What shipped

Failure Model Labs launches at failure.fail.

Failure Tempest 1 API surface

• Pre-answer brief. A fast lane writes TASK, DIFFICULTY, GOAL, TRAPS, PLAN and CHECK before drafting, and can raise the difficulty. Briefs are cached for 24 hours.
• Lensed drafting. Drafts are written under lenses — direct, verify by an independent method, adversary/traps first, first principles. Deep prompts get at least three drafts.
• Response and prefix caching. Exact repeats at temperature 0 (or with x-tempest-cache: on) are served from a one-hour cache at cached rates. Earlier conversation turns resent within an hour bill at the cached-input rate.
• Per-key spend caps. Past the cap, requests return 402 insufficient_quota until the 1st (UTC).
• Visible drafting. Every brief, draft and audit streams as reasoning_content and bills as reasoning_tokens; late drafts are cancelled and not billed.
• Login with Failure ID. OAuth 2.0 + PKCE and OpenID Connect provider, a zero-dependency SDK at /sdk/failure-id.js, and OAuth apps in the console.
• Face check. A second factor with no typing: one button, follow the arrows. Only a face template is stored.
• Invite-only sign-up. Accounts need an invite key issued by the lab.
• Usage endpoint. GET /v1/usage returns the key's month spend, cap, remaining and per-model breakdown.

Failure ID

• Accounts with email and password, TOTP two-factor with recovery codes, and passkeys.
• Console for API keys, usage and spend caps.

• Route. Easy prompts stream straight from the best-fit model.
• Fuse. Disagreements between drafts are settled by checking, never by majority.
• Audit. A different model audits the fused answer; concrete issues trigger a revision.
• Refusal swap. A refusal from any lane is swapped before the user sees it.