| Claude Opus 5.5 and Claude Haiku 5.5 | Free plan with a limit that resets every 5 hours; you must be 18 or over. | Projects with your own documents, even on the free plan (up to 5). | Pro from €18 per month in Spain; via the API, Opus 5.5 costs USD 4.00 input and USD 20.00 output per million tokens. | No image generation of its own (still to be confirmed). | Haiku 5.5 for repetitive tasks via the API, from USD 0.10 per million input tokens. | It does not train on your chats unless you allow it; check the model-improvement setting and use incognito mode. | Anthropic · United States | No · proprietary | 1,000,000 tokens | No | Free plan | Pro: €18 per month | USD 0.10 per million input tokens | Only if you allow it | Sources: 5 |
| DeepSeek-V4.1-Flash-1017 (and DeepSeek-V4-Pro) | Free and powerful, but chats are stored in China: nothing personal. | Avoid it with client data; alternative: the same model hosted in Europe or run locally. | API at USD 0.30 input and USD 1.20 output per million at peak time, and MIT weights. | No tests of our own. | For low-cost automation via the API, bearing in mind that data goes to China; or deploy the MIT weights in your own cloud. | The app, no; the open weights on your own infrastructure, yes. | DeepSeek · China | Yes · MIT | 1,000,000 tokens | Yes | Free downloadable weights | not officially published | USD 0.15 per million input tokens | It may; opt out on request or by email | Sources: 3 |
| Gemini 4 Argon | Free Gemini plus Gemini Notebook to study with your own notes; Argon needs a paid plan. | A good fit if you work in Gmail and Docs; Argon from Google AI Pro (€21.99 per month). | 1M-token context via the API. | Images with Nano Banana. | Google AI Plus (€4.99 per month) is the cheap way in, but it does not include Gemini 4 Argon. | Turn off activity saving or use temporary chats. | Google · United States | No · proprietary | 1,048,576 tokens | No | Paid only | Google AI Pro: €21.99 per month | USD 2.00 per million input tokens (from 1 December 2026: USD 4.00) | Yes, by default; you can turn it off in settings | Sources: 6 |
| GLM-5.3 (and GLM-5.3-Flash) | No tests of our own yet. | Useful for long documents thanks to its 1M context; be careful with personal data in the chat. | Built for coding and long tasks; via the API, USD 1.40 input and USD 4.40 output per million tokens. | Not its stated strength. | A very permissive licence for almost any company; it can be hosted on your own infrastructure. | Better self-hosted or via the API (content is not stored) than in the web chat. | Z.ai (Zhipu) · China | Yes · GLM-5.3 License (MIT-style; security review for model-as-a-service operators with more than USD 10 billion a year) | 1,000,000 tokens | Yes | Free downloadable weights | not officially published | USD 0.15 per million input tokens | It may; opt out on request or by email | Sources: 6 |
| GPT-6 (Luna, Sol, Astra) and GPT-6.1 Sol | The free plan already includes GPT-6 Luna and study mode. | Go (€8 per month) or Plus (€23 per month), depending on how much you use it. | Codex with GPT-6.1 Sol from Plus; via the API, USD 2.00 input and USD 10.00 output per million tokens. | Image generation included, with limits. | Business, Enterprise and Edu do not train on your data by default; via the API you can choose EU data residency. | Turn off “Improve the model for everyone” and use temporary chats. | OpenAI · United States | No · proprietary | 1,050,000 tokens | No | Free plan | Go: €8 per month | USD 0.10 per million input tokens | Yes, by default; you can turn it off in settings | Sources: 6 |
| Grok 4.7 | The free plan has limited use, with no published figure. | xAI says it is better at documents and presentations; our own test is pending. | API at USD 2.00 input and USD 6.00 output per million tokens (prompts under 200K); 500K context. | Grok Imagine marks its images with “grok”. | Prices only in USD and excluding VAT; it uses your chats for training unless you change that. | Check the data controls before using it; private chats are deleted after 30 days. | xAI (named SpaceXAI LLC in its policy) · United States | No · proprietary | 500,000 tokens | No | Free plan | Plus: USD 8 per month | USD 2.00 per million input tokens | Yes; how to turn it off is not officially published | Sources: 4 |
| Kimi K3 | Possible for long texts, but app data is stored in China; do not upload notes with personal data. | Not recommended for client documents because of where the data is stored. | Interesting via the API (it does not train on API data); compare prices once they are verified. | No tests of our own yet. | Open weights with a permissive licence up to very high thresholds; for production, better hosted by a European provider. | The app stores data in China and trains by default; if that worries you, use the API or host it yourself. | Moonshot AI · China | Yes · Kimi K3 License (MIT-style base with conditions) | 1,048,576 tokens | Yes | Free downloadable weights | not officially published | not officially published | It may; opt out on request or by email | Sources: 6 |
| Qwen3.8-27B | The 27B runs on your own computer without sending your notes to anyone, if you have a powerful machine. | Alibaba says it excels at office tasks; our own test with the large model is pending. | A good local candidate with an Apache 2.0 licence. | It understands images and video; for generating images, the Qwen-Image family is separate. | Apache 2.0: commercial use with no caps; a good base to host on your own server. | If you run it locally, your data does not leave your computer. | Alibaba Cloud (Qwen) · China | Yes · Apache 2.0 | 262,144 tokens | Yes | Free downloadable weights | not officially published | not officially published | It may; opt out on request or by email | Sources: 3 |