For every task, three models: the one that can do the job, the best value, and the best regardless of price. With what each costs a month, for the volume you choose.
WoBench ratings are based on human evaluations.
319 models10 tasks9 public sourcesno sponsor in the verdicts
11.3 times cheaper than the best on this task, which costs $22.00 a month.
At the Small volume: 20 runs per month, of 100,000 tokens in and 2,000 out.
This model was measured with a "high" reasoning effort: its output is counted 3 times.
redo the calculation
22.7 times cheaper than the best on this task, which costs $44.40 a month.
At the Small volume: 500 runs per month, of 12,000 tokens in and 600 out, 30% of it cached.
This model was measured with a "xhigh" reasoning effort: its output is counted 3 times.
redo the calculation
Preferred 5 times out of 10 to the one that does the job
$11.00a month
402 times the price of the one that does the job, which costs $0.0274 a month.
At the Small volume: 200 runs per month, of 3,000 tokens in and 500 out.
redo the calculation
Prices are those published by the providers and can change without notice. Our verdicts guide a choice, they do not guarantee it: test the model on your own cases and validate your calculation, economics included.
Our method
240
The price gap between the one that does the job and the best.
On Sort and extract, sort, tag and extract structured data at high volume. The, the cheapest that does the job costs $0.0534 a month, the best $12.80. 132 models measured, 56 meet the bar.
Scores from October 20, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
7.6
The price gap between the one that does the job and the best.
On Search the web, search and summarise information on the web. The, the cheapest that does the job costs $0.95 a month, the best $7.20. 27 models measured, 14 meet the bar.
1st of the 27 models this task judges. 8 times the price of Can do the job, for 56.3 measure points more.
This model was measured with a "xhigh" reasoning effort: its output is counted 3 times.
Scores from October 13, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
374
The price gap between the one that does the job and the best.
On Summarise, summarise a short to medium text, article or report. The, the cheapest that does the job costs $0.0321 a month, the best $12.00. 147 models measured, 59 meet the bar.
Scores from October 20, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
329
The price gap between the one that does the job and the best.
On Synthesize, summarise or analyse 50K to 1M tokens. The, the cheapest that does the job costs $0.0669 a month, the best $22.00. 104 models measured, 58 meet the bar.
Scores from October 20, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
22.7
The price gap between the one that does the job and the best.
On RAG, answer questions from the documents you supply. The, the cheapest that does the job costs $1.96 a month, the best $44.40. 31 models measured, 18 meet the bar.
23 times cheaper than The best. 3 cheaper models do not touch the bar, not even with the upper end of their range.
This model was measured with a "xhigh" reasoning effort: its output is counted 3 times.
1st of the 31 models this task judges. 23 times the price of Can do the job, for 59.3 measure points more.
This model was measured with a "high" reasoning effort: its output is counted 3 times.
Scores from September 13, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
509
The price gap between the one that does the job and the best.
On Write, write texts, emails and editorial content. The, the cheapest that does the job costs $0.0147 a month, the best $7.50. 147 models measured, 63 meet the bar.
77% cheaper than The best, with no measurable gap, on 8,136 votes against 8,040. Statistically tied with Can do the job as well: the three published intervals overlap.
Scores from October 20, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
289
The price gap between the one that does the job and the best.
On Web Dev, a whole website or web app from a brief, judged on the result. The, the cheapest that does the job costs $0.0364 a month, the best $10.50. 82 models measured, 28 meet the bar.
Scores from October 19, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
8.9
The price gap between the one that does the job and the best.
On Agent, long loops with tool calls and recovery from errors. The, the cheapest that does the job costs $1.47 a month, the best $13.05. 28 models measured, 9 meet the bar.
No middle ground here. Between $1.47 and $13.05, 4 models meet the bar, but none does measurably better than Can do the job.
The furthest along covers 28% of the way to the best, where 35 is needed, and its range overlaps that of Can do the job.
Those close enough to the best cost as much as it does, or more.
Scores from October 20, 2026, prices from October 22, 2026.
Small volume. The measure details and the full model list are on the task page.
WoBench ratings are based on human evaluations.
402
The price gap between the one that does the job and the best.
On Analyse images, describe, explain or answer a question about a photo, a screenshot or a chart. The, the cheapest that does the job costs $0.0274 a month, the best $11.00. 72 models measured, 36 meet the bar.
86% cheaper than The best, with no measurable gap, on 4,524 votes against 11,304.
This model was measured with a "high" reasoning effort: its output is counted 3 times.
See what they cost, from the cheapest to the most expensive.
One dot per model and per task.
the cheapest that does the job
best value
the best
free right now (in its zone, position follows the score, unrelated to any price)The bigger the dot, the higher the score.
Compared with the entry price, scale fitted to wide gaps
6 tasks of 10 have a free model that clears the bar.
Providers sometimes give access to a model for free. We show it apart, without letting it compete, because a free price says nothing about the service delivered. Each cell answers a single question: on this task, does this model reach the bar? Same measures, same bars as the verdicts.
Free offers and prices are those published by providers, and may change or stop without notice. Our verdicts guide a choice, they do not guarantee it: test the model on your own cases and check your numbers, including the economics, planning for what the same work will cost on the day the offer stops.
Duration
Until when?
No offer announces its end. The 2 we have seen stop lasted from 7 to 10 days.
Limits
How many requests?
Free access may be capped in requests or throughput. Our sources do not publish these caps.
Your data
Who reads your requests?
Free access may allow the provider to reuse what you send. Read its terms before sending customer data.
Tracked since September 30, 2026, and still ongoing.
No model scored since the previous publication. Here are the latest three added to the catalogue.
Nova 2 Lite
too little data
Amazon, added to the catalogue on September 3, 2026. Not enough votes to place it.
Nova Lite 1.0
too little data
Amazon, added to the catalogue on September 3, 2026. Not enough votes to place it.
Nova Micro 1.0
does not do the job
Amazon, added to the catalogue on September 3, 2026. On none of our 10 tasks.
×509
Figure of the day
The cost gap between the best and the cheapest that does the job, on Write.
The cheapest that does the job, GLM 5.3 Flash, costs $0.0147 a month. The best, Claude Fable 5, costs $7.50. At 100 runs per month, for a task both can do.
Prices: OpenRouter (providers), collected on October 22, 2026.
Follow the models you use: on the task you care about, each one says where it stands, and what it costs compared with the one that does the job. The rest of the time, we stay quiet.
Free. Leave whenever you like, we delete everything.
Where these numbers come fromEuropean Central BankEpoch AILiteLLMLMArenamodels.devOpenRouter (providers and per provider prices)OpenRouter (models)tau²-bench (Sierra Research)Vectara (hallucination leaderboard)All sources, licences and datesPrices are collected once a day, at 3:00 am (Paris time).
These figures come from the public sources cited and reflect the last import, not the state at this second. Prices are those published by the providers and can change without notice: check them there before committing.
Costs are computed on a reference volume per task, shown on the page of each task: your usage will give another amount.
Our method