Math-first comparison of every consumer AI subscription on the planet — ChatGPT, Claude, Gemini, Mistral, DeepSeek, Qwen, GLM, Kimi and more — against the same spend on the raw API. Cold research, zero affiliate links.
🌐 Live → bossincrypto.github.io/best-ai-plan
Most "AI plan comparison" pages list features and prices and call it a day. That tells you what you pay, not what you get.
This index flips the question. For every consumer plan we map the model you actually receive when you hit send, multiply your monthly usage profile by the public list API price of that model, and divide by the subscription price. The result is a single, comparable number: the leverage multiple — how many dollars of raw API the subscription hands you for every dollar you spend.
Headline finding (SemiAnalysis methodology): pushing a single Claude Max 20× account and a single ChatGPT Pro account against long agentic coding tasks until the weekly caps slammed shut implied $8,000/mo of Sonnet API value (Claude) and $14,000/mo of GPT-5 API value (OpenAI). Per seat. We checked the rest too.
- Value leaderboard — every tracked plan ranked by leverage multiple, filterable by usage profile (coder, chat-heavy, research…).
- $5–$25 budget zone — the cheapest credible plans, so you don't overpay for the flagship tier.
- API analytics — $/M-token pricing for input & output across all tracked models, with sort & insights.
- SemiAnalysis case study — the real-world "max the cap" numbers, with caveats.
- Method & caveats — the model mapping, the weekly-cap landmines, and the Chinese-model caveat, stated up front.
- Daily auto-refresh — a GitHub Actions cron re-snapshots every 6 hours; the page revalidates on every visit.
- SEO baked in — prerendered HTML, Open Graph, Twitter cards,
Dataset+FAQPageschema, sitemap & robots.
usage_profile_tokens × list_API_price_$per_token
leverage = ───────────────────────────────────────────────────
subscription_price_$/mo
- Map the model the user actually gets (Plus → GPT-5, Max 20× → Sonnet 4.5 w/ Opus burst…).
- Cost it against the public API meter for a usage profile.
- Divide by the sticker price → the leverage multiple.
- Snapshot daily so historical drift is auditable.
⚠️ Read the caveats (in-app Method section). Subscription throughput ≠ API throughput — subs degrade silently (model swaps, smaller context, weekly caps). The big multiples are ceilings pushed by agents, not typical chat use.
git clone https://github.com/BOSSincrypto/best-ai-plan.git
cd best-ai-plan
bun install
bun run dev # http://localhost:5173Build a static bundle (mirrors what Pages serves):
BASE_PATH=/ bun run build
bunx serve dist/clientUses Bun.
bun.lockb/bun.lockis committed for reproducible CI.
| Layer | Choice |
|---|---|
| Framework | TanStack Start (file-based routing, SSR/prerender) + React 19 |
| Build | Vite 8, static prerender to dist/client |
| Data | Supabase (Postgres) — plans, api_prices, providers, price_snapshots, RLS-locked public read |
| State | TanStack Query (suspense + refetchOnMount: "always") |
| Styling | Tailwind CSS v4 + tw-animate-css, terminal / editorial aesthetic |
| UI primitives | Radix UI, lucide-react, recharts, sonner, cmdk, vaul |
| Forms | react-hook-form + zod |
| Deploy | GitHub Actions → GitHub Pages |
src/
├── routes/
│ ├── __root.tsx # app shell, <Outlet/>, global <head>
│ ├── index.tsx # the Value Index homepage (loader + prerender)
│ ├── sitemap[.]xml.ts # dynamic sitemap
│ └── api/ # server routes (nitro disabled for Pages)
├── components/
│ ├── value-index/ # leaderboard, budget-zone, api-analytics,
│ │ # chrome (Hero/Header/Footer/Methodology), shared
│ └── ui/ # Radix-based design system (shadcn-style)
├── lib/value-index/ # queries, profiles, math, model-mapping
├── integrations/supabase/ # typed Supabase client
└── styles.css
.github/workflows/deploy.yml # build → upload-pages-artifact → deploy-pages
public/og-image.jpg # social share cover
The site is a fully static SPA deployed via GitHub Actions
(.github/workflows/deploy.yml). It rebuilds on
every push to main, every 6 hours (cron), and on manual dispatch. Data stays
fresh three ways:
- At build time — the prerender step bakes real Supabase rows into
index.html(great for SEO + share previews). - At view time — the browser re-fetches on mount, so visitors always see the latest numbers.
- On a schedule — the cron keeps prerendered HTML within ~6h of the database.
Full one-time setup (Pages source + repo Variables) is in
DEPLOY.md. The Pages site publishes at
https://<user>.github.io/<repo>/ and the workflow auto-detects the correct
base path via actions/configure-pages.
Pricing snapshots come from provider pricing pages and the SemiAnalysis methodology. The dataset is released under CC BY 4.0 — credit the AI Subscription Value Index if you reuse it. Source code is provided as-is for the project maintainers.
- Methodology: SemiAnalysis — the "max the weekly cap" teardowns.
- Sources: vendor pricing pages (OpenAI, Anthropic, Google, Mistral, DeepSeek, Moonshot, Zhipu, Alibaba).
- Built with Lovable · TanStack · Supabase.
#AI #LLM #ChatGPT #Claude #Gemini #AIPricing #LLMpricing
#APICost #TokenEconomics #ValueIndex #SemiAnalysis #Supabase
#React19 #TanStack #Vite #TailwindCSS #GitHubPages
cold research · no affiliate links · refreshed daily