connic model catalog

Exact model IDs, supported inputs, context limits, and EUR token prices for connic-managed inference.

Configuration

Use an exact catalog ID as the primary or fallback model. A Project can mix connic-managed and BYOK models.

model: connic/glm-5.2
fallback_model: connic/gpt-5.6-luna # or anthropic/claude-sonnet-4-6 with Anthropic BYOK configured
temperature: 0.3
reasoning_effort: high

Available models

Prices are net EUR per one million tokens. A dash means no cached-input discount is available. The cache write price applies to input tokens the model writes to its prompt cache. Listed reasoning effort values are the model-specific YAML overrides. Omit reasoning_effort or use auto for model-managed behavior. Standard and Fast IDs share one row and the same prices. Open the info icon beside a Fast ID for its supported reasoning settings, context limit, and input modalities. Models with audio input and output prices, such as Gemini 3.8 Live, only serve voice agents.

Provider labels beside a model name report whether the provider is EU-native. EU-native provider marks an EU-native provider, while Non-EU-native provider marks a model whose provider is not EU-native. The additional 30-day provider retention marks one whose provider retains request data for 30 days under its own policy, and 24-hour provider retention marks a voice model whose provider keeps a call's session for up to 24 hours so that long calls can continue when the connection to the model is renewed. Models without these labels run only on providers with a zero-data-retention policy. Inference still runs on EU capacity in every case.

Fast IDs

An ID ending in -fast selects a latency-optimized model variant. Most models list standard and Fast IDs at the same price; some list only a Fast ID.

model: connic/glm-5.2-fast
reasoning_effort: high   # the catalog lists the efforts each Fast ID accepts

• Fast and standard IDs can produce different output for the same prompt.
• Fast and standard IDs can differ in supported reasoning_effort values, context window, and input modalities. Setting a value the Fast ID does not accept is rejected.
• The info icon beside a Fast ID lists its differences from the standard ID. No listed differences means the supported settings match.

File input

Models with the Files label read PDFs natively, including tables and scanned pages. connic sends the file to the model unchanged. The label’s tooltip lists the accepted file types. Some of these models also accept office documents:

• GPT-5.4 and GPT-5.6: PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), RTF, OpenDocument (.odt, .ods, .odp), and EPUB.
• Every other model with the Files label: PDF.

Text files (.txt, .csv, .tsv, .md, .html, .json, .jsonl, .xml) reach every model as text. Any other file in the latest message fails with unsupported_file_type. Convert the file to PDF, or extract its content in a tool or middleware. connic replaces unsupported attachments in earlier messages with text notes so later requests can continue.

• A model request can be up to 48 MiB after base64 encoding. Supported files stay in the conversation for later requests in the run, so several large files can reach this limit.
• Return files from tools with ToolFile(data=...). See Tools.
• Pages count as input tokens. The number of tokens per page differs between models, and scanned pages can cost more than text pages.
• A fallback_model should accept the same file types as the primary model. Otherwise, after a switch, such a file in the latest message fails on the fallback too. The deployment log warns about such a pair.

Fallbacks and retries

All failed model requests follow retry_options, except that unsupported_file_type is not retried on the same model. Without a fallback, the primary gets up to attempts requests, including the first. With fallback_model, the primary gets one attempt and the fallback gets that budget, which defaults to three. A switch marks the run with context.fallback_model_used.

Each request has a 120-second timeout by default and is capped by the run’s remaining time. Context-window errors first use configured compression and one extra same-model repair request, without resetting the attempt counts or deadline. Cancellation, the run deadline, or output already delivered by the request stops further attempts. Completed tools keep their results. See execution limits and retries for timeout, backoff, context repair, and streaming behavior.

Cancelled requests

When a non-streaming managed request is cancelled by a run timeout or a manual stop, its reserved Project credit is released. After a streaming response begins, connic continues receiving the model response if the client disconnects and bills the reported token usage. If that stream ends without usable usage data, including at the request timeout, its reserved credit is not automatically released.

EU inference and data handling

• Every connic/* request runs in the EU. A request fails when EU execution is unavailable.
• Managed inference is not used to train models. connic records model and token counts for billing rather than prompt or response content; narrow security and error-retention exceptions are described in the Privacy Policy.

EU Project boundary

The connic/* guarantee covers the model call. A complete Project remains EU-only when every customer-configured component, including BYOK providers, tools, guardrails, judges, and external data destinations, also stays in the EU.