Utpal Das
    Free apps
    All Labs

    Labs · Open source · For anyone who uses AI at work

    Token SaverAutomating the model picker.

    Most of what we ask AI to do in a week is routine: tag this, rename that, draft a caption, write the test. These skills teach your AI assistant to pick the model for you — a script, a small model, a mid model or the flagship — and to move up only when a check fails. Pick your platform and the way you work, and get the exact package for it.

    Find your packageWatch 38-second demoGitHub(opens in a new tab)

    Anthropic · OpenAI · Google · DeepSeek · Qwen · Perplexity · Kimi

    Headline savings

    lower AI cost
    68%
    fewer tokens
    55%
    AI platforms
    7
    personas
    13

    Average across 7 platforms × 13 personas. Compared with running every task on each platform's strong everyday model, with the full conversation in context. Even in the pessimistic case, cost drops 51% and tokens 17%. Prices checked 2026-10-26.

    Product demo · 38 seconds

    See the model picker at work

    Four short chapters: the cost problem, smart routing, choosing your platform and workflow, and the modeled savings.

    Open video(opens in a new tab)

    Find your package

    Pick your platform, then the way you work

    Every persona skill is packaged for each platform, with that platform's own models and prices. Install the base skill plus your persona; the installer does both.

    01 · Your AI platform

    02 · The way you work

    03 · Your package

    Vibe coder · Anthropic

    −68%

    AI cost · 65% fewer tokens

    vs. running everything on Claude Opus 5.5. Pessimistic case: −50% cost.

    Your models

    • T0 ToolScript, no model
    • T1 MicroClaude Haiku 5.5
    • T2 StandardClaude Sonnet 5.5
    • T3 StrongClaude Opus 5.5
    • T4 FrontierClaude Fable 5.1

    Busiest tasks it routes

    • Commit message or PR description from a diff
    • UI tweak, copy or style change
    • Summarise a log or stack trace

    Install

    • Claude CodeSwitches models automatically

      Skills go in ~/.claude/skills/, the three worker subagents in ~/.claude/agents/. Each subagent pins its own model and effort, so delegated work really runs on the cheaper model. Check with /skills and /agents.

      ./install.sh anthropic vibe-coder-routing

    • Claude.ai

      Upload a skill zip under Customize → Skills → + → Upload a skill. The skill prints a routing plan and tells you which model to switch to at each step.

    • Claude API

      Load routing.json, look up the task's tier and call that model ID.

    Commands run from the cloned repo: git clone https://github.com/UtpalDasLabs/token-saver-automated-model-picker.git. Haiku 5.5 costs 5x more per token once a prompt passes 100k tokens. Split or compact before that line.

    Compare

    Cost saved, every persona on every platform

    Tap a cell to open that package. Savings differ with model prices, effort controls and context use; Kimi's savings come from lower effort and smaller context rather than a cheaper micro model.

    Swipe sideways to see all 7 platforms →

    PersonaAnthropicOpenAIGoogleDeepSeekQwenPerplexityKimi
    Data analyst
    Ops assistant
    Research analyst
    Sales and BD
    Social creator
    Support and success
    Vibe coder
    Educator and course creator
    Growth marketer
    Job seeker
    Designer
    Product strategy
    Writer and editor
    Average69%61%72%74%78%76%45%

    Show your working

    How the numbers are calculated

    These are modeled estimates, not a lab benchmark. Savings differ by platform because the price gap between small and strong models differs. Kimi has no supported cheaper micro model, so its model uses effort and context savings instead. Every assumption is in the repo — change the workload to match your week and re-run the calculator.

    Read the methodology(opens in a new tab)Task-by-task results(opens in a new tab)Anthropic price list(opens in a new tab)
    1. 01Each persona has a modeled week of real task types — how often, how much context, which tier the skill routes it to.
    2. 02Baseline: every task on the platform's strong everyday model at default effort, with the full conversation in context.
    3. 03Routed: each task on its tier's model, with failed attempts paid in full and redone one tier up, plus the cost of writing briefs and loading the skill.
    4. 04Prices are official API list prices, checked through 2026-10-26. Caching, batch and off-peak discounts are not counted — they stack on top. Kimi Code membership quota is not directly comparable to API spending.
    5. 05Pessimistic case (failure rates doubled, half the context savings): 51% lower cost, 17% fewer tokens. Cost savings hold up because they come from routing; token savings depend on keeping context lean.

    Free · MIT licence

    Fork it, measure it, improve it

    Add a persona or a platform with one YAML file. Pull requests with measured before-and-after usage are the most welcome contribution.

    Open on GitHub(opens in a new tab)