Model Price

Compare 650+ LLMs side by side — real pricing, real capabilities.

modelprice.boxtech.icu · built for devs who read configs more than marketing pages

English · 简体中文

---

Why

Every LLM vendor publishes their own pricing table in their own format, with their own units, their own cache semantics, and their own marketing copy. The moment you try to answer "should I use Sonnet 4.5 or DeepSeek V3 for this feature" you end up with six open tabs and a spreadsheet.

Model Price collapses that into one keyboard-first page that:

• Knows about 650+ models across Anthropic, OpenAI, Google, xAI, Meta, DeepSeek, Moonshot AI, Alibaba, Z.AI, MiniMax, Mistral, Cohere and 40+ other labs
• Normalizes everything to per 1M tokens so side-by-side numbers are actually comparable
• Sources data from the community-maintained LiteLLM registry + direct provider scrapes — so new models land within days of release without manual curation
• Deduplicates across providers: "Claude Sonnet 4.5" is one entity with multiple offerings (Anthropic / Bedrock / OpenRouter), not three confusing rows
• Is shareable: every model has /m/:slug, every comparison has /compare/:ids, both are proper URLs you can paste in a tweet

Screens

Home — browse and filter 650+ models

Drawer — zero-click-away detail with same-tier alternatives

Compare — up to 4 models side by side, shared capabilities highlighted

⌘K command palette — client-side fuzzy search, zero network

Light theme — warm off-white, not institutional gray

Highlights

• Keyboard first — ⌘K opens a fuzzy search palette, ↑↓ to navigate, Enter to open, ⌘C to copy the model_id, ⌘D to add to compare. Runs entirely client-side.
• Cold-start resilient — every visit renders instantly from a bundled v2-fallback.json snapshot (≈90 KB gzipped); the live backend refreshes in the background and silently swaps in once ready. No blank page, ever — even if the Render free-tier backend is cold-starting.
• "Same tier, cheaper" recommendations — every detail view surfaces 3 alternatives ranked by capability_overlap × savings (Jaccard over capability sets, weighted by input-price delta). The algorithm lives in backend/services/alternatives.py and is mirrored in the build-time fallback generator.
• Drift report — every refresh emits a drift.json listing unmatched provider models, price deltas > 5%, new/removed entities. Self-healing data quality.
• Official source links — every detail page links out to the maker's official pricing and docs. We provide the index, vendors provide the truth.
• Dark / light / system theme, EN / 中文 localization, both persist to localStorage.
• Open Graph + Twitter Card meta tags with a custom OG image — shareable on X, Slack, WeChat (copy link), Discord, Feishu with a real preview card.

How it works

         ┌────────────────────────────────────────────────┐
         │  LiteLLM registry                              │
         │  github.com/BerriAI/litellm  (single SoT)      │
         └────────────────┬───────────────────────────────┘
                          │  fetched raw JSON on every refresh
                          ▼
         ┌────────────────────────────────────────────────┐
         │  Canonical Entity Layer                        │
         │  slug, family, maker, context, capabilities,   │
         │  modalities, primary_offering_provider         │
         └────────────────┬───────────────────────────────┘
                          │  provider aliases via reverse index
                          ▼
  ┌──────────┬──────────┬──────────┬──────────┬──────────┬──────────┐
  │ Anthropic│ Bedrock  │  Azure   │  OpenAI  │OpenRouter│ xAI/etc. │
  │  fetch() │  fetch() │  fetch() │  fetch() │  fetch() │  fetch() │
  └────┬─────┴────┬─────┴────┬─────┴────┬─────┴────┬─────┴────┬─────┘
       └──────────┴──────────┴────┬─────┴──────────┴──────────┘
                                  ▼
                  ┌──────────────────────────────────┐
                  │  Offering Merger                 │
                  │  Each provider record resolves   │
                  │  to a canonical_id via strict    │
                  │  slug+prefix+version stripping   │
                  └──────────────┬───────────────────┘
                                 ▼
                  ┌──────────────────────────────────┐
                  │  entities.json + offerings.json  │
                  │  + drift.json (data quality)     │
                  └──────────────┬───────────────────┘
                                 ▼
                  ┌──────────────────────────────────┐
                  │  /api/v2/* endpoints             │
                  │  + v2-fallback.json snapshot     │
                  │    shipped inside the Vite bundle│
                  └──────────────────────────────────┘

Two-layer data model: Entity (logical model like claude-sonnet-4-5) + Offering (one specific (entity, provider) pair with its own price and last-updated timestamp). One canonical claude-sonnet-4-5 entity with three offerings (Anthropic, Bedrock, OpenRouter) — not three separate rows with the same name.

Stack

Layer	Tech
Backend	Python 3.12, FastAPI, Pydantic v2, httpx, Playwright (for provider scrapes), Pillow (OG image generation), pytest
Frontend	React 19, TypeScript 5.9, Vite 7, React Router 7, vitest + @testing-library/react + happy-dom
Data	LiteLLM model_prices_and_context_window.json + direct provider APIs / scrapes, merged into entity/offering JSON files, shipped as a static snapshot alongside the SPA
Hosting	Vercel (frontend, auto-deploy on release push) · Render free tier (backend, Docker) · GitHub Actions keepalive cron

Local development

Backend

cd backend
uv sync
uv run playwright install chromium   # first time only

uv run main.py                        # http://localhost:8000
uv run pytest                         # 86 backend tests

Frontend

cd frontend
npm install

npm run dev                           # http://localhost:5173
npm test                              # 32 frontend tests
npm run build                         # prod bundle into dist/

The frontend proxies /api/* → http://localhost:8000 in dev. Use VITE_PUBLIC_BASE_URL to override the canonical public origin used by Copy Link / Share on X (defaults to https://modelprice.boxtech.icu).

Refresh the data

curl -X POST http://localhost:8000/api/v2/refresh

This re-fetches LiteLLM, re-runs all provider scrapes, writes fresh entities.json / offerings.json / drift.json. Check GET /api/v2/drift for unmatched models and price-drift items.

Sanity check

cd backend
uv run --active python scripts/sanity_check.py

Prints a hit-rate matrix of ~80 popular first-party models against the v2 entity set. Currently at 79% — the misses are long-tail edge cases logged in drift.json.

Regenerate the OG cover

cd backend
uv run --active python scripts/generate_og_cover.py

Writes frontend/public/og-cover.png (1200×630). Regenerate after copy or palette changes.

Take README screenshots

cd backend
uv run --active python scripts/take_screenshots.py

Uses the Playwright Chromium to capture the production site into docs/screenshots/ — home / drawer / entity page / compare / ⌘K palette in both dark and light themes.

Testing

118 tests total, all hermetic, ~3 seconds to run the whole suite:

• Backend (pytest, 86) — slugify / strip_version_suffix, family-detection ordering (guards the "Codex before GPT" and "Cogito before Llama" regressions), canonical resolver cascade, offering merger helpers, compute_alternatives ranking math, /api/v2/* contract tests via FastAPI TestClient.
• Frontend (vitest + @testing-library/react + happy-dom, 32) — formatPrice / formatContext / formatPct, CompareBasketProvider (capacity cap, sessionStorage persistence, provider-scoping error), ThemeProvider (dark/light/system cycle, matchMedia listener, <html data-theme> reflection), LocaleProvider (default EN, {name} interpolation, <html lang> sync).

Run them:

cd backend && uv run pytest
cd frontend && npm test

Deployment

The release branch is the production branch.

• Vercel auto-builds and deploys the frontend on every release push. Config lives in frontend/vercel.json: it proxies /api/* → Render and has a catch-all rewrite to index.html so React Router deep links work.
• Render rebuilds the backend Docker image on every release push. Config in backend/render.yaml. Free plan sleeps after 15 min of inactivity — the SWR fallback snapshot in the frontend makes this invisible to users.
• GitHub Actions keepalive in .github/workflows/keepalive.yml pings /api/health every 10 min as a second line of defense against sleep.

Previous v1 production is tagged release-v1-backup for rollback. The v1 routes (/api/models, /api/providers, /api/families, /api/stats, /api/refresh/metadata) were removed on 2026-04-15. POST /api/refresh remains as a compatibility alias that delegates to the v2 refresh pipeline.

Design notes & philosophy

Read the v2 rebuild design document at docs/plans/2026-04-14-v2-redesign-design.md — it documents the product positioning (D quick lookup + A comparison as foundation, C "same-tier cheaper" as killer differentiator), the target audience (independent developers + agent/toolchain devs), the single-point breakthrough (Cmd+K + shareable URLs), the two-layer data model, the cold-start strategy, and every non-goal we intentionally didn't ship.

The v2 API contract is frozen at docs/plans/v2-api-contract.md.

License

MIT — see LICENSE.

Data sourced from BerriAI/litellm (MIT). Thanks to the LiteLLM community for maintaining the registry that makes this site possible.