All Models

Below is a detailed overview of usage-based pricing. All prices are in Euro.

Text generation

Prices per 1M tokens in Euro, grouped by provider.

For models with tiered pricing, the default auto-compaction threshold is at most 90% of the first price boundary when the feature is enabled in the workspace. Compaction starts after a response; a large new input or a manually raised threshold can still trigger the higher rate. See Workspace Configuration for the setting.

OpenAI

GPT-6 Astra runs at OpenAI with EU data residency. It only becomes selectable once your organization has signed DPA version 2.0, which lists OpenAI Ireland Ltd. as a sub-processor — until then it stays locked in Settings → Models.

Above 272,000 tokens of context the long-context rate applies: €22.73 per 1M input tokens and €85.24 per 1M output tokens for GPT-6 Astra, €5.28 and €19.80 for GPT-6.1 Sol.

Anthropic

Google

Mistral

Open-source models

Cached input tokens

Several of our models support prompt caching: when you send the same (or very similar) context repeatedly, the already-processed part isn't billed again at the full rate. How this is actually billed differs between two mechanisms:

Models with an explicit cache-write charge (Claude models, the GPT-5.6 family, GPT-6 Astra, GPT-6.1 Sol): these models bill two separate rates - a heavily discounted rate for reusing cached context (cache read, currently 10% of the standard input price, i.e. a 90% discount; 5% for GPT-6.1 Sol) and a one-time surcharge for the first time a block gets cached (cache write, currently 125% of the standard input price). The surcharge pays for itself once the cached block is read back more than once. This applies to all current Claude models (Opus, Sonnet, and Haiku lines) as well as GPT-5.6 (Sol/Terra/Luna), GPT-6 Astra, and GPT-6.1 Sol.

Models with a read-only cache discount, no separate cache-write charge (most other OpenAI models, Gemini): for these models, caching happens automatically in the background - there's no separate charge for creating the cache, only a discounted rate for reusing it. The size of the discount varies by model:

• Google Gemini (all current Flash and Pro variants): 90% discount on cached input tokens.
• GPT-5, GPT-5.1, GPT-5.2, GPT-5.3, GPT-5.4, GPT-5.5 and their Chat/Thinking variants (excluding GPT-5.4 Pro, see below): roughly 80-90% discount, depending on the exact model version.
• GPT-4.1, o3, o4-mini: 75% discount.
• GPT-4o, GPT-4o Mini: 50% discount.
• DeepSeek-V4 (Flash/Pro): its own, vendor-set cache rate.

Models with no cache discount: for a few models (including Mistral, self-hosted models like Qwen, GPT-OSS, as well as GPT-5.4 Pro and Kimi K3), no cache discount is currently configured - cached tokens are billed at the full input rate for these.

The exact rates are set globally per model, the same for every organization, and can change as models are updated - the standard input/output prices listed above remain the reliable reference point that the cache discount is calculated as a percentage of.

Image generation

Prices are for a single image at standard resolution (1024×1024). For FLUX.2 models, the price may vary depending on resolution and number of reference images.

Translation

Other features

• Web reading: token-based consumption
• Workflow Wizard: €0.10 per workflow
• Audio transcription: €1.00 per hour
• PII filter: €1.50 per 1M tokens (input and output) the filter processed, on top of the model price; plus a monthly base fee as set in your offer

Related pages