Model performance on MindsHub

See how models served through MindsHub Inference perform: response times, availability and cost, with current prices, API aliases and fallback routes.

Measurements refresh every 2 to 6 hours · About the data

All makers
Makers
Reasoning effort

Fallback map

When the requested model cannot answer, MindsHub tries these backup models in order. The grid labels backups 1st, 2nd and so on. Counts show how many API names use each backup across the catalog. Filters match source names and makers; every backup for those sources stays listed.

Showing 70 names with fallbacks

What each name falls back to, and in which place of its list
Asked for deepseek-v4-flash 23 names list it sonnet 19 names list it gpt 14 names list it gemini-flash 13 names list it kimi 10 names list it gemini 9 names list it opus 6 names list it gpt-mini 6 names list it haiku 5 names list it glm-5-3-flash 3 names list it gpt-nano 2 names list it gpt-oss-fireworks 1 name lists it
mindshub_air 1st
mindshub_blaze 1st
sonnet and sonnet-5-5 , sonnet-5 1st
opus and opus-5-5 , opus-5 1st
fable and fable-5-1 , fable-5-2 , fable-5 1st
haiku and haiku-4-5 , haiku-5-5 2nd 1st
gpt and gpt-6-astra , gpt-5-6-sol 2nd 1st
gpt-sol and gpt-sol-6 , gpt-sol-6-1 2nd 1st
gpt-terra and gpt-terra-5-6 , gpt-terra-6-1 1st 2nd
gpt-luna and gpt-luna-5-6 , gpt-luna-6 1st 2nd
gpt-codex and gpt-codex-5-3 1st
gpt-mini and gpt-mini-5-4 1st 2nd
gpt-nano and gpt-nano-5-4 1st
gemini and gemini-3-1-pro 1st 2nd
gemini-argon and gemini-argon-4 1st 2nd
gemini-flash and gemini-flash-3-8 , gemini-flash-3-7 , gemini-flash-3-6 , gemini-flash-3-5 , gemini-flash-3 2nd 1st
gemini-flash-lite and gemini-flash-lite-3-1 2nd 1st
kimi and kimi-k3 1st
kimi-pro and kimi-k3-1-pro 1st
deepseek-v4-flash and deepseek-v4-1-flash-1017 , deepseek-v4-1-flash 1st
qwen and qwen-3-8-a95b 1st
glm and glm-5-3 1st
glm-5-3-flash and glm-flash-5-3 1st
muse-spark and muse-spark-1-3 , muse-spark-1-2 , muse-spark-1-1 1st 2nd
muse-wonder and muse-wonder-1-0 1st 2nd
muse-ember and muse-ember-1-0 2nd 1st
grok and grok-4-7 , grok-4-6 , grok-4-5 1st 2nd

No fallback: gpt-oss-fireworks

About the data

What does this page show?

Performance and reliability measured through MindsHub Inference, alongside current token prices, API aliases and fallback routes. The figures describe the models as served through MindsHub. They are not a ranking of every provider’s deployment of those models.

Where do the measurements come from?

We run automated API checks and writing tasks through MindsHub Inference. Complex task, writing speed and cost per task figures refresh every two hours, and the other measurements every six hours. Complex task performance and cost cover up to 35 days of tests. Each chart shows its measurement dates. If an update fails, we keep previously saved measurements with their original dates. Prices, aliases and fallback routes come from the same live catalog as our pricing page. A dash means no usable measurement is available.

How is Ping measured?

Every 10 minutes, checks from multiple regions ask for a one-word reply at the lowest reasoning effort. We time the first answer token, including any thinking before it. Ping measures basic responsiveness and tracks availability.

Sample prompt

Reply with the single word: ok
How is the Simple task measured?

Every six hours, the model receives a short briefing about a fictional market hall and is asked to describe a Saturday morning in at least 150 words. It must interpret the facts, decide how to present them and follow the writing constraints. We time the first answer token at the default reasoning effort, including thinking time.

Sample prompt, abridged; briefing omitted

Using only the supplied market-hall briefing, describe a typical Saturday morning in at least 150 words. Use plain paragraphs, with no headings or lists, and do not add facts.
How is the Complex task measured?

The model receives a longer briefing about the same fictional market hall and must turn it into a complete visitor’s guide of at least 800 words. It must analyze and organize the supplied facts, cover them in order and obey content and formatting constraints. We measure the first answer token, total response time and writing speed at each tested effort level. The first-answer and total times include thinking; writing speed starts when the answer begins. Repeated tests run over several weeks; each model’s details show the test count and interval. These are speed tests, not scores for reasoning quality.

Sample prompt, abridged; briefing omitted

Using only the supplied briefing, write a visitor’s guide of at least 800 words. Cover every part in the given order, in plain paragraphs with no headings, lists or tables. Do not add facts.
What are p50, p05 and p95?

P50 is the median: half of measurements are at or below it and half at or above it. P05 and P95 mark the lower and upper ends of the middle 90% of measurements. The leaderboard shows p50. Open Performance to explore the spread; Ping and Simple task have p50 and p95 only.

Does availability below 100% mean requests failed?

Not necessarily. A configured fallback route can still answer when the model you chose does not. Availability before fallback is the share of our automated checks answered by that model, including retries, over the selected period. Other checks were answered by a fallback or got no answer. These checks do not measure how often customer requests fail. Failures attributed to our monitoring network are excluded.

What does the intelligence score mean?

MindsHub’s read is a subjective estimate from 1 to 10 based on public benchmarks and rankings, not a precise measurement. Performance depends on your workload; test models on your own tasks.

For detailed benchmarks, refer to third party services such as Artificial Analysis, Epoch AI or similar.

How is cost per task compared?

The most expensive measured model and effort level is the 100% reference, updated automatically as results change. The leaderboard range covers a model’s measured effort levels; expand it for individual values. In Cost per task, choose another reference to compare against. Dollar estimates use mean input and output tokens at current standard rates, excluding cache discounts. Your bill depends on your requests.

What changes with reasoning effort?

Intelligence, Complex task timings and writing speed follow the selected effort. Ping always uses the lowest effort and Simple task the default. The cost range covers all measured levels. Views showing all effort levels disable the selector because every level is already visible.

What are fixed and rotating aliases?

An alias is the name sent in the API’s model field. Fixed aliases select a specific model version. Rotating aliases follow a model family as it updates. MindsHub Air and MindsHub Blaze are managed rotating aliases: MindsHub can change the model behind each name. The Aliases tab explains their purpose and lists the names you can use.

How do token prices, context limits and the free tier work?

Tokens are pieces of text. The context limit is how much input a model accepts in one request. Token prices are US dollars per million input or output tokens, from the same catalog as the pricing page. MindsHub Air is free within the free-tier limits; its listed token rates apply to usage beyond those limits. TypeSafe’s Jev is free. See Prices for cached-input, cache-write and other rates.

Which models are included?

The leaderboard covers chat models and their managed aliases. Other models lists decision and embedding models with their own response-time measurements and availability. See mindshub.ai/unified-inference#free-tier for the free tier and mindshub.ai/pricing for rates and credits.