# Which AI model?

- Link: https://vibecad.nl/en/ai-models/
- Published: 2026-10-22T18:45:24+02:00
- Author: Marijn van den Heuvel \+ AI

_This page is also available in Dutch: [Welk AI-model?](https://vibecad.nl/ai-models-for-vibecadding/)_

Which AI model do you use for what? Not the leaderboard, but what works in practice:
reliability on agentic work, long context, tool and MCP steering, speed and cost.

**Read this first:** the classic benchmarks are saturated. On SWE-bench Verified
the top models now sit within a few points of each other (the independent Vals AI
list was archived in September 2026 because it no longer separated the leaders).
Chasing that last point is pointless, _choose based on your workflow_ and test on
your own project.

## The three tiers

 * **Frontier (closed):** the hard 20%, with complex refactors, autonomous agents
   and difficult debugging. Claude Opus (Anthropic), GPT-5 generation (OpenAI), 
   Gemini 3 generation (Google).
 * **Cheap & good:** everyday work for a fraction of the price. Claude Sonnet, Gemini
   Flash, DeepSeek V-series Flash, MiniMax M-series, Grok.
 * **Open-weight:** self-hosting, no vendor lock-in. Qwen3 Coder, Kimi (K2/K3), 
   GLM-5, DeepSeek V-series, Llama.

A pattern that is common in 2026: most of your agent traffic on an open-weight or
cheap model, and only the hardest tasks escalated to a frontier API.

## Model families at a glance

| Family | Provider, weights and access | 
| Claude Opus | Anthropic, closed. Best agentic coding, long refactors, hard reasoning. Via API, Claude apps, Cursor, Cline. | 
| Claude Sonnet / Haiku | Anthropic, closed. Everyday production work and many small tasks. Via API, Claude apps, IDE agents. | 
| GPT-5 generation | OpenAI, closed. A balanced all-rounder with a mature tool ecosystem. Via API, ChatGPT, Copilot, Cursor. | 
| Gemini 3 Pro / Flash | Google DeepMind, closed. Very large context (whole codebases), multimodal and fast. Via API, AI Studio, Vertex. | 
| DeepSeek V-series | DeepSeek, open weights (MIT). Near-frontier coding at low cost. Via API, OpenRouter or locally on heavy hardware. | 
| Qwen3 Coder | Alibaba / Qwen, open weights (Apache-2.0). The default open-weight coding model, in a wide range of sizes. Locally (Ollama, LM Studio) or via API. | 
| Kimi (K2 / K3) | Moonshot AI, open weights (adapted MIT). Agentic coding and long-context question answering. Via API, OpenRouter or self-hosted. | 
| GLM-5 | Zhipu AI, open weights. GPT-class quality in the open. Via API or self-hosted. | 
| MiniMax M-series | MiniMax, open weights. Cost-conscious agent loops. Via API or self-hosted. | 
| Llama | Meta, open weights. Self-hosting, very large context and a broad ecosystem. Locally or via API. |

_Exact version numbers change monthly, so treat the families above as the stable
layer and check the current releases before you standardise._

## Short advice

 * **Hardest tasks and highest reliability:** a frontier model from the Claude Opus
   class.
 * **Best price/quality from a frontier lab:** the Gemini Flash class.
 * **Best open-weight (self-hosted):** Qwen3 Coder for flexibility; Kimi or GLM-
   5 for quality close to frontier.
 * **Cheapest workable agent loops:** the MiniMax M-series and DeepSeek Flash-class
   APIs.
 * **Huge codebases in one go:** the Gemini Pro class (million-token context).

## What matters more than the leaderboard

 * **The harness.** Cursor, Cline, Aider, Claude Code and your own scripts get very
   different results out of the same model. The agent loop often weighs more than
   the model.
 * **Tool and MCP steering.** For vibecadding, whether a model can reliably steer
   your CAD through MCP (sketches, constraints, features) matters more than a few
   benchmark points.
 * **Cost per task**, not per token: a cheap model that needs five attempts can 
   end up more expensive than a frontier model that gets it right in one.
 * **Context window** and how well the model uses it.
 * **Local versus API:** self-hosted open weights mean no cost per seat and full
   data sovereignty, but only models under roughly 120B fit on hardware you own 
   yourself, and the trillion-parameter open models are “open” through an API, not
   on your desk.

## Check the current figures

Because scores shift weekly, you are better off verifying yourself before you standardise:

 * [swebench.com](https://www.swebench.com/), the official SWE-bench leaderboards
 * [Vals AI](https://vals.ai/benchmarks/swebench), one standardised harness (archived
   September 2026, saturated)
 * [Aider Polyglot leaderboard](https://aider.chat/docs/leaderboards/), code editing
   in multiple languages
 * [LMArena](https://lmarena.ai/), human preference (Elo ranking)

_Last updated: November 2026._