Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Pickwise: Best AI Model for Any Task

    1

    Returns the best AI model for a task: model ID, cost, savings and why, ranked by quality per dollar.

    $9.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Pickwise: Best AI Model for Any Task

    Pickwise: Best AI Model for Any Task

    Example session with this skill installed

    I'm building a support bot for an e-commerce site. It handles about 50,000 queries a month. Which model should I use if I want good quality but need to keep costs down?

    • Read your context and instructions
    • Compiled the pickwise

    Use Claude Haiku 5.5 — anthropic/claude-haiku-5.5 (Anthropic API: claude-haiku-5-5)

    About

    $21 a month for 50,000 queries. Compared with Claude Opus 5.5, the top model for this task, that saves

    $799 a month ($9,585 a year): it is 97% cheaper and keeps 83% of Opus's task score.

    ModelModel IDScoreCost / month
    PickClaude Haiku 5.5anthropic/claude-haiku-5.565$21.25
    Top qualityClaude Opus 5.5anthropic/claude-opus-5.578$820.00
    Fallback (other vendor)DeepSeek V4.1 Flashdeepseek/deepseek-v4.1-flash-101764$27.15

    Haiku 5.5 is also the cheapest model that clears the quality bar.

    Why: Haiku 5.5 clears the quality bar for a routine support task (42) with a score of 65, at 3% of Opus's cost. Moving up to Claude Sonnet 5.5 adds 5.4 useful points for 19× the price (1.3 points per doubling of cost), short of the 5.7 this task needs to justify the spend.

    Assumptions (correct any of them and I'll re-run):

    • Ongoing workload, 50,000 requests a month
    • About 5,000 input and 300 output tokens per request
    • 60% of input served from the prompt cache (already included in the cost)
    • Real-time replies, balanced priority

    Next steps

    1. Call anthropic/claude-haiku-5.5. On errors or an Anthropic outage, fail over to deepseek/deepseek-v4.1-flash-1017.
    2. Send low-confidence or sensitive conversations (refunds, complaints) to anthropic/claude-sonnet-5.5. At 10% of traffic that adds about $41 a month.
    3. Cheaper models such as MiMo V2.6 Flash fall below the quality bar for customer-facing replies.

    Prices are each vendor's standard API rate; quality data from Epoch AI and LMArena (CC BY 4.0), catalog of 2026-10-25. Charts: https://lihaila.com/models/en/

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Pickwise does one thing: your agent describes a task, and Pickwise returns the best model for it, with the model ID to call and an analysis the agent can act on.

    What you get back

    • The chosen model and its model ID (OpenRouter ID, plus the vendor's native API name where known)
    • Why it wins, what it costs per request and per month (or for a one-time job), and how much it saves versus the top model
    • Alternatives: top quality, cheapest that works, and a fallback from a different vendor for outages
    • A model stack for multi-part projects. A short film, for example, gets a video model, a script model, a voice-over model and a music model, with the total cost

    Coverage

    Current text LLMs from Anthropic, OpenAI, Google, xAI, DeepSeek, Alibaba Qwen, Moonshot Kimi, Z.ai GLM, ByteDance Doubao Seed, MiniMax, Xiaomi MiMo, StepFun, Meta and Mistral, plus video, image, voice and music generation models (Wan, Seedance, Veo, GPT Image, Nano Banana, FLUX, ElevenLabs, Suno and more).

    How it decides

    pick = argmax over models of [useful quality − λ · log2(cost)]

    • Quality comes from the Epoch AI Capabilities Index and LMArena ratings (CC BY 4.0), weighted by expert presets for each task type
    • Cost uses each vendor's own standard API price and the real workload, including cached input and hidden reasoning tokens
    • A quality bar set by difficulty: a model that fails the task is never the cheap option
    • λ is how many points each doubling of cost must buy. It rises with the money at stake and falls for hard tasks
    • It tells one-time jobs from ongoing workloads

    For agent platforms

    • --route returns just the routing view: the model to call, a different-vendor fallback for errors and outages, and an escalation model for hard cases, with model IDs and cost
    • --batch builds a whole routing table, one route per kind of work, in a single call
    • Every output carries schemaVersion, so your router can parse it safely
    • Guide with examples: https://lihaila.com/models/en/#agents

    Private

    No API key, no per-use cost. Your task never leaves your machine: the skill only downloads the day's prices and scores from lihaila.com, and works offline without them. Needs Node 18+.

    Tested in Claude Code. Free web version with charts: https://lihaila.com/models/

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Excludes 3D models and embeddings. Uses standard vendor API pricing; does not account for third-party hosting, free tiers, or volume-based enterprise discounts.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 22 days ago

    Software
    Node.js 18+

    Permissions

    Terminal / Shell
    Network Access

    Terminal: runs one local Node script (scripts/pick.mjs). Network: one GET request to https://lihaila.com/models/api/updates for the day's public prices and quality scores (2-second timeout); nothing about the user's task is sent, and the skill works offline (or with --offline) on its bundled catalog.

    Listed23 days ago

    What's inside

    Frequently Asked Questions

    Pickwise solves the problem of overspending on frontier models by identifying which LLMs or media generators provide the best quality for your specific task budget. It eliminates "dominated" models that are both more expensive and less capable than their peers.

    The skill includes a standalone Node.js engine and a comprehensive model catalog covering text (Anthropic, OpenAI, DeepSeek, Google) and media (Veo, Seedance, GPT Image, ElevenLabs, Suno). You receive the full logic for calculating the quality-per-dollar frontier for your projects.

    Pickwise is designed for any environment running Node.js 18+ and provides OpenRouter IDs for easy integration. It is ideal for developers building RAG pipelines, support bots, or automated media workflows who need data-backed model routing.

    A catalog ships inside the skill, and each run first downloads the day's update from lihaila.com: current prices, quality scores and newly added models (2-second limit; offline it uses the bundled catalog). Prices are each vendor's own standard API rate; quality comes from the Epoch AI Capabilities Index and LMArena ratings (CC BY 4.0). The result shows the catalog's date and whether it was updated live.

    The engine uses a specialized quality-doubling metric and real-world arithmetic that standard LLMs cannot perform accurately due to stale training data. It accounts for complex variables like prompt caching, reasoning tokens, and task difficulty that a simple prompt would miss.

    New in LLM Ops