Track AAPI surface built, lanes awaiting keys

Four drafts. One checked answer.

Failure Tempest 1 fuses Claude Fable 5.5, Claude Opus 5.5, GPT-6 Astra 6.2 and Gemini 4 Argon. Easy prompts go straight to the best-fit model. Hard ones get a brief, several lensed drafts, a fusion that checks rather than votes, and an audit by a different model.

Model card Quickstart

Route

Easy prompts stream straight from the best-fit model. Nothing else runs.

Brief

A fast lane writes TASK, DIFFICULTY, GOAL, TRAPS, PLAN and CHECK. Cached for 24 hours.

Draft

Lensed drafts in parallel: direct, verify, adversary, first principles.

Fuse

Agreement ships at once. Disagreement is settled by checking, never by majority.

Audit

A different model audits. Concrete issues trigger a revision.

The panel

Four models, live status.

Each lane comes online when its API key lands. Until the panel is ready the API returns 503 "not live yet".

Lenses

Same question, different angles.

A draft is a model plus a lens. Deep prompts get at least three drafts; if there are fewer free models than lenses, a model is reused under a different lens.

Lens 1

Direct

Answer the task as briefed, the shortest correct way.

Lens 2

Verify

Reach the answer by an independent method, then compare.

Lens 3

Adversary

Start from the traps in the brief and try to break the obvious answer.

Lens 4

First principles

Rebuild the problem from definitions, ignoring the usual shortcut.

The pipeline in detail

1. Route

Not every prompt needs four models. The router sends easy prompts straight to the best-fit model and streams its answer back with no fusion overhead. failure-tempest-1-flash always takes this path; failure-tempest-1-max never does; failure-tempest-1 decides per prompt. See models.

2. Pre-answer brief

For anything harder, a fast lane writes a short brief before any drafting starts: TASK (what is actually being asked), DIFFICULTY, GOAL (what a good answer looks like), TRAPS (where answers usually go wrong), PLAN and CHECK (how to tell the answer is right). The brief can raise the difficulty the router assigned, but never lowers it. Briefs are cached for 24 hours, so a repeated task does not pay for a second one.

3. Lensed drafting

Each drafter receives the prompt, the brief and one lens. Drafts run in parallel. Once two drafts have landed, the pipeline waits only max(8 s, half the elapsed time) for stragglers, so one slow provider cannot hold the answer hostage.

4. Fusion

If every draft states the same final answer, it ships immediately. Otherwise the merger takes each disagreement in turn and settles it by checking the claim against the brief's CHECK, not by counting which side has more drafts. It also looks for shared errors: places where all drafts agree on something that does not survive a check.

5. Audit

A model that did not write the fused answer reads it. If it finds concrete issues — a wrong figure, a broken step, a missed requirement — the answer is revised before it is returned. Vague unease does not trigger a revision; a specific defect does.

Refusals and tools

If any lane refuses, its output is swapped out before the user ever sees it, and the remaining lanes carry the answer. Requests that include tool calls are not fused: they go to a single tool-capable model, because tool calls do not merge cleanly.

What you are billed

You pay for the tokens you send and the tokens you receive. Every brief, draft and audit streams to you as reasoning and bills as reasoning tokens, so nothing runs that you cannot read. Drafts cancelled before they land are not billed. See pricing and caching.

Observability

Watch it work.

When streaming, the working is sent as reasoning_content deltas. The final chunk carries usage plus a tempest object saying which path the request took.

// final stream chunk (abridged)
{
  "usage": {
    "prompt_tokens": 412, "completion_tokens": 988,
    "prompt_tokens_details": { "cached_tokens": 0 },
    "cost": 0.03170
  },
  "tempest": { "task": "…", "difficulty": "…", "stage": "…", "cache": "…", "billed_usd": 0.03170 }
}

Illustrative shape only; token counts are not a measurement. Full fields in the API reference.

Try it when the panel lands.

Invite-only for now — ask for an invite key, then make API keys in the console.

Get API access Status