The model observatory
Intelligence. At the right cost.
Explore the trade-offs. Find the model that fits your work.
AA classes are quartiles within each output modality · ⚡ speed: A 3, B 2, C 1, D 0 · — insufficient AA data · ? speed unavailable · $–$$$$ consolidated price quartiles (15:1); free = zero cost
Using reported catalog speeds. Provider and test conditions may vary.
More intelligence. Less spend.
Estimated cost per task × intelligence. Bigger bubbles mean faster output.
No comparable measurements for this selection. Try another task or maker.
Hover, focus or click a bubble to reveal its name. Click the name to open the model card.
0/0 models with comparable cost and quality. Hollow dots: speed unavailable. Dashed line: efficient frontier.
Cost, speed and intelligence. Together.
Lower cost is farther left; the top is faster. Larger bubbles: more intelligence.
Speed (tokens/s): select a measure below. Task time is output tokens ÷ tokens/s; latency is time to first token.
No models with all three measurements in this selection.
0/0 models plotted with complete data. Color = maker.
How the balanced ranking works
Each model receives a normalized position among distinct values for lower cost, higher intelligence and faster output. The balanced score is the average of the three (equal weights). Ties share the same position. Only models with all three measurements enter this ranking.
This is a selection aid, not a new intelligence benchmark. Scores change with the task, workload and selected makers; a 90 does not mean 90% task success.
What does scale cost?
Estimated spend for 10,000 tasks, among models on the efficient frontier.
No comparable measurements for this selection. Try another task or maker.
The wait behind the answer
Estimated output generation time for this workload. Excludes first-token latency, tools and reasoning overhead.
No verified speeds for this provider selection. Reported measurements may be available using the speed selector.
Does the provider change the speed?
The current catalog does not provide matched official-provider and faster-provider measurements with source and test conditions. A speedup comparison is unavailable; reported speeds above must not be interpreted as a provider advantage.
How to read these comparisons
Cost estimates use 1,000 input + 300 output tokens, at Muitto catalog USD token rates (which may include markup and cheapest-provider routing). Both rates are required. Cache discounts, hidden reasoning tokens, tool calls and volume discounts are excluded.
All charts and their intelligence and balanced rankings use only the Artificial Analysis Intelligence Index supplied by the catalog. Task selection filters models; it does not turn this general index into a task-specific benchmark. Models without the AA index are excluded from intelligence comparisons. No Muitto score is substituted.
Bubble area is proportional to tokens per second; the same scale is used across the selection. Missing measurements are never zero. Official-provider mode requires explicit attribution; reported mode may mix providers and test conditions.
The efficient frontier contains models for which no other model is both cheaper and at least as capable, or more capable at the same price.
Chart inspiration: Artificial AnalysisCatalog: Muitto · Scores retain their original source and scale.