CrofAI OpenAI-compatible

Aggregator that resells open-source models behind one cheap key.

⚠️Heads up: Subscriptions are being reworked (announced 2026-05-31) — likely off the weekly bench from 2026-W23. CrofAI says flat-rate subs are no longer sustainable. Either subs convert to credit-based plans worth 1.15× their PAYG value (a $5 Hobby sub becomes $5.75 of credit/month) or they are removed entirely in favour of rock-bottom PAYG. Existing subs keep working until they end or renew; the change rolls out 2026-05-31 through ~06-03. The flat, unlimited $5 Hobby plan that made CrofAI our volume daily driver is going away, so under metered PAYG we expect to stop running the full weekly suite against it. Historical data stays on this page for context.

CrofAI is an inference aggregator: one API key, many open-source models (DeepSeek, GLM, Kimi, Qwen, MiniMax, Gemma). Two of its models are completely free, and the paid ones are positioned as the cheapest places to run their respective open weights.

Strengths

  • One key, many open-source models
  • Two fully free models (Qwen3.5-9B, GLM-4.7-Flash)
  • Standard OpenAI endpoint — works with any compatible client

When to use it

  • Trying many open-source models under one key
  • Exploration before committing to a single provider
  • Quick MVPs and prototypes

Subscription plans

PlanPriceQuotaAvailable
Free / PAYG$0/moPay-per-token, no recurring chargeyes
Hobby$5/mo500 daily requests · access to all modelsyes
Pro$10/mo1,000 daily requests · priority supportyes
Intermediate$20/mo2,500 daily requestsyes
Scale$50/mo7,500 daily requestsyes
Max$100/mo15,000 daily requestsyes
Notes: Subscription tiers add daily request quotas on top of the PAYG endpoint; all tiers see the same model catalogue. See crof.ai/pricing for current rates.

Models tested on CrofAI

Speed numbers below are specific to CrofAI's routing and hardware. The same model may appear on other providers' pages with different throughput.

2026-04-26 2026-05-10 peak 614 tok/s
Best tok/s observed on CrofAI per weekly snapshot (2 points).
Model Best tok/sAvg tok/s RunsSuccess Longest output (chars)
kimi-k2.5-lightning614.4558.43100%5,196
qwen3.5-9b174.7156.93100%2,592
qwen3.5-9b-chat158.0148.93100%3,644
qwen3.6-27b150.7127.23100%2,895
kimi-k2.5114.092.43100%5,467
glm-4.7-flash113.2113.2333%3,245
glm-5.1108.074.53100%4,600
kimi-k2.6105.873.73100%5,201
glm-5105.1105.1333%2,255
greg103.8103.8333%4,380
minimax-m2.5102.883.53100%3,352
gemma-4-31b-it101.250.23100%3,469
deepseek-v3.293.068.63100%3,232
kimi-k2.6-precision80.765.13100%3,901
glm-5.1-precision72.264.23100%4,217
deepseek-v4-pro59.658.13100%4,699
glm-4.757.850.23100%3,521
qwen3.5-397b-a17b52.752.7333%3,018