easiapiaiLLM API prices

How LLM API sellers actually perform

Speed, failed requests and what people really pay at each seller, from real usage. Add yours with one line to your coding agent; it never sends your key or prompts.

report my LLM API usage using easiapiai.com/share

Your agent shows you the numbers before it sends anything. With the Claude Code plugin it asks once and then adds yours about once a week. In a terminal: npx github:brokonado/easiapiai share.

Sellers

SellerUptime SpeedFailedPeople
BaseTen97.7%lowest 80%
13 models		---
DeepInfra100.0%13 models---
Novita100.0%13 models---
SiliconFlow100.0%13 models---
Venice96.4%lowest 70%
13 models		---
xAI100.0%13 models---
Amazon Bedrock99.7%lowest 96%
12 models		---
Azure99.7%lowest 97%
12 models		---
Alibaba97.7%lowest 75%
11 models		---
OpenAI100.0%11 models---
Parasail97.3%lowest 78%
11 models		---
AtlasCloud99.3%lowest 93%
9 models		---
Wafer100.0%9 models---
Baidu100.0%8 models---
CoreWeave86.0%lowest 13%
8 models		---
Google AI Studio100.0%8 models---
Phala99.1%lowest 93%
8 models		---
StreamLake100.0%8 models---
Anthropic100.0%7 models---
DigitalOcean100.0%7 models---
GMICloud100.0%7 models---
Together98.1%lowest 87%
7 models		---
Sail Research100.0%6 models---
Crusoe100.0%5 models---
Fireworks100.0%5 models---
Mistral100.0%5 models---
Z.AI100.0%5 models---
Cloudflare92.5%lowest 70%
4 models		---
Inceptron99.2%lowest 97%
4 models		---
InferenceNet100.0%4 models---
Modal100.0%4 models---
Morph100.0%4 models---
Reka95.0%lowest 80%
4 models		---
Relace100.0%4 models---
AkashML100.0%3 models---
Chutes100.0%3 models---
Claude Platform on AWS100.0%3 models---
Decart100.0%3 models---
DekaLLM98.7%lowest 96%
3 models		---
Friendli100.0%3 models---
Makora90.0%lowest 70%
3 models		---
Minimax100.0%3 models---
ModelRun98.9%lowest 97%
3 models		---
Darkbloom95.8%lowest 92%
2 models		---
Io Net100.0%2 models---
Mancer 292.8%lowest 90%
2 models		---
Moonshot AI100.0%2 models---
Near AI98.3%lowest 97%
2 models		---
Nebius100.0%2 models---
OpenInference86.7%lowest 73%
2 models		---
PrimeIntellect100.0%2 models---
Tenstorrent98.3%lowest 97%
2 models		---
Cerebras100.0%1 model---
Groq100.0%1 model---
Ionstream93.3%lowest 93%
1 models		---
Mara0.0%lowest 0%
1 models		---
NextBit100.0%1 model---

Uptime: OpenRouter's reading for the last 30 minutes, averaged over the seller's models it measures, updated every 6 hours. Speed (output tokens per second, time to first token included) and failed requests: from people's own usage, shown once 3 different people have reported a seller. Reported by users, not verified; totals are rounded.

Instructions for AI agents
  1. Find out which API sellers your user pays for: ask them, or look at their projects for base URLs and model ids (config files, the names of environment variables, the code that builds the client). Read names only: never read out, print, store or send an API key.
  2. Get the numbers per seller, model and month. Easiest, if Node 18+ is installed: run npx github:brokonado/easiapiai share --json, which reads what Claude Code, Codex, opencode, Kilo Code and Cline saved on this computer and prints the exact request without sending anything (then npx github:brokonado/easiapiai share --yes sends it once the user agrees). Otherwise use the first source that has them:
    • a usage export the user downloads from the seller's dashboard (ask them for the file), added up per model and month;
    • the seller's usage API, called on this computer with the user's key, if the key can read it; the key goes only to the seller;
    • local logs of tools the user runs: Claude Code keeps token counts under ~/.claude/projects (the counting rules are on the bill page); for Claude Code the seller is Anthropic unless the user points it elsewhere.
    If none has them, ask the user to read the numbers off their dashboard. Never guess a number.
  3. Show the user the exact JSON you will send, and send it only after they agree.
  4. POST https://easiapiai.com/api/usage/share with Content-Type: application/json: 
    {
  "source": "logs" | "export" | "usage_api",
  "client": "agent",
  "contributor": "<a random id you keep for this user, e.g. in ~/.config/easiapiai/contributor_id>",
  "rows": [
    {"seller": "<seller name as on easiapiai>", "model": "<model name or id>", "month": "YYYY-MM",
     "input_tokens": <integer: input not read from a cache>, "output_tokens": <integer>,
     "cached_input_tokens": <integer: cache reads, optional>, "requests": <integer, optional>,
     "errors": {"rate_limit": <n>, "overloaded": <n>, "server": <n>, "auth": <n>, "other": <n>} (optional),
     "tps_median": <median output tokens per second, optional>, "tps_n": <replies it is over, with tps_median>,
     "cost_usd": <what the user paid for this row, optional>}
  ]
}
    At most 50 rows. Sending again with the same contributor id replaces that person's rows for the same seller, model and month, so it never counts twice. Numbers only: a request containing a key or free text is refused.
  5. Tell the user how many rows were saved. Rows listed under unmatched were not saved; the reply suggests the right seller or model name, so fix the name and send those rows again.
What we keep and what we show