--- title: "Models" description: "The language models available in meinGPT, their strengths, and recommendations for usage" canonical_url: "https://meingpt.com/en/docs/platform/models" language: en --- # Models meinGPT brings together leading language models (LLMs) from multiple providers under a single interface. For each chat, you choose the model that fits the task — based on the work to be done, the speed you need, and your data protection requirements. ## Indicators in the model picker The model picker shows several indicators to help you decide: - **EU flag** — the model is hosted in the EU and is GDPR-compliant without further measures. - **US flag / globe** — the model runs outside the EU. - **Quality and speed indicators** — visible on hover, each rated 1–5. - **Reasoning tag** — marks models that use an explicit thinking step before answering. - **Preview tag** — the model is still in a preview phase (functionality and behavior may change). - **Free indicator** — the model is not billed and is always available (currently GPT-5 Nano). - **Cost tier (`$` to `$$$$`)** — hovering shows roughly how expensive a model is compared to others (`$` = cheap to `$$$$` = expensive). - **Best model** — selects GPT-5 first. Only if the conversation, including any attached documents, exceeds GPT-5's context window does it automatically switch to Gemini 2.5 Flash instead. The choice is based solely on how much input there is, not on task type or the reasoning depth required — for demanding tasks, you should still select a specific model manually. By default, only EU-hosted models are used this way. **Data protection:** EU-hosted models process all data within the EU. For assistants, the model choice defaults to **Automatic** — an explicit selection is optional. Left on Automatic, it follows the same logic as the "Best model" behavior in regular chat, and an EU-hosted model is used. **Accuracy for factual lookups:** For structured factual questions (e.g. specific transaction codes, part numbers, or figures), a "bigger" or "better" model alone doesn't guarantee correctness — any language model can state a wrong answer just as confidently as a correct one, and may initially stand by it when challenged. Don't rely on **Best model** for these — pick a model deliberately, enable [web search](/en/docs/platform/web-search) to cross-check if needed, and verify the result against a primary source. **Exact prices:** You only see the exact price per million tokens in this tooltip if you're a workspace admin. As a regular member, the tooltip shows only the cost tier (`$` to `$$$$`). The full price table can be found under [All models](/en/docs/admin-guide/pricing). ## Recommendations for everyday use | Use case | Recommendation | | -------------------------------------- | ------------------------------------------------- | | Standard chat (default) | GPT-5.1, GPT-5.5, Claude Sonnet 4.6 | | Deep reasoning, strategy, audits | GPT-5 Thinking, o3, Gemini 2.5 Pro | | Coding and refactoring | Claude Sonnet 4.6, Claude Opus 4.8, GPT-5.4 | | Very long documents / 1M context | GPT-5.4, GPT-5.5, Claude Opus 4.8, Gemini 2.5 Pro | | High volume, high throughput, low cost | GPT-5 mini, GPT-5 Nano (free), Gemini 2.5 Flash | | Web research with source citations | Web Search (Linkup) | | Default for assistants with tool calls | o3, GPT-5 Thinking | When in doubt, start with a regular (non-reasoning) model. Only switch to a reasoning model if the answers aren't good enough — reasoning consumes noticeably more credits. ## Reasoning models vs. standard LLMs Reasoning models perform a visible thinking step before producing the actual answer. This delivers significantly better results on multi-step or logic-heavy tasks — at the cost of speed and credits. | Task type | Standard LLM | Reasoning model | | ------------------------------------------- | ------------ | ------------------ | | Simple text, emails, summaries | Yes | No (too expensive) | | Creative writing | Yes | No | | Complex logic, mathematics, formal analysis | No | Yes | | Multi-step planning | Sometimes | Yes | | Coding (simple) | Yes | No | | Coding (complex / architectural) | Sometimes | Yes | | Compliance analysis, audits | No | Yes | Many models (the GPT-5 family, Claude, and Gemini 3.5 Flash) let you choose the **reasoning effort** — minimal, low, medium, or high — directly in the chat via the model selector. Higher effort means more thorough answers but takes longer and uses more credits. This lets you dial reasoning up or down per message without switching models. Dedicated reasoning models like GPT-5 Thinking and o3 remain available as separate, always-reasoning choices. ## Context window and file uploads The **Context** column in the model tables below shows the context window in tokens - the total amount of text a model can consider at once in a request (chat history **plus** all attached files combined). As a rough rule of thumb, one A4 text page is about 500-750 tokens; a model with a 200k context window handles roughly 250-400 pages of plain text, one with 1M context correspondingly more. **The context window isn't the only limit.** In addition to the token limit, AI providers set their own independent limits for PDFs and images: a maximum page count per PDF, a maximum file size or resolution per image, and a maximum number of images per request. These provider limits are **not directly** tied to context window size - a model with a very large context (e.g. 1M tokens) can still reject a PDF with very many pages, or heavily illustrated documents (e.g. patent filings with many drawings), once the provider's page or image limit is reached. The exact values are set by each AI provider and can change; they appear in the error message when relevant. **Image input is a per-model file type, not a separate "vision" tier.** Whether a model accepts an attached image works the same way as PDF support: some models can, some can't, and it isn't broken out as its own capability in the tables below. Practically the whole current flagship lineup accepts images (JPEG, PNG, GIF, WEBP) as file attachments: every GPT-4o and GPT-5-family model, every Claude Sonnet/Opus/Haiku model, every Gemini model, and Mistral's Medium and Magistral models. DeepSeek, GPT-OSS, CodeStral, and the older Kimi K2 models do not accept images - check with your workspace admin if you're unsure about a specific model. Over the API, this only applies to the [Assistants Run endpoint](/en/docs/api/assistants) (multipart/form-data file upload) - the [chat completions endpoint](/en/docs/api/completions) never accepts image content in the request body, regardless of which model you select. If you hit one of these limits, switching to a different model (possibly from a different provider) usually helps - see [Chat error messages](/en/docs/platform/chat-errors) for details on the individual error messages. Models differ considerably in consumption: more capable models (currently including o3-pro, GPT-5.4, GPT-5.4 Pro, and the Claude Opus models) cost several times more per request than cheaper ones. The difference sits entirely in the consumption rate — there are no special classes. The exception is models explicitly marked as **free**, whose use consumes no credits. ## OpenAI OpenAI models mostly run via Azure OpenAI Service — for EU hosting, exclusively in EU data centers (Sweden Central). Individual US variants go directly via OpenAI. Data is not used for training. | Model | Hosting | Reasoning | Context | Good for | | -------------- | ------- | --------- | ------- | -------------------------------------------------------------------------- | | GPT-4o Mini | EU | — | 128k | Classification, tagging, simple summaries, FAQ bots | | GPT-4o | EU | — | 128k | General assistant, coding, multimodal, multilingual | | GPT-4.1 | EU | — | 1M | Knowledge work, consistent output, tool use, long documents | | o4-mini | EU | ✓ | 200k | Efficient reasoning tasks, many requests with a logic component | | o3 | EU | ✓ | 200k | Structured reasoning, multi-step workflows, tool calls in assistants | | o3-pro | US | ✓ | 200k | Highly demanding logic, formal analysis, compliance, audits | | GPT-5 | EU | — | 400k | High-quality text, code, agents, long contexts | | GPT-5 Thinking | EU | ✓ | 400k | Difficult logic, planning, mathematics, deep-dive analysis | | GPT-5 mini | EU | — | 400k | Many requests, simple tasks, classification, pipelines | | GPT-5 Nano | EU | — | 400k | **Free** for basic conversations and highly cost-sensitive bulk processing | | GPT-5.1 | EU | — | 400k | Standard model for business and development, agents, planning | | GPT-5.2 | EU | ✓ | 400k | Workflows that need reasoning (EU) | | GPT-5.2 (US) | US | ✓ Coding | 400k | Same class as GPT-5.2, US region | | GPT-5.3 Chat | US | ✓ Coding | 128k | Chat-oriented variant with reasoning and coding focus | | GPT-5.4 | EU | ✓ Coding | 1M | Very long contexts, heavy reasoning and coding tasks (EU) | | GPT-5.4 Pro | US | ✓ Coding | 1M | Maximum reasoning + 1M context for the most demanding workloads | | GPT-5.5 | EU | ✓ Coding | 1M | Previous EU flagship: reasoning, coding, long contexts | | GPT-5.6 Sol | EU | ✓ Coding | 1M | New EU flagship: the most demanding reasoning, coding, and agentic work | | GPT-5.6 Terra | EU | ✓ Coding | 1M | Balanced everyday model in the GPT-5.5 class at lower cost | | GPT-5.6 Luna | EU | ✓ Coding | 1M | Fast and affordable for high throughput, still strong reasoning | ## Anthropic (Claude) Claude models run via Google Vertex. Anthropic pioneered the **Artifacts** concept — accordingly, Claude models produce especially clean artifact output (interactive documents, diagrams, code snippets). | Model | Hosting | Reasoning | Context | Good for | | ----------------- | ------- | --------- | ------- | ------------------------------------------------------------------------------------------------------ | | Claude Haiku 4.5 | EU | — | 200k | Fast, cheap Anthropic responses for simple tasks | | Claude Sonnet 4 | EU | — | 200k | Advanced coding, good balance of intelligence and speed | | Claude Sonnet 4.5 | EU | — | 200k | Production coding, complex agents, multi-tool workflows | | Claude Sonnet 4.6 | EU | — | 200k | Efficient Sonnet all-rounder with high speed | | Claude Sonnet 5 | EU | — | 1M | Current Sonnet with 1M context, balanced speed and quality for coding, analysis, and agentic workflows | | Claude Opus 4.1 | US | — | 200k | Agentic tasks, large refactorings, software engineering | | Claude Opus 4.5 | EU | — | 200k | Opus class with EU hosting for high-quality requirements | | Claude Opus 4.6 | EU | — | 200k | Current EU Opus for the most demanding tasks | | Claude Opus 4.7 | EU | — | 1M | Opus with 1M context for very long documents and codebases | | Claude Opus 4.8 | EU | — | 1M | Current Opus with 1M context, sharper agentic coding and stronger reasoning | | Claude Opus 5 | EU | — | 1M | Newest Opus with 1M context, the family's highest quality tier | ## Google (Gemini) Gemini models run via Google Vertex and stand out with the largest context windows (1M+ tokens). Ideal for long documents, entire knowledge bases, or extensive file uploads. | Model | Hosting | Reasoning | Context | Good for | | ------------------------ | ------- | --------- | ------- | ----------------------------------------------------------------------------------------------------------- | | Gemini 2.5 Flash | EU | ✓ | 1M | Bulk processing, fast APIs, image analysis | | Gemini 2.5 Pro | EU | ✓ | 1M | Complex code projects, STEM, large datasets, agentic workflows | | Web Search | EU | ✓ | 200K | Fast, cited web answers using Claude Haiku 4.5 and Linkup search | | Gemini 3 Flash (Preview) | US | ✓ | 1M | Being phased out — use Gemini 3.5 Flash instead | | Gemini 3.5 Flash | EU | ✓ | 1M | Latest Flash, combines reasoning at Pro level with high speed, multimodal input (text, image, audio, video) | | Gemini 3.6 Flash | EU | ✓ | 1M | Newest Flash, reasoning at Pro level at nearly halved latency, multimodal input (text, image, audio, video) | | Gemini 3.5 Flash-Lite | EU | ✓ | 1M | Budget-optimized for speed and high volume, with reasoning and multimodal input | | Gemini 3.1 Pro (Preview) | US | ✓ | 1M | Multimodal reasoning, long contexts, agentic workflows — preview | ## Mistral Mistral models are hosted in European data centers and offer a strong EU-native alternative for business applications. | Model | Hosting | Reasoning | Context | Good for | | -------------------- | ------- | --------- | ------- | ------------------------------------------------------------------- | | Mistral Medium 3 | EU | — | 128k | Enterprise deployments, multimodal analysis, long documents, coding | | Mistral Medium 3.5 | EU | — Coding | 256k | Latest Mistral Medium generation with larger context | | Magistral Medium 1.2 | EU | ✓ | 128k | EU reasoning alternative for logic and analysis tasks | | CodeStral | EU | — Coding | 256k | IDE code completion, fill-in-the-middle, code tests | ## DeepSeek DeepSeek models run via Azure AI Foundry and offer cost-efficient reasoning and coding options. | Model | Hosting | Reasoning | Context | Good for | | ----------------- | ------- | --------- | ------- | ------------------------------------------------------------------ | | DeepSeek-V3.2 | Global | — Coding | 128k | Cost-efficient batch processing, coding (Python/C++), multilingual | | DeepSeek-V4-Flash | Global | ✓ Coding | 1M | Affordable reasoning and coding with a long context | | DeepSeek-V4-Pro | Global | ✓ Coding | 1M | More demanding reasoning and coding with a long context | ## Other models | Model | Provider | Hosting | Reasoning | Context | Good for | | ------------ | -------------------------------------------- | ------- | --------- | ------- | -------------------------------------------------------------------- | | GPT-OSS 120B | OpenAI open weights (hosted on Nebius) | EU | ✓ | 128k | Self-hosting, RAG, on-prem agents | | Kimi K2.5 | Moonshot (hosted on Azure AI Foundry) | Global | ✓ Coding | 131k | Reasoning and coding tasks with large context | | Kimi K2.6 | Moonshot (hosted on Azure AI Foundry Sweden) | Global | ✓ Coding | 256k | Successor to K2.5 with doubled context window | | Kimi K3 | Moonshot (hosted on Nebius) | EU | ✓ Coding | 1M | Largest context window in the family, EU-hosted reasoning and coding | ## Image generation Image models (FLUX family, GPT Image, Gemini image models) are documented separately under [Image model management](/en/docs/admin-guide/image-models) — including hosting region, premium status, and recommendations. ## Enabling or disabling models Which models your users can choose from in the workspace is controlled by you as an admin in the settings. Details: [Workspace configuration](/en/docs/admin-guide/workspace-configuration). **Can I bring my own LLM keys?** No. All language models come from meinGPT's own managed, GDPR-compliant provider accounts — you don't supply (and aren't billed through) your own OpenAI, Azure OpenAI, or Anthropic keys. As an admin you control **which** of the built-in models your workspace can use (enable/disable and default), but you can't add a model or a custom provider key to the picker. The only "bring your own key" option is for the DeepL **translation** service — see [Translation Settings](/en/docs/admin-guide/translation-settings). If you need a specific model that isn't in the list, contact us. ## Why do answers differ from ChatGPT, Claude or Gemini directly? Even when you pick the same model in meinGPT (e.g. GPT-5.5), the answers can differ from what you get in ChatGPT, the Claude app or Gemini directly. This is not a matter of lower model quality - it is the same model from the same provider. The differences come from the environment the model runs in: - **A different system prompt.** ChatGPT and others silently give the model their own heavily prepared system prompt and a fixed answer format. meinGPT uses a more neutral system prompt so the model follows your workspace and assistant settings. For consistent results, set tone and format directly in your prompt or in an [assistant](/en/docs/platform/assistants). - **No ChatGPT personalization.** In ChatGPT, saved "custom instructions" and memory from earlier conversations often shape the answer - without you noticing. In meinGPT you start without that invisible context. You can bring your saved context across using the [memory feature](/en/docs/platform/memory). - **Randomness is built in.** Language models are not deterministic - the same question yields slightly different wording from one run to the next, even in the same tool. - **Tools and web search.** Whether the model may search the web or use tools strongly affects the answer. Use the ["+" menu](/en/docs/platform/chats/eingabefeld) to check that the right tools (e.g. web search) are enabled. If meinGPT seems to give "worse" results than ChatGPT, it is almost never the model. State tone, role and format explicitly in your prompt, enable the tools you need - or store both permanently in an [assistant](/en/docs/platform/assistants). ## Why can meinGPT feel slower than ChatGPT or Claude natively? It's the same model with the same weights - answer quality is identical. Small speed differences from the native app usually come down to: - EU hosting: EU models deliberately run in European data centers (e.g. Azure Sweden Central, Google Vertex EU) rather than the fastest US regions. This routing path through our aggregation layer typically costs only fractions of a second, but can become noticeable. - Provider load: response times at the AI provider fluctuate during peak times - this affects native apps just the same. - Large input size: the more text the model must consider at once (long chat history plus attached documents, e.g. ~50,000 tokens), the longer it takes just to read the request — before the first word of the answer even appears. - Tools and agent steps: when an assistant uses tools (web search, document search, connectors), each tool call is its own complete model round-trip. Several steps in sequence add up — so a demanding task with a large context and multiple tool calls can legitimately run for several minutes. **A slow answer is not an error.** If a request runs for a long time and eventually returns a result, it worked successfully — it is not stuck. If nothing appears during the wait, the answer may have finished in the background; reload the chat to see it (see [Connection interrupted](/en/docs/platform/chat-errors#stream-interrupted)). If a task is too slow for you, pick a faster model (e.g. Claude Sonnet 4.6 or a Gemini Flash model), reduce the context you attach, or lower the assistant mode for assistants. Sustained or sudden slowdowns are almost never a meinGPT server issue - they're provider-side. For a specific error message in chat, see [Chat error messages](/en/docs/platform/chat-errors). ## Related pages - [Image models](/en/docs/admin-guide/image-models) — image generation models and their use cases - [Workspace configuration](/en/docs/admin-guide/workspace-configuration) — control which models are enabled per workspace - [Privacy](/en/docs/privacy-security) — GDPR compliance, EU hosting, provider overview - [Pricing](/en/docs/admin-guide/pricing) — token prices per model