The best AI models, by task
Eight rankings from benchmark evidence and live prices, on the same quality projection that routes Inferbase traffic. Missing evidence is shown as missing, never guessed.
Quality index is a 0 to 100 percentile among measured models. The full overall ranking →
Eight rankings, one method
Each card states how its list is ordered: measured quality, verified capability then quality, or specification and price where no quality evidence exists.
Ranked by measured coding benchmarks, not editorial opinion, with current pricing and context windows alongside every entry.
View the rankingRanked by knowledge and graduate-level reasoning benchmarks (MMLU-Pro, GPQA, HLE), with pricing and context windows for every entry.
View the rankingThe general-purpose ranking: ordered by composite intelligence evidence, with the price and context data needed to judge the tradeoff.
View the rankingOnly models with affirmatively verified tool support, ranked by measured quality, because a silently mangled tool call is worse than a refusal.
View the rankingModels with verified structured-output support, ranked by measured quality, for pipelines where a malformed response is a production incident.
View the rankingModels that accept image input, ranked by general quality evidence, since no independent vision benchmark suite is ingested yet, and this page says so.
View the rankingRanked by advertised context window, an objective spec, with the quality and price data needed to judge whether the headline number is worth it.
View the rankingEmbedding models compared on price and specs. No quality benchmark suite for embeddings is ingested yet, and this page makes no quality claim.
View the rankingHow these rankings are made
Generated from the Inferbase catalog, with the same quality projection the router consults for live requests.
Independent benchmark suites are ingested, normalized within each suite, and projected into per-task quality scores. Rankings combine that quality signal with objective specifications and current pricing, and they regenerate automatically as the data changes.
The index is a 0 to 100 percentile among the models with measured evidence for the task: relative, never absolute. No ranking is published for a task without measured evidence, which is why some popular tasks do not have a page yet, and a model without results on the covered suites has no index and no rank rather than a guess.
Pages regenerate at least hourly; benchmarks refresh weekly and prices on the catalog sync cycle. The full methodology.
The fainter marks are the other two floors, permissive lower and strict higher. Price is compared among the models marked in; a preset that leaves nobody in relaxes one step, and the relaxation is on the decision record.
Frequently asked questions
What best means per task, how benchmarks rank models, what the index is, and how to act on the rankings.
It depends on the task, which is why this site publishes one ranking per task rather than one list. The overall page orders models by a general intelligence index; the coding, research, agent, structured-output, vision, long-context and embeddings pages each rank on evidence specific to that job. The model at the top of one ranking is often several places down another.
Results from independent benchmark suites are ingested per model, normalized within each suite so that a score is comparable across suites of different difficulty, and combined into a per-task quality score. That projection is the same one the Inferbase router consults when it routes production traffic, so the rankings and the routing rest on one set of numbers.
A 0 to 100 percentile among the models with measured evidence for that task: a model at 90 sits in the top tenth of the measured pool for that job. It is relative, not absolute, and it is only computed where evidence exists; a model with no published result on the covered suites has no index and is not ranked.
Each page names its suites: code-generation evaluations such as SciCode, Terminal-Bench Hard and LiveCodeBench for coding; MMLU-Pro, GPQA and Humanity's Last Exam for research; a general intelligence index for the overall ranking; tool-use and long-context reasoning evaluations where they exist. Capability pages (function calling, structured output) first verify support and then order by quality; the long-context and embeddings pages order on specifications and price and make no quality claim.
Read the tables rather than take a position: every ranking lists open-weight and proprietary models side by side on the same evidence, and on most tasks several open-weight models sit within a few index points of the leader at a fraction of the per-token price. The price gap is usually larger than the quality gap, which is the case for routing.
Because no covered suite has published results for it on that task. A missing model is not a judgment; there is no measured evidence to rank it on, and these pages do not substitute estimates for evidence. It appears as soon as covered evidence for it lands.
Because no independent benchmark evidence for them is ingested yet. Pages exist only for tasks where the ranking can be defended with data; new task pages are added as evidence coverage grows.
The ordering is computed from measured evidence by the engine that routes production traffic, refreshed automatically as benchmarks and prices change, and each page states its method and how many models qualified. Where evidence does not exist, the pages say so explicitly instead of substituting opinion.
Pages regenerate at least hourly from live catalog data, and each ranking shows when it was last refreshed. Benchmark results are refreshed weekly, prices and model availability on the catalog sync cycle, and a new model appears in a ranking as soon as covered evidence for it lands.
Not necessarily. The rankings regularly show models within a few index points of the leader at a tenth of the price, and most workloads mix easy and hard requests. The economically rational pick is the cheapest model that clears the bar for each request, which is what routing does per request rather than once.
No. Models are ranked on evidence whether or not Inferbase serves them; the ones it does serve carry a badge, and the rest stay in the list with their scores. The rankings are a view of the catalog, not of the managed pool.
Models marked as served on Inferbase are available through one OpenAI-compatible API: pin any of them by id, or send model="auto" and let the router pick per request from the same quality data these pages display. Models not served can still be compared and, with your own provider keys connected, served on your account where the provider offers them.
Yes. Send model="auto" and each request is classified for its task, narrowed to the models that clear the quality bar for it (the same projection behind these rankings), and served by the one that wins on your objective: cost, balanced, quality or latency. The decision comes back with the response and is kept by request id.
Put intelligence in the middle.
One gateway in front of every model, with your policies applied and every decision on record. Start on the free plan with 5,000 requests a month, no card required.