Failure Tempest 1
Four frontier models, one answer. Tempest 1 routes easy prompts to the best-fit model and sends hard ones through a brief, lensed drafts, fusion and an independent audit. It is our first merged model, and it merges at inference.
POST /v1/chat/completions returns 503 service_unavailable with the message "not live yet". The status page shows each lane as it joins.At a glance
| Field | Value |
|---|---|
| Name | Failure Tempest 1 |
| Kind | API fusion — merged at inference, not in the weights |
| Panel | Claude Fable 5.5 · Claude Opus 5.5 · GPT-6 Astra 6.2 · Gemini 4 Argon |
| Model ids | failure-tempest-1-flash · failure-tempest-1 · failure-tempest-1-max |
| Interface | OpenAI Chat Completions compatible, streaming supported |
| Base URL | https://api.failure.fail/v1 |
| Reasoning trace | Streamed as reasoning_content deltas |
| Tool calls | Sent to a single tool-capable model, not fused |
| Caching | Conversation prefix (1 h), exact-repeat responses (1 h), briefs (24 h) |
| Training on your traffic | No |
| Evaluation | — (not yet run; see evaluation) |
| Status | Lanes awaiting keys |
Route, brief, draft, fuse, audit
The pipeline is spent where it earns its keep. A greeting does not need four models. A proof does.
Route
Easy prompts stream straight from the best-fit model.
Brief
TASK, DIFFICULTY, GOAL, TRAPS, PLAN, CHECK. The brief can raise the difficulty. Cached 24 h.
Draft
Deep prompts get at least three drafts under different lenses, reusing a model under a new lens if needed.
Fuse
Same final answer everywhere? Ship now. Otherwise check each disagreement and hunt shared errors.
Audit
A different model audits. Concrete issues trigger a revision.
What makes it a merge, not a vote
Lenses, not copies
Four drafts of the same approach share the same mistakes. Each draft answers under a lens: direct, verify by an independent method, adversary with traps first, or first principles.
Checking, never majority
When drafts disagree, the merger settles each point by working it out. Three wrong drafts do not outvote one right one.
Shared-error hunt
Agreement is not proof. The merger looks for the error every draft might share, and a different model audits the result.
Bounded waiting
Once two drafts land, Tempest waits only max(8 s, half the elapsed time) for stragglers. One slow lane cannot hold an answer hostage.
Refusals swapped
If one lane refuses a prompt the others answer, the refusal is swapped out before you see it.
Tools stay simple
Requests with tools go to a single tool-capable model, so function calls behave the way your agent expects.
Every answer reports back
The final streamed chunk carries standard usage plus a tempest object: the task, the difficulty, the stage the answer came from, whether it was cached, and the billed amount in USD.
// final chunk of a stream (abridged) { "usage": { "prompt_tokens": 412, "completion_tokens": 905, "prompt_tokens_details": { "cached_tokens": 256 }, "completion_tokens_details": { "reasoning_tokens": 712 }, "cost": 0.0287 }, "tempest": { "task": "...", "difficulty": "...", "stage": "...", "cache": "...", "billed_usd": 0.0287 } }
Illustrative values. Token counts are estimated at about four characters per token.
What we will not claim
Tempest 1 has not been evaluated yet. We have no benchmark numbers to show, and we will not borrow any. Our target is #1 on Artificial Analysis; until our own runs are in, that is a target and nothing more.
Fusion costs time. A deep prompt waits for several drafts, a merge and an audit, so Max is slower than any single model in its panel. Flash is there for when speed matters more than depth.
Upstream providers process prompts under their own terms. Tempest inherits their availability: if a lane is down, the router works with the lanes that remain.