Many minds.
One answer.
Failure Model Labs merges models. Failure Tempest 1 briefs, drafts, fuses and audits across four frontier models, then hands you a single answer through an API you already know how to call.
OpenAI Chat Completions compatible · api.failure.fail/v1
The four models fused in Failure Tempest 1. Named so you know what you are paying for, not as endorsements.
Two ways to merge a model
A merged model is any model built from other models. We work on that from both ends: at inference time across closed APIs, and in the weights of open models.
Failure Tempest 1
Frontier models cannot share weights, so we merge them at inference. Each one drafts under a different lens, a merger settles every disagreement by checking, and a different model audits the result before it reaches you.
Expert transplant and weight merging
Open models can be merged for real. We transplant expert layers from other mixture-of-experts families into a frozen spine, and we merge fine-tunes of one base into a single checkpoint.
Five steps between your prompt and the answer
Easy prompts skip straight to one model. Hard prompts get the whole pipeline. You see the working stream past as reasoning, and you pay only for the answer.
Route
Easy prompts stream straight from the best-fit model. Nothing is fused that does not need to be.
Brief
A fast lane writes the task, difficulty, goal, traps, plan and a check. It can raise the difficulty. Cached for 24 hours.
Draft
Models draft under different lenses: direct, verify by an independent method, traps first, first principles.
Fuse
Agreeing drafts ship at once. Otherwise the merger settles each disagreement by checking, never by majority.
Audit
A different model audits the fused answer. Concrete issues send it back for revision.
Four models, live
This panel reads the API's own status endpoint. A lane joins the panel when its key is in place. Until every lane is ready, the API answers 503 with "not live yet" rather than serving a partial Tempest.
Target #1 on Artificial Analysis
That is the bar we set for Tempest 1. We have not run it yet, so every score here is a dash. When the numbers arrive they will be our own runs, and we will publish the losses alongside the wins.
Change the base URL. Keep your code.
Tempest speaks OpenAI Chat Completions. Point any OpenAI SDK at our base URL, pick a model id, and stream. Every brief and draft arrives as reasoning_content; the final chunk carries usage, cost and what the pipeline did.
Prompt and response caching
Resent conversation turns bill at the cached-input rate.
Spend caps per key
Set a monthly ceiling. Past it, requests stop, not your budget.
from openai import OpenAI client = OpenAI( base_url="https://api.failure.fail/v1", api_key="YOUR_API_KEY", ) stream = client.chat.completions.create( model="failure-tempest-1", messages=[{"role": "user", "content": "Prove there are infinitely many primes."}], stream=True, ) for chunk in stream: if chunk.choices: print(chunk.choices[0].delta.content or "", end="") if chunk.usage: # tokens, cached tokens, cost in USD, and the tempest stage print("\n", chunk.usage)
You pay for the answer
No hidden drafts
Every brief, draft and audit streams to you as reasoning, and bills as reasoning tokens. Nothing runs that you can't see.
Caching built in
Exact repeats at temperature 0 come back from cache at cached rates. Long conversations stop paying full price for old turns.
No training on your traffic
We do not train on API traffic. Request content lives only in short-lived caches measured in hours.
Fusion, configured for you
Dedicated spend caps, priority access to new lanes, custom routing and lens configuration, invoiced billing by arrangement, and straight answers to your security review.
Ask four models.
Get one answer.
Invite-only while Tempest 1 rolls out. Ask for an invite key, create your Failure ID, and make keys in the console.