See the evidence
behind the comparison.

01 / Benchmark evidence

Compare the raw scores
and their point gaps.

The Arena figures on this website are reference records presented as scores and point gaps. Evaluate models on your own tasks before choosing.

A point gap is the cost candidate’s recorded score minus the other candidate’s recorded score. For example, the text records are 1475 and 1504: a gap of −29 points. Score points do not measure a proportional difference in accuracy, quality or usefulness.

An API test establishes only the capability and request settings it exercised. It does not verify an Arena result. Evaluate candidate models on your application before choosing.

02 / Candidate eligibility

Available candidates.
Visible limits.

A candidate must have an available API route for the workload. Review its input formats and requirements: text input does not imply native image or PDF support.

GLM is listed for text input and is excluded from the vision comparison. Both Qwen image entries are unavailable for their advertised workloads and are excluded from actionable comparisons.

When there is one eligible candidate, it appears once. A missing alternative or a model appearing in both recorded positions does not establish a performance or value winner.

03 / Scoped prices

Compare the same billing settings.

Published prices apply to the stated model settings. Compare the same billing units, input types and output settings on both sides.

Token comparisons use uncached input and output prices per million tokens, with each side shown separately. Cache reads, cache writes and other billable features can change the charge.

Input or output cost reduction (%) =
  (reference rate − candidate rate) ÷ reference rate × 100

Optional mixed reference: 750,000 input + 250,000 output tokens.
Use the same billing units and request settings on both sides.

The optional mixed reference uses 75% input / 25% output. That ratio describes a cost estimate only; use your application’s input and output quantities for budgeting.

Media prices require matching settings. GPT Image 2’s displayed rate is for 1024×1024, low-quality image generation only. It is not an image-editing quote. Contact Roten for image-editing pricing. Wan 3.0’s displayed rate covers 720p without a reference video; MiniMax H3’s covers 768p without a reference video. Other resolutions, quality settings and reference types can change the rate.

No media savings percentage is published when an eligible alternative or a matching billing specification is missing.

04 / Usage and wallet charges

Exact usage cost.
Integer wallet units.

The audited GLM-5.3 Flash request used 24 input and 128 output tokens at $0.15 and $0.50 per million tokens. Its calculated usage cost was $0.0000676. Upstream settlement rounded up to 68 micro-USD units, giving an actual wallet deduction of $0.000068.

24 input tokens × $0.15 / 1,000,000 = $0.0000036
128 output tokens × $0.50 / 1,000,000 = $0.000064
Calculated usage cost                         $0.0000676

Upstream wallet units = ceil($0.0000676 × 1,000,000) = 68
Wallet deduction for this audited request = 68 / 1,000,000
                                         = $0.000068

Roten displays upstream account and usage information. Token counting, rate selection, charging and wallet settlement remain upstream; Roten does not independently reprice requests.

Calculated usage cost and wallet deduction can differ at the settlement unit. Historical customer usage responses do not expose a frozen USD conversion for every request, so a historical wallet charge is not reconstructed from today’s prices or conversion settings.

05 / Workload comparisons

Recorded scores.
Scoped costs.

The table includes eligible candidates only. Arena figures show reference scores and point gaps; prices apply to the stated model settings.

Raw score records for excluded entries are retained below as historical evidence, with their availability limits. They are not actionable recommendations.

06 / Updates

Evidence can change the selection.

Model availability, input requirements and prices can change. Evaluate models on tasks that reflect your application.

Review the model’s input formats, output settings and pricing when integrating. Contact Roten if you need help choosing a request format.