How the models are scored: Arena quality, Efficiency, the FinOps badge and quality bars

Published 2026-10-18 · Prose last reviewed 2026-10-19 · Figures recomputed nightly from the 2026-10-22 catalogue snapshot

A FinOps guide by OptimNow · figures from the live OptimToken catalogue

Besides price, the site judges a model four ways: an Arena score for quality, an Efficiency score for quality per dollar, the FinOps Friendly badge, and the quality bars that match a use case to the models good enough for it. Only the first is measured, by someone else. The other three are OptimNow's arithmetic on it, and all three are relative: they rank models against each other, so they move when the market does, without anyone deciding they should.

1. The quality score: Arena

The quality score is the model's rating on LMArena, a public ranking built from blind votes: two anonymous models answer the same prompt, a person picks the better answer, and the ratings are fitted from those votes, the way chess players are rated from games. The site reads Arena's overall text ranking. Arena also publishes rankings by category, coding among them; the site does not read those today.

The scores are copied once a week, and every score is rewritten from the same read, never a few at a time. The reason is the Efficiency score below: it ranks models against each other, so mixing this week's scores with last month's would invent movement nobody measured. The date shown beside the scores is the leaderboard's own, not the day it was read: on this page, 2026-10-08.

Our model names and Arena's do not always match, so the link between the two is a written list, reviewed by hand, with three rules. When Arena scores a dated build, the most recent build counts. When Arena ranks a model only at a given reasoning setting, that score is used, and it reads somewhat higher than the default setting would. When Arena does not rank a model at all, the model has no score, and no score is ever guessed or borrowed from a sibling.

On the 2026-10-22 snapshot, 156 of the 316 models carry a score and 160 do not; by price tier, 17 of 31 Premium, 64 of 113 Mid-range and 75 of 164 Budget models are ranked, and the 8 image models are not ranked on token price at all. Most of the cheapest tier is unranked: 89 of its 164 models. No score means not measured, not measured low. Arena does not rank every model, ranks new releases only after enough votes, and drops old ones.

The best ranked model scores 1525 (Gemini 4 Argon) and the median ranked model 1426. A gap under 20 points does not separate two models on Arena, so the site treats it as noise. Scores also drift upward as votes accumulate, which is why nothing on the site uses a fixed score as a threshold.

What the score does not tell you: how a model handles your documents, your formats or a structured-output task, how fast it answers, or what it costs. It is one general benchmark of human preference on open-ended prompts.

2. The Efficiency score: quality per dollar, 0 to 100

The Efficiency score in the Business view blends two ranks across the 156 models that have both an Arena score and a token price:

• Quality percentile: 100 for the best Arena score, 0 for the lowest.
• Price percentile: 100 for the cheapest, 0 for the dearest, on a fixed blend of 30% input and 70% output price per 1M tokens.
• Score = 60% of the quality percentile + 40% of the price percentile, rounded.

Ranks, not values: a model one cent cheaper than its neighbour gains as much as one ten dollars cheaper. That keeps a $60 model and a $0.10 model on one scale. The 60/40 split favours quality on purpose, so a cheap weak model does not outrank a strong one on price alone. The price blend is fixed, so the score does not move with the use case selected, and the ranks are taken across every ranked model, so filtering the table never changes a score. Deprecated models and models quoted at an outlier listing are ranked too, because the table shows them; the table marks them, and the badge below refuses the deprecated ones.

Two examples on this snapshot. GLM 5.3 Flash has the highest Efficiency score: 60% of its quality percentile (84.5) plus 40% of its price percentile (96.1) gives 89. Gemini 4 Argon, the best ranked model, scores 67: 60% of its quality percentile (100.0) plus 40% of its price percentile (18.7) gives 67. A model with no Arena score has no Efficiency score; the table shows a dash.

3. The FinOps Friendly badge

A green "F" beside a model name in the Business view, for models that are good enough, priced well and safe to build on. A model needs all four:

• Quality: an Arena score in the top 40% of ranked models, which today means 1442 or more.
• Value: an Efficiency score in the top 30%. Scores are whole numbers, so models tied on the cut-off score all clear it, and slightly more than 30% can.
• Cost: a list price among the cheapest 70% of ranked models, which today means $3.41 or less per 1M tokens on the 30/70 blend, and that price not an outlier listing: one reseller's near-free input against a dear output, which is not what most hosts charge.
• Stability: no preview, beta or experimental label, not a model in the Premium price tier released in the last 3 months, and no retirement announced by its maker.

The cost rule is a gate, not extra weight: at a 60/40 blend the best ranked model earns 60 points on quality alone, which is how the dearest model in the catalogue once earned the badge. Prices are list prices only, because cache and batch discounts depend on the customer using them. All four rules are percentiles, so the badge recalibrates as the market moves. On this snapshot 23 of the 156 ranked models carry it: MiMo-V2.6-Pro, MiMo-V2.5-Pro, DeepSeek V4.1 Flash, MiMo-V2.6-Flash, Gemma 4 31B, Gemini 3.8 Flash, Hy3, DeepSeek V4 Pro 0423, Muse Spark 1.2, Muse Spark 1.3, Gemini 3.7 Flash, GPT-6 Luna, Muse Spark 1.1, Gemini 3.6 Flash, Grok 4.20, DeepSeek V4 Pro 0813, Qwen3.7 Plus, GPT-5.6 Luna, Grok 4.20 Multi-Agent, Gemini 3.5 Flash Lite, Kimi K2.6, GLM 5 and Qwen3.6 Plus.

4. Quality bars: how good a model has to be for a use case

The Business view prices 8 use cases. The quality-bar method gives each one a minimum Arena score and finds the cheapest model above it. A use case's level is decided from the work, not the model: what a wrong answer costs, who catches it before it does harm, and whether one error feeds the next. Each level is a distance below the best ranked model, never a fixed score, so it survives Arena's upward drift. Measured from 1525, the best score among models that are not deprecated:

• Frontier-grade, within 25 points of the best: Arena 1500 or more today, for coding task and agent workflow.
• Strong, within 60 points of the best: Arena 1465 or more today, for support ticket, knowledge Q&A and marketing content.
• Standard, within 110 points of the best: Arena 1415 or more today, for meeting summary, invoice processing and call summary.

The first distance is the benchmark's own noise, 20 points, plus 5, so it means "as good as the best, as far as Arena can tell, or nearly". The other two are OptimNow's working assumptions from AI cost reviews, not measurements, and so is the choice of level for each use case: an evaluation on a team's own traffic replaces them, and should. The levels describe quality only; they are not the Premium, Mid-range and Budget price tiers, and a model in any price tier can clear any bar. The quality-bar guide lists the cheapest model above each bar and what each use case costs on it.

What none of the scores do

• They do not replace an evaluation. They shortlist; a test on your own prompts decides.
• They are relative. A model's Efficiency score, its badge and the bars can move because another model was added, repriced or rescored, while nothing about the model itself changed.
• They read one price per model. The price is one host's published rate, the one the site quotes; most models are sold by several hosts at different rates, and the model page lists them.

Figures on this page come from the 2026-10-22 price snapshot and Arena scores as of 2026-10-08, and rebuild nightly.

Data snapshot: 2026-10-22 · Compare all models · Cloud compute pricing · All guides