Fix GPT-5.6 fast pricing

[robinebers]

TL;DR

Price GPT-5.6 Sol, Terra, and Luna fast variants at the owner-confirmed 2.5× rate, matching GPT-5.5 fast mode, and stop unknown fast variants from silently falling back to standard-speed pricing.

What was happening

• GPT-5.6 base pricing existed, but its fast slugs had no aliases or model-specific multiplier.
• An unrecognized -fast slug could fuzzy-match its base model and be estimated at 1×, understating spend without warning.
• Fast-mode pricing is model-dependent: OpenAI currently documents GPT-5.5 at 2.5× and GPT-5.4 at 2× in its Codex Speed documentation.

What this changes

• Adds explicit 2.5× fast multipliers for GPT-5.6 Sol, Terra, and Luna.
• Adds fast aliases for each tier and its supported reasoning-effort suffixes.
• Keeps exact fast prices from LiteLLM or models.dev authoritative when present.
• Leaves unknown fast variants unpriced so the existing warning triangle exposes incomplete spend instead of silently using base rates.
• Documents the unknown-fast behavior and adds end-to-end regression coverage for every GPT-5.6 tier and token bucket.

Heads-up

• GPT-5.6's 2.5× multiplier is an owner-confirmed pricing decision; the public GPT-5.6 preview describes fast mode but does not yet publish its multiplier.
• This supplement update publishes to installed apps through GitHub Pages after merge; no app release is required.

Tests

• Supplement JSON and regex validation
• DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer swift test --scratch-path .build-test --filter 'ModelPricing|PricingBundledResource' — 41 tests passed

---

Note

Medium Risk
Changes core model resolution logic and spend totals for any -fast slug lacking a known multiplier; supplement data is owner-confirmed for GPT-5.6.

Overview
Fixes GPT-5.6 Sol, Terra, and Luna fast spend estimates by adding 2.5× fast multipliers (aligned with GPT-5.5) and alias rules so tiered Cursor slugs like gpt-5.6-sol-ultra-fast resolve and scale all token buckets correctly.

ModelPricing lookup no longer prices unrecognized -fast slugs by fuzzy-matching the standard-speed base model (which understated fast usage). Those slugs stay unpriced unless a multiplier, scaled base entry, or exact fast row exists in LiteLLM/models.dev—so incomplete totals surface via the existing warning UI instead of silent 1× rates.

Docs and bundled-resource/unit tests cover the new behavior and GPT-5.6 fast end-to-end.

^{Reviewed by Cursor Bugbot for commit f8e02b5. Bugbot is set up for automated code reviews on this repo. Configure here.}