Prices checked 2026-11-17 · 88 models

Which AI model should you use?

Describe your task in plain words. We read it, weigh quality, speed and price the way an expert would, and pick the model that gives you the most for your money.

ClaudeGPTGeminiGrokDeepSeekQwenKimiGLMDoubaoMiniMaxMiMoStepMuseMistral
Video · image · voice · musicWanSeedanceKlingHailuoVeoNano BananaFLUXElevenLabsSuno

How we pick

No sponsored placements, no vibes. The same transparent method runs for every task.

  1. Read the task

    A small AI model classifies your description: task type, difficulty, language, whether someone waits on each answer, and any hard requirements.

  2. Weigh what matters

    Each task type has expert weights over eight capabilities: reasoning, coding, agentic tool use, writing, vision, Chinese, other languages and speed. A support bot weighs speed; a research task weighs reasoning; a Japanese translation weighs language quality.

  3. Score every model

    Capability scores combine the Epoch Capabilities Index (benchmarks) with LMArena human-preference ratings for coding, writing, Chinese and vision, plus speed; where a model has no rating, known vendor strengths adjust it lightly.

  4. Price the real workload

    Your tokens × list prices, with cached input discounts and the hidden reasoning tokens that hard tasks burn.

  5. Set a quality bar

    Below the bar a model fails the task and only looks cheap, so it is heavily penalized. Above the bar, extra quality counts less on easy tasks.

  6. Make price earn its place

    Dominated models drop out. A pricier model wins only if each doubling of cost buys enough extra quality. That threshold rises with volume and falls for hard tasks.

The formula

Every pick on this page comes from this one expression. Dominated models (worse and pricier than another) are removed first.

task score of model m (0–100)
model m's score on capability d (reasoning, coding, agentic, writing, vision, Chinese, other languages, speed)
how much the task needs capability d (expert preset per task type, sums to 1)
quality bar for the task's difficulty: 25 / 42 / 60 / 76 / 88
value of quality above the bar: 0.3 for trivial tasks up to 1.0 for frontier ones
cost of the workload per month, or for the whole one-time job
requests (or seconds, images, characters, tracks) in the period; t = tokens, p = vendor prices, c = cached share
hidden reasoning tokens per visible output token: 0 to 3.5 by difficulty
takes per kept result for generative media
points a doubling of cost must buy: 1.5 quality-first, 4.5 balanced, 9 cost-first
difficulty discount: 0.75 for hard tasks, 0.5 for frontier ones
monthly spend on the strongest model (a one-time job counts as 1/12 per month)

For AI agents and agent platforms

The Pickwise skill gives your agent the same answer as this page, as structured JSON it can route on.

  • Route on the result

    Each answer names the model to call (primary), a different-vendor fallback for errors and outages, and an escalation model for hard cases, with OpenRouter and vendor model IDs and the cost.

  • A whole routing table in one call

    List the kinds of work your agents do (tag emails, draft replies, plan, research) and get one route per kind in a single batch call. Store it and route each request by its type.

  • Fresh every day, stable to parse

    Each run picks up that day's prices, scores and new models from lihaila.com, and works offline on the bundled catalog. Every output carries a schemaVersion, so your parser knows what to expect.

$ node pick.mjs --route --analysis \
    '{"category":"support","requests":90000,"realtime":true}'
{
  "schemaVersion": "1.0",
  "primary":  { "modelId": "anthropic/claude-haiku-5.5",
                "cost": 38.25, "period": "month" },
  "fallback": { "modelId": "meta/muse-ember" },
  "escalate": { "modelId": "anthropic/claude-fable-5.2" },
  "catalogUpdated": "2026-11-17"
}
Get the skill on Agensi ($9.99)

Frequently asked questions

Where do the numbers come from?

Prices are each vendor's official list prices for standard API calls, checked automatically every day. Quality comes from the Epoch AI Capabilities Index (epoch.ai/benchmarks) and the LMArena leaderboard dataset (huggingface.co/datasets/lmarena-ai/leaderboard-dataset), both CC BY 4.0 and refreshed daily; we rescale and combine them. Speed, voice and music quality, and anything else we had to estimate are marked with *.

Why not always the smartest model?

Because for most tasks it is overkill. A ticket tagger that is right 99% of the time gains almost nothing from a model that costs 40× more. We pay for quality only where the task can use it.

Do you store my task description?

No. It is sent once to the AI reader to classify it and is not saved; we log counts and categories, not text. If you click Share, we save the result and the one-line summary the AI wrote (never your original words) so the link works.

Can I use this from my AI agent?

Yes: Pickwise is also an agent skill for Claude Code, Codex, Cursor and other agents ($9.99 on Agensi: agensi.io/skills/pickwise-best-ai-model-for-any-task). It returns the chosen model, a different-vendor fallback and an escalation model as structured JSON, for one task or a whole routing table in one call, with the same engine and that day's data.

Prices checked 2026-11-17.