Tools · Models · 02

Compare models for your workload, not with one universal ranking.

Set request size and filter by measured intelligence, latency, context or cost. The explorer keeps provider pricing separate from independent Artificial Analysis measurements so you can see the trade-off among quality, speed and spend.

QualityIntelligence Index
Performancetokens/s + TTFT
Pricecurrent first-party API

Workload

Displayed cost uses these quantities and current standard rates. Published long-context pricing rules are applied automatically when relevant.

Filters

0 = no limit.
0 = no limit.
0 = no limit.

View

Visible
On frontier

Loading data…

Highest index
Lowest cost
Highest speed
Lowest TTFT
Comparison mapselect a point to inspect its provenance
Intelligence Index

Cost per request

Price/intelligence frontier: a model is included when no other model achieves both an equal-or-higher index and an equal-or-lower cost for the workload you defined.

Comparable data

The same comparison, without hiding dimensions.

Reasoning variants are compared as distinct configurations. A higher index does not mean a model is better for every task.

ModelCostIndextokens/sTTFTContextRead
Method

What the numbers mean.

Cost. The explorer applies each provider's current public standard price to the token counts you define. For GPT-5.6 it also applies the published multiplier when input exceeds 272k tokens. Cache, tools, search, discounts, batch pricing and premium tiers are excluded.

request_cost = Tinput × Pinput / 10⁶ + Toutput × Poutput / 10⁶

Intelligence Index. This is Artificial Analysis's composite index for the named configuration. It aggregates multiple evaluations covering reasoning, knowledge, coding and tool use. It is a useful comparison signal, not a universal measure of "intelligence" and not a substitute for your own evals.

Output speed and TTFT. Artificial Analysis measures them against live APIs. tokens/s describes generation speed after output starts; TTFT measures time to the first token. For reasoning models, latency can be dominated by deliberation and vary substantially with the selected reasoning effort.

Frontier. There is no hidden composite score and no arbitrary weighting. The Pareto frontier only identifies configurations that are not simultaneously dominated on both cost and Intelligence Index.

Data verified August 21, 2026. Performance measurements and benchmarks change; this is deliberately a small curated set containing only configurations for which price, specifications and performance can be maintained with explicit provenance.

Sources and verification date