💸 vibe-budget

Know the cost before you code.
Translate any project description into token estimates and real LLM costs — before you burn your budget.

vibe-budget plan "saas with supabase, oauth, docker, stripe and kubernetes"

The Problem

You open Cursor, describe your project, and start vibing.
Three hours later — $47 gone. On a landing page.

Most developers have no idea how many tokens a project actually costs until after the fact. Different models have wildly different prices. A task that costs $0.01 with Gemini Flash costs $5.80 with Claude Opus.

vibe-budget fixes this.

What It Does

Paste any project description — in plain English or Spanish — and vibe-budget will:

• Detect all the technical tasks involved (auth, database, deployment, UI, infra...)
• Estimate the token count for each task (input + output separately)
• Fetch real-time prices from OpenRouter across 85+ models
• Recommend the best model per task based on quality/price ratio
• Show the cheapest, best value, and premium options side by side

No opinions. No "you should use X". Just the numbers — you decide.

Demo

Project: saas with supabase, oauth, docker, stripe and kubernetes

Detected Tasks (8):
------------------------------------------------------------
  1. PostgreSQL/Prisma    high        75k tokens
  2. Supabase Integration medium      67k tokens
  3. Docker Container     medium      53k tokens
  4. OAuth/OpenID         high        60k tokens
  5. Stripe Payment       high        90k tokens
  6. Kubernetes           high        115k tokens
  7. Authentication       medium      70k tokens
  8. Next.js App          high        120k tokens

Combined Token Estimate:
------------------------------------------------------------
   Input:  265,000 tokens
   Output:  520,000 tokens
   Total:  785,000 tokens

Token Cost Breakdown (OpenRouter Pricing):
------------------------------------------------------------
   PostgreSQL/Prisma      25,000 in +   50,000 out = $0.53
   Supabase Integration   22,000 in +   45,000 out = $0.48
   Docker Container       18,000 in +   35,000 out = $0.37
   OAuth/OpenID           20,000 in +   40,000 out = $0.43
   Stripe Payment         30,000 in +   60,000 out = $0.69
   Kubernetes             40,000 in +   75,000 out = $0.80
   Authentication         25,000 in +   45,000 out = $0.48
   Next.js App            40,000 in +   80,000 out = $0.92

Model Recommendations by Task:
-----------------------------------------------------------------------------------------------
Task                Tokens    Recommendation                               Cost
-----------------------------------------------------------------------------------------------
   PostgreSQL/Prisma    75,000     DeepSeek R1              [BEST] $0.15 (Q: 96%)
   Supabase Integration 67,000     Gemini 2.5 Flash-Lite    [BEST] $0.01 (Q: 68%)
   Docker Container     53,000     Gemini 2.5 Flash-Lite    [BEST] $0.01 (Q: 68%)
   OAuth/OpenID         60,000     DeepSeek R1              [BEST] $0.12 (Q: 96%)
   Stripe Payment       90,000     DeepSeek R1              [BEST] $0.18 (Q: 96%)
   Kubernetes           115,000    DeepSeek R1              [BEST] $0.23 (Q: 96%)
   Authentication       70,000     Gemini 2.5 Flash-Lite    [BEST] $0.01 (Q: 68%)
   Next.js App          120,000    DeepSeek R1              [BEST] $0.24 (Q: 96%)
-----------------------------------------------------------------------------------------------
   TOTAL                785,000                             $0.95

Overall Best Models:
-----------------------------------------------------------------------------------------------
   [BEST]:   DeepSeek R1               Quality: 96.1%   Cost: $1.57
   [REC]:    Claude Sonnet 4.5         Quality: 77.0%   Cost: $9.42
   [CHEAP]:  Gemini 2.5 Flash-Lite     Quality: 68.0%   Cost: $0.12

Top 10 Cheapest Models:
-------------------------------------------------------------------------------------
   Gemini 2.5 Flash-Lite      Google      68.0%     $0.12
   GPT-5 Nano                 OpenAI      85.0%     $0.24
   Gemini 2.5 Flash           Google      71.0%     $0.24
   Qwen 2.5 Coder 1.5B        Qwen        N/A       $0.24
   Gemini 2.0 Flash           Google      70.0%     $0.29
   GPT-4o-mini                OpenAI      87.0%     $0.36
   Gemini 3 Flash Preview     Google      90.8%     $0.57
   Qwen 2.5 Coder 7B          Qwen        88.4%     $0.57
   DeepSeek V3                DeepSeek    82.6%     $0.15
   DeepSeek R1                DeepSeek    96.1%     $1.57

Showing top 10 of 85 models

⚠️  ESTIMATED PRICES & TOKENS
------------------------------------------------------------
Prices shown are estimates based on OpenRouter costs.
Token estimates are based on standard prompting patterns.
Actual usage may vary based on prompt style and complexity.
View updated prices: https://openrouter.ai/models
------------------------------------------------------------

Install

# Clone the repo
git clone https://github.com/gaboexe0/vibe-budget.git
cd vibe-budget

# Install dependencies
npm install

# Run
npm run dev -- plan "your project description"

npm package coming soon — npx vibe-budget will work once published.

Usage

plan — Full project breakdown

vibe-budget plan "e-commerce with react, supabase and stripe"
vibe-budget plan "landing page with nextjs and vercel deployment"
vibe-budget plan "mobile app with firebase, react native and oauth"

Works in English and Spanish:

vibe-budget plan "necesito un saas con oauth, docker y postgresql en supabase"

estimate — Quick single task estimate

vibe-budget estimate "kubernetes cluster"
vibe-budget estimate "stripe payment integration"
vibe-budget estimate "supabase with rls policies"

tokenize — Count tokens in any text

vibe-budget tokenize "your prompt text here"

How It Works

1. Task Detection

vibe-budget parses your description and maps it to known technical task categories using keyword matching across 13 domains:

| Category | Examples | |----------|---------| | Cloud Integrations | Supabase, Firebase, Vercel, Netlify | | Databases | PostgreSQL, MongoDB, MySQL, Redis, Prisma, DynamoDB | | Deployment | Vercel Edge, Netlify Functions, Railway, Render, Fly.io, AWS Lambda, Cloudflare Workers | | Infrastructure | Docker, Kubernetes, Terraform, CI/CD, GitHub Actions | | Auth & Security | OAuth, NextAuth, JWT, SSL/TLS, Secrets | | Frontend | React, Next.js, React Native, Flutter, UI Components | | Backend | REST APIs, GraphQL, gRPC, Microservices | | Payments | Stripe, PayPal, Webhooks | | AI/ML | LLM Integration, RAG, AI Agents, Vector DBs | | Mobile | iOS/Android, Push Notifications | | Testing | Unit, E2E, Integration | | DevOps | Monitoring, Logging, Alerts | | SaaS | Multi-tenancy, Billing, User Management |

2. Token Estimation

Each task has calibrated base estimates for input and output tokens, derived from real-world prompting patterns across 84 LLM models. Models with different efficiency profiles (reasoning models, smaller models, etc.) get adjusted estimates automatically.

Base estimates per task type:
  Supabase Integration:    22K in + 45K out = 67K total
  Firebase Integration:    25K in + 50K out = 75K total
  PostgreSQL Setup:        18K in + 35K out = 53K total
  Kubernetes Cluster:      40K in + 75K out = 115K total
  Stripe Payment:          30K in + 60K out = 90K total
  Vercel Deployment:        8K in + 15K out = 23K total
  SSL/TLS Setup:            7K in + 14K out = 21K total
  Environment Variables:    5K in + 10K out = 15K total
  ... and 15 more

Reasoning models (DeepSeek R1, o3, QWQ) use ~2.3x more output tokens due to chain-of-thought. vibe-budget accounts for this automatically.

3. Real-Time Pricing

Prices are fetched live from the OpenRouter API and cached for 1 hour. This covers 85+ models across all major providers.

4. Quality-Price Optimization

Each recommendation uses a quality/price ratio based on HumanEval, SWE-bench, and LiveCodeBench benchmark scores. The goal is not always the cheapest model — it's the best model for the money, given the task complexity:

• Simple tasks (config files, env vars, deployments) → cheap fast models
• Complex tasks (auth logic, payment flows, infra) → higher quality models

Models Covered

85+ models across all major providers, including:

| Provider | Models | |----------|--------| | OpenAI | GPT-5.2, GPT-5.1, GPT-5 Nano, GPT-4o, GPT-4o-mini, o3, o4-mini | | Anthropic | Claude Opus 4.5/4.6, Sonnet 4.5, Haiku 4.5, Claude 3.5 Sonnet | | Google | Gemini 3 Pro, Gemini 2.5 Pro/Flash/Flash-Lite, Gemini 3 Flash | | DeepSeek | R1, R1-0528, V3, V3.2, V4, Coder V2 | | xAI | Grok 4, Grok 4 Code, Grok 3 variants | | Qwen | Qwen3 Coder 480B, Qwen 2.5 Coder (32B/14B/7B/3B/1.5B), QWQ 32B | | Zhipu AI | GLM-5, GLM-4.7, GLM-4.6 | | MiniMax | M2.5, M2.1, Text-01 | | Moonshot AI | Kimi K2.5 | | 01.AI | Yi-Large, Yi-Medium, Yi-Coder-9B |

Benchmark Data

Quality scores used for recommendations come from three coding benchmarks:

| Benchmark | What It Measures | |-----------|-----------------| | HumanEval | Code generation from docstrings | | SWE-bench Verified | Real GitHub issue resolution | | LiveCodeBench | Competitive programming problems |

Top performers:

DeepSeek R1         HumanEval: 96.1%   SWE-bench: —
GLM-5               HumanEval: 94.2%   SWE-bench: 77.8%
Claude Opus 4.5     HumanEval: 93.8%   SWE-bench: 80.9%
Qwen 2.5 Coder 32B  HumanEval: 92.7%   SWE-bench: 69.6%
Gemini 3 Pro        HumanEval: 92.5%   LiveCode:  91.7%

Philosophy

vibe-budget is a calculator, not an advisor.

It shows you the numbers. It doesn't tell you which model to use. Like Google Flights showing all airlines — you see the prices, you make the call.

What it won't do:

• Connect to your Cursor/Windsurf/Codeium accounts
• Access your API keys or usage data
• Tell you "use this model because it's better"

What it will do:

• Give you the most accurate estimates it can
• Be honest about uncertainty (hence the ⚠️ disclaimer)
• Let you make an informed decision before spending

Disclaimer

⚠️ All token counts and costs are ESTIMATES.
   Actual usage depends on your prompting style,
   project complexity, and model behavior.
   Prices update from OpenRouter hourly.
   Use as a planning tool, not a guarantee.

Roadmap

• [ ] npx vibe-budget (npm publish)
• [ ] --budget $5 flag — filter models by max cost
• [ ] --export csv/json — export plan for client quotes
• [ ] Multi-project planning — optimize across phases
• [ ] Free tier tracker — manually track Cursor/Windsurf trials
• [ ] vibe-budget actual <tokens> — compare estimate vs real usage

Contributing

Found a task category missing? Wrong token estimate? Open an issue or PR.

The core data lives in data/tasks.json and data/integrations.json — adding new tasks is straightforward.

License

MIT — use it, fork it, build on it.