▸ free ai models usage
new: genuinely free AI models are live — try surp/free + see live budgets · token-gating prototype · vote on SRP

free AI models

surp/free — real responses, zero payment, treasury-sponsored.

surp/free gives users genuinely free LLM inference through the same OpenAI-compatible endpoint. There is no x402 payment challenge and no API key. The surp treasury pays the upstream model cost, with strict daily budgets, per-IP limits, cheap-model guardrails, and automatic fallback.

important: We do not resell or proxy personal free-tier credentials from OmniRoute's catalog. Many provider free tiers prohibit proxying or third-party access. OmniRoute's MIT-licensed methodology powers our catalog intelligence; actual public inference is purchased through our normal Surplus account and sponsored by surp.

try it now

curl -X POST https://surp.ivc.lol/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"surp/free",
       "messages":[{"role":"user","content":"Explain x402 in one sentence"}],
       "max_tokens":64,
       "stream":false}'

Limits: 20 requests/IP/day · 128 max output tokens · non-streaming · no tool calls · availability is best-effort.

live sponsored budget (UTC)

500
requests remaining today
100.0K
tokens remaining today
22
eligible live models
0ms
average free latency
requests served today0 / 500
tokens served today0 / 100,000
failed upstream attempts0
maximum sponsored model price$0.1000 / 1M tokens
fallback attempts per requestup to 3 live models

live models we may sponsor

These are real, liquid Surplus marketplace models currently below the configured treasury ceiling. The router tries the cheapest first and falls back when a seller is unavailable.

#modelUSD/1Mhealthy sellersrequests 24hvolume 24h
1glm-5.2$0.012814849,54919,299,738
2glm-4.7$0.02151777,4841,808,669
3gemma-3-27b-it$0.025210912
4deepseek-v4-flash-0731$0.02764216,6254,489,995
5qwen3-235b-a22b-2507$0.028012200
6deepseek-v4-flash$0.0297114281,68573,498,700
7mistral-small-3.2-24b-instruct$0.0344105480
8glm-4.5-air$0.0390365053
9glm-4.7-flash$0.0414123631,205
10e2ee-qwen-2-5-7b-p$0.043283180
11qwen3-5-9b$0.05008162,237
12gemini-2.5-flash-image$0.05002300
13e2ee-gpt-oss-20b-p$0.05758100
14glm-4.6$0.065417754293
15qwen3-235b-a22b-thinking-2507$0.08008370815
16gemma-4-31b-it$0.08258516095,584
17glm-4.5$0.084028501,711
18llama-3.3-70b-instruct$0.084010464369
19gpt-5-nano$0.090038581,272
20openai-gpt-oss-120b$0.092411454378
21nvidia-nemotron-3-nano-30b-a3b$0.09371015625
22gpt-5.6-luna$0.09809053,31614,211,585

models actually served today

modelfree requeststokens
No free requests served today yet.

what we adopted from OmniRoute

OmniRoute is an MIT-licensed AI gateway with an unusually careful free-tier catalog. We incorporated the parts that improve honesty and transparency:

  • Pool deduplication: if ten models share one provider quota, count the pool once using its maximum—not ten times.
  • Recurring vs one-time separation: signup credits never inflate the steady monthly headline.
  • Uncapped-provider honesty: permanently free providers with no published token cap are listed but never multiplied by RPM × 24/7.
  • ToS risk labels: each catalog entry is classified as ok, ambiguous, caution, avoid, or unknown.
  • Quota visibility: used, remaining, reset/budget limits and actual served models are public.
  • Automatic fallback: failed free models are skipped and the next live candidate is tried.

OmniRoute free-tier catalog snapshot

Source: OmniRoute release/v3.8.50, curated 2026-07-22, MIT License. Snapshot stored with license attribution in this repository.

1.53B
documented recurring tokens/month
2.15B
realistic first month + credits
43
deduped recurring pools
519
catalog model entries
providers represented81
permanently free, uncapped providers13 (listed, never summed)
ToS generally ok49
ToS ambiguous38
ToS caution334
ToS avoid93
catalog entries after excluding avoid426

why not directly pool all those free tiers?

Because "free" often means free for one developer's account—not free to resell through a public proxy. OmniRoute itself surfaces these ToS risks. surp will only integrate a third-party free provider into public routing when its terms explicitly allow proxying/resale or we have written permission. Until then, the catalog is informational and our public free requests are treasury-sponsored.

failure behavior

  • If a model has no healthy sellers, it is excluded before routing.
  • If the first model returns an error, the router tries up to two more candidates.
  • If the daily global budget is exhausted, the API returns HTTP 429 with a paid x402 alternative.
  • If one IP reaches its daily allowance, only that client is throttled.
  • If every sponsored model fails, the API returns structured JSON—never a paid settlement followed by an HTML 500.

related: live system status · cache-aware routing · reward proposal · token-gated access · API docs