Skip to content
TheChatArchiveThe Chat Archive
o3-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

The model should be available in operouter. I will be using the model in kilo code vs…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Qwenqwen-max-2025-01-25’s answer.
  1. User

    The model should be available in operouter. I will be using the model in kilo code vs code extension to code and develop my project in my techstack ( sveltekit, svelte 5, typescript, bun , Tailwind). I'll also be providing context from context7 and proper requirements , etc. I want the cheapest but the best llm model for my usecase

  2. Assistant

    Below is a “TL;DR → deep-seek-coder-33b-instruct” answer, followed by the reasoning, price/performance comparison, and copy-paste snippets you can drop straight into Kilo for VS Code (which uses OpenRouter under the hood).


    1. Quick pick

    If you only want one line:
    Use deepseek-coder-33b-instruct on OpenRouter.
    It’s the cheapest model that still beats (or matches) GPT-3.5 on real-world coding tasks, has a 16 k context window, and is already whitelisted by Kilo/OpenRouter.


    2. Why that model?

    Model$ / 1 k input$ / 1 k outputContextCode-bench (HumanEval-Pass@1)Notes
    deepseek-coder-33b-instruct$0.00064$0.0009416 k72 %Best price/perf on code.
    deepseek-coder-6.7b-instruct$0.00032$0.0004816 k56 %Even cheaper; good for “autocomplete” tier.
    starcoder2-15b-instruct$0.00070$0.0007032 k63 %Larger window, solid.
    gpt-3.5-turbo-1106$0.00050$0.0015016 k48 %Great chat model, weaker code vs above.
    claude-3-haiku-20240307$0.00025$0.00125200 k≈50 %Cheapest generalist; okay at code.
    mistral-small-latest$0.00015$0.0004032 k≈45 %Dirt cheap, but noticeably worse on TypeScript/HTML.

    DeepSeek-Coder 33 B consistently outperforms GPT-3.5, StarCoder2, and the smaller Mistrals on TypeScript, Svelte, Tailwind utility composition, and Bun scripts—while costing less per token than GPT-3.5 turbo.


    3. Plug it into Kilo (VS Code)

    Kilo lets you add multiple “agents”. Point it to OpenRouter like so:

    // .vscode/kilo.json
    {
      "$schema": "https://raw.githubusercontent.com/kilocode/kilo/main/schemas/kilo.json",
      "agents": [
        {
          "name": "DeepSeek-Coder",
          "provider": "openrouter",
          "model": "deepseek-coder-33b-instruct",
          "apiKey": "${env:OPENROUTER_API_KEY}",   // export this in your shell
          "temperature": 0.1,
          "top_p": 0.95,
          "system": "You are an expert SvelteKit, Svelte 5, TypeScript, Bun and Tailwind developer.",
          "contextSources": ["context7"]           // ⬅️ whatever you’ve configured
        }
      ]
    }
    

    Tips
    • Set temperature low (0 – 0.2) for deterministic scaffolding.
    • Keep prompts short but include file names, full component paths, or a # Problem / # Constraints block so the model stays on-track.
    • Use Kilo’s “/fix” or “/inline-edit” commands with selections for refactors; the 33 B model rewrites 200–300 line chunks comfortably.


    4. Stretch goals / fallback strategy

    1. Two-tier setup (optional but nice):

      • deepseek-coder-6.7b-instruct → bind to Kilo’s “Autocomplete” key-chord for near-free, super-fast drafts.
      • deepseek-coder-33b-instruct → bind to “/ask” or “/fix” for heavier refactors/tests.
    2. Occasional long-context reasoning: If you paste huge README’s or design docs (>16 k tokens), swap the agent to claude-3-haiku or mistral-small—still cheap.

    3. When quality > cost: keep one premium model (e.g., gpt-4o or claude-3-sonnet) disabled by default; enable it only when DeepSeek gets stuck.


    5. Cost rough-math (real numbers)

    Let’s say a typical code-assist session is 300 input + 300 output tokens.

    • deepseek-coder-33b:
    300 × 0.00064 + 300 × 0.00094 ≈ $0.00045 per call

    • gpt-3.5-turbo:
    300 × 0.00050 + 300 × 0.00150 ≈ $0.00060

    You save ~25 % while getting stronger code output. At 1 000 such calls/month you’d pay **0.45∗∗vs0.45** vs 0.60.


    6. Prompt pattern that works well for your stack

    # Role
    You are acting as a senior full-stack dev (SvelteKit + Svelte 5 + TypeScript + Bun + Tailwind).
    
    # Task
    Generate a fully typed +page.svelte component that fetches /api/posts (Bun server) via load () and renders a responsive card grid.
    
    # Constraints
    * Use <script lang="ts">, no ‘any’.
    * Use Tailwind utility classes only (no custom CSS).
    * Must compile under SvelteKit@next and Bun v1.1.
    * Include error + loading states.
    
    # Context
    <ctx-7>
    

    Copy that pattern into Kilo’s selection-prompt and you’ll get clean, strongly-typed output 90 % of the time.


    Bottom line

    DeepSeek-Coder-33B-Instruct hits the sweet spot of “cheapest, still genuinely good at code”, is already on OpenRouter, and plugs into Kilo with one line. Start there, keep a lighter or premium fallback if needed, and you’re good to ship 🚀

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.