Models
What each privacy mode offers, the presets, how TEE models pass attestation, and how pricing works.
All top models live in one place, and your mode sets the menu. Anonymous brings text, image and video models. TEE brings text models running in confidential-compute enclaves. Image questions, image generation and video are Anonymous only.
Each mode does different work
| Anonymous | TEE | |
|---|---|---|
| Text chat | 20 models | 12 open-weight models |
| Image questions (vision) | Yes, on vision models | No, encryption carries text only |
| Image generation | 13 models, 8 of which keep nothing | No |
| Video generation | 7 models | No |
| Tool calling | Yes, on supporting models | No, tool definitions would travel in the clear |
| Prompt readers | The provider, which may log it | The attested enclave alone |
Presets cover the common choices
A few presets sit up front in each mode, with All models behind them.
| Preset | TEE (text) | Anonymous (text) | Anonymous (image) | Anonymous (video) |
|---|---|---|---|---|
| Fast | GPT OSS 120B | DeepSeek V4.1 Flash | FLUX.2 Klein 4B | MiniMax H3 Max Turbo |
| Smart | GLM 5.3 | Claude Sonnet 5.5 | Seedream 4.5 | Flux 3 Max |
| Top | GLM 5.3 (see below) | Claude Opus 5.5 | Nano Banana (Gemini 2.5 Flash Image) | Grok Imagine 1.5 |
| Uncensored | Gemma 4 26B | DeepSeek V4.1 Flash Uncensored | None | None |
| Vision | None (TEE is text only) | GPT-6 Luna | None | None |
- Top under TEE means the strongest model that passes attestation at that moment. The preset is GLM 5.3, matching Smart.
- No Vision under TEE, since encryption carries text and photos can't be attached. Qwen3 VL 30B A3B appears there for text alone. For picture questions, use Vision in Anonymous, where the provider sees the image.
- Uncensored refuses less. Acceptable use rules hold all the same, and media prompts still get moderated.
Anonymous mode covers text, image and video
| Type | Models |
|---|---|
| Text (20) | Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6 Luna, GPT-5.6 Luna, GPT-4o Mini, Gemini 3.1 Flash Lite, Grok 4.7, DeepSeek V4 Pro, DeepSeek V4.1 Flash, DeepSeek V4.1 Flash Uncensored, DeepSeek V3.2, Kimi K3, Kimi K2.7 Code, GLM 5.3, Qwen3.7 Flash, Qwen3 VL 235B A22B, MiniMax M3.1 Flash, Mistral Large 3, Llama 3.3 70B |
| Image (13) | Nano Banana (Gemini 2.5 Flash Image), GPT-5 Image Mini, Seedream 4.5, FLUX.2 Flex, FLUX.2 Klein 4B; not stored: Grok Imagine 2.0, Grok Imagine High Quality, Grok Imagine, Krea 2 Turbo, Hunyuan Image 3.0, Z-Image Turbo, Chroma, Anime (WAI) |
| Video (7) | Grok Imagine 1.5, Grok Imagine 1.5 Lite, Flux 3 Max, MiniMax H3 Max, MiniMax H3 Max Turbo, LongCat, LongCat Distilled |
Image models tagged Not stored use providers that hold no copy of prompt or output. That rests on the provider's word, whereas TEE rests on hardware attestation. Any video model can also start from an uploaded image used as frame one.
TEE mode offers 12 text models
Phala and NEAR AI run the enclaves behind all of them. Attestation is per model.
| Model | Context | Enclave operator |
|---|---|---|
| Kimi K3 | 1M | Phala |
| GLM 5.3 | 1M | Phala |
| GLM 5.2 | 524K | Phala |
| Kimi K2.6 | 262K | NEAR AI |
| Qwen 3.8 27B | 262K | NEAR AI |
| Qwen3 VL 30B A3B (text only in TEE) | 128K | Phala |
| Gemma 4 26B A4B Uncensored | 64K | Phala |
| Qwen 3.6 35B A3B FP8 | 32K | NEAR AI |
| DeepSeek V4 Flash | 1M | NEAR AI |
| GLM 5.3 Flash | 1M | NEAR AI |
| GPT OSS 120B | 128K | Phala |
| Qwen 2.5 7B | 32K | Phala |
A model that fails remains selectable, minus the badge. In that case no hardware evidence shows the model sits in a genuine enclave. Attestation details the checks.
Credits open every model
With credits you reach all models in both modes, media included. Main and Bonus behave identically everywhere, and tiers resize the Bonus without changing access.
Beta models fail more often
The beta label signals limited availability and a higher failure rate. When a model can't serve you, a "model unavailable" error appears and no money moves.
Prices are per use and shown up front
The picker lists a public rate per million tokens, per image or per video. Tier makes no difference, and subscriptions don't exist.
- Anonymous text: a secondary compute market serves it, usually under the official list price. The cheapest offer available for the model goes first, and you owe what that request really cost, capped at the official list price.
- TEE: the official list price, exactly, on real usage.
- Images and video: the listed price.
More on Credits.
The picker is the full catalogue
Open it for prices, beta labels and, under TEE, attestation status. Our server retests each TEE model about once an hour, and your browser tests again before displaying a badge.