay visible so the calculation can be reproduced. Arithmetic computed from supplied values does not establish independent source verification.
An Arena score difference does not describe the same percentage difference in real-world quality, accuracy or usefulness. The leaderboard is one signal for choosing a model.
05 / Recommendation snapshot
All eight comparisons.
One shared data source.
The model names, scores, prices and calculated differences below use the same canonical configuration and helpers as the model pages.
| Functionality | Best Performance | Arena Score | Best Value | Arena Score | Arena Difference | Price Comparison | Cost Difference |
|---|---|---|---|---|---|---|---|
| Text | Claude Fable 5 High | 1504 | GLM-5.3 Flash | 1475 | -29 / -1.9%Points / percent | Performance$10.00 input / $50.00 output · per 1M tokensValue$0.06 input / $0.20 output · per 1M tokens | 99.4% lower input cost99.6% lower output costCompared separately per 1M tokens. |
| Coding | Claude Fable 5 High | 1551 | GLM-5.3 Flash | 1525 | -26 / -1.7%Points / percent | Performance$10.00 input / $50.00 output · per 1M tokensValue$0.06 input / $0.20 output · per 1M tokens | 99.4% lower input cost99.6% lower output costCompared separately per 1M tokens. |
| Vision | Claude Fable 5 High | 1308 | GLM-5.3 Flash | 1273 | -35 / -2.7%Points / percent | Performance$10.00 input / $50.00 output · per 1M tokensValue$0.06 input / $0.20 output · per 1M tokens | 99.4% lower input cost99.6% lower output costCompared separately per 1M tokens. |
| Documents | Claude Fable 5 High | 1496 | Kimi K2.6 | 1451 | -45 / -3.0%Points / percent | Performance$10.00 input / $50.00 output · per 1M tokensValue$0.95 input / $4.00 output · per 1M tokens | 90.5% lower input cost92% lower output costCompared separately per 1M tokens. |
| Image Generation | GPT Image 2 | 1383 | Qwen Image 2512 | 1125 | -258 / -18.7%Points / percent | Performance~$0.06 / imageValue~$0.02 / image | ~66.7% lower costApproximate supplied unit prices. |
| Image Editing | GPT Image 2 | 1462 | Qwen Image Edit | 1242 | -220 / -15.0%Points / percent | Performance~$0.06 / imageValue~$0.03 / image | ~50% lower costApproximate supplied unit prices. |
| Text → Video | Wan 3.0 | 1476 | Wan 3.0 | 1476 | Same modelNo score gap. | ~$0.07 / secondSame price for both recommendations. | Same costSame model and reference price. |
| Image → Video | MiniMax H3 | 1494 | MiniMax H3 | 1494 | Same modelNo score gap. | ~$0.08 / secondSame price for both recommendations. | Same costSame model and reference price. |
Source: project owner’s supplied Arena and pricing snapshot. Leaderboard URL, benchmark date and final billing specifications have not been supplied. Approximate prices are marked with ~.
06 / Updates
Recommendations can change.
Selections are reviewed as benchmark rankings change, Roten pricing changes, new models are supported or workload-specific evaluations provide better evidence.
Model IDs, rates, capabilities, scores and recommendation metadata are maintained in one shared configuration. Dated source information can be added when it is supplied, so the website stays consistent.
Explore the current recommendations