connic/claude-opus-5.5Anthropic · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €4.20
- Output / 1M
- €20.80
The right model for every agent.
No contract with each provider.
Use EU-hosted connic models billed per token from Project credit, your own provider keys or a model server you run. Choose per agent, and name a fallback for when a request fails.
Pick the model by what the job needs.
Three jobs at a hotel group and a tour operator, each with the model setup that fits it.
The guest chat names a second model as its fallback. When requests to the first fail, the fallback answers, and the guest gets a normal reply.
Fallback model usedguest-replyconnic/claude-sonnet-5.5 · 503→ connic/gpt-5.6-terra answered
Usesfallback_modelretry_optionscontext.fallback_model_used
Tour groups send guest lists as Excel files, PDFs or Word documents. A model that reads these files directly turns each into rooms and names and flags the rows it can't place.
UsesFile inputconnic/gpt-5.6-terraEmail connector
Reviews mention names and room numbers, so the hotel group tags them with a model on its own GPU server. connic reaches it through a bridge, with no inbound ports.
UsesCustom providervllm/review-taggerBridge
Claude, GPT, Gemini and open-weight models, with no provider account.
connic/* models work in every Project from day one. You pay per token in EUR from Project credit, the same balance that pays for runs. Prices are net EUR per million tokens.
Already have a provider contract? Add your own key for OpenAI, Azure OpenAI, Anthropic, Google Gemini, OpenRouter, AWS Bedrock or Google Vertex AI, or point agents at an OpenAI-compatible server you run.
| Model | Context | Inputs & strengths | Input / 1M | Output / 1M |
|---|---|---|---|---|
connic/claude-opus-5.5Anthropic · Provider terms | 1M | TextVisionFilesReasoning | €4.20 | €20.80 |
connic/claude-sonnet-5.5Anthropic · Provider terms | 1M | TextVisionFilesReasoning | €2.20 | €10.70 |
connic/claude-haiku-5.5Anthropic · Provider terms | 1M | TextVisionFilesReasoning | €0.10 | €0.50 |
connic/gpt-6-astraOpenAI · Provider terms | 1M | TextVisionReasoning | €11.60 | €57.90 |
connic/gpt-6.1-solOpenAI · Provider terms | 1M | TextVisionReasoning | €2.30 | €11.40 |
connic/gpt-6-lunaOpenAI · Provider terms | 1M | TextVisionReasoning | €0.15 | €0.60 |
connic/gemini-4-argonGoogle · Provider terms | 1M | TextVisionFilesReasoning | €2.20 | €10.70 |
connic/gemini-3.8-flashGoogle · Provider terms | 1M | TextVisionFilesReasoning | €0.80 | €3.90 |
connic/gemini-3.5-flash-liteGoogle · Provider terms | 1M | TextVisionFilesReasoning | €0.35 | €2.60 |
connic/kimi-k3-fastMoonshot AI · Kimi K3 License | 1M | TextVisionReasoning | €3.20 | €15.80 |
connic/glm-5.3Z.ai · GLM-5.3 License | 1M | TextReasoning | €1.50 | €4.60 |
connic/deepseek-v4-pro-0813DeepSeek · MIT | 1M | TextReasoning | €1.90 | €3.80 |
connic/qwen3.8-2.4t-a95bAlibaba Cloud · Qwen3.8-Max License | 262k | TextReasoning | €2.40 | €5.70 |
connic/mistral-medium-3.5-128bMistral · Modified MIT | 180k | TextVisionReasoning | €1.60 | €7.90 |
connic/mistral-small-2603Mistral · Provider terms | 256k | TextVisionReasoning | €0.15 | €0.60 |
connic/kolibri-1Aleph Alpha · Apache 2.0 | 262k | TextReasoning | €0.20 | €0.60 |
connic/claude-sonnet-5.5Anthropic · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €2.20
- Output / 1M
- €10.70
connic/claude-haiku-5.5Anthropic · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €0.10
- Output / 1M
- €0.50
connic/gpt-6-astraOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €11.60
- Output / 1M
- €57.90
connic/gpt-6.1-solOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €2.30
- Output / 1M
- €11.40
connic/gpt-6-lunaOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €0.15
- Output / 1M
- €0.60
connic/gemini-4-argonGoogle · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €2.20
- Output / 1M
- €10.70
connic/gemini-3.8-flashGoogle · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €0.80
- Output / 1M
- €3.90
connic/gemini-3.5-flash-liteGoogle · Provider terms
TextVisionFilesReasoning
- Context
- 1M
- Input / 1M
- €0.35
- Output / 1M
- €2.60
connic/kimi-k3-fastMoonshot AI · Kimi K3 License
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €3.20
- Output / 1M
- €15.80
connic/glm-5.3Z.ai · GLM-5.3 License
TextReasoning
- Context
- 1M
- Input / 1M
- €1.50
- Output / 1M
- €4.60
connic/deepseek-v4-pro-0813DeepSeek · MIT
TextReasoning
- Context
- 1M
- Input / 1M
- €1.90
- Output / 1M
- €3.80
connic/qwen3.8-2.4t-a95bAlibaba Cloud · Qwen3.8-Max License
TextReasoning
- Context
- 262k
- Input / 1M
- €2.40
- Output / 1M
- €5.70
connic/mistral-medium-3.5-128bMistral · Modified MIT
TextVisionReasoning
- Context
- 180k
- Input / 1M
- €1.60
- Output / 1M
- €7.90
connic/mistral-small-2603Mistral · Provider terms
TextVisionReasoning
- Context
- 256k
- Input / 1M
- €0.15
- Output / 1M
- €0.60
connic/kolibri-1Aleph Alpha · Apache 2.0
TextReasoning
- Context
- 262k
- Input / 1M
- €0.20
- Output / 1M
- €0.60
Catalog updated November 4, 2026See the full model catalog
When a model request fails, the fallback finishes the run.
Rate limits, timeouts and provider errors are retried per request. With fallback_model set, the retry goes to a second model, even one from another provider.
- One attempt budget
- retry_options.attempts allows 1 to 10 requests, 3 by default. With a fallback, the primary gets one attempt and the fallback gets the budget.
- Same request, other model
- The fallback gets the identical request, including reasoning_effort and attached files. Pick a model that supports both.
- Finished work stays
- Only the failed request is sent again. Tools that already ran keep their results, and the run keeps its deadline.
connic/* requests stay in the EU.
Every connic/* request runs on EU capacity and fails if none is available. Managed inference is not used to train models.
The guarantee covers the model call. A Project stays EU-only when its BYOK providers, tools, guardrails, judges and data destinations do too.
connic/claude-sonnet-5.5- Runs onEU inference capacityCostProject credit, in EUR
anthropic/claude-sonnet-5-5- Runs onAnthropic, under your accountCostYour provider contract
vllm/review-tagger- Runs onYour server, through a bridgeCostYour own infrastructure
Each agent names its model in one line.
The three agents behind the hotel examples on this page. Switching a model is a one-line change your team reviews like any other.
- A small model where it's enough
- guest-triage sorts every message on a small model. Only rebookings and complaints reach the larger one.
- One model for a whole folder
- _defaults.yaml sets the model for every agent in its folder and below, and an agent file can still override it.
- Cost per model
- Token Usage groups spend by model and by agent, so you see what each choice costs.
agents/guests/guest-triage.yaml
version: "1.0"
name: guest-triage
description: Sorts guest messages and answers routine ones
model: connic/mistral-small-2603agents/guests/guest-reply.yaml
version: "1.0"
name: guest-reply
description: Drafts replies for rebookings and complaints
model: connic/claude-sonnet-5.5
fallback_model: connic/gpt-5.6-terra
reasoning_effort: mediumagents/reviews/review-tagger.yaml
version: "1.0"
name: review-tagger
description: Tags guest reviews by topic
model: vllm/review-tagger # your own GPU serverKeep exploring
composer SDKName the model in the same file as the agent's prompt and tools.A/B testingTry a cheaper model on part of live traffic first.TestingRun your test cases against a new model before you switch.ObservabilityCompare token cost by model and agent, and see each model call's duration in the trace.Private networkingReach a model server inside your network.Custom providersPoint agents at any OpenAI-compatible endpoint.
Frequently Asked Questions
No. connic/* models are on in every Project, need no provider key and bill from Project credit. The free basic plan gives your first Project €20 credit for 14 days, the same balance that pays for runtime, storage and retrieval.
No. Managed inference is not used to train models, and connic records the model and token counts for billing, not the content of prompts and responses. The privacy policy describes narrow exceptions for security and error retention. The catalog marks which providers are EU-native and which keep request data for 30 days under their own policy.
Timeouts, rate limits and provider errors follow the agent's retry_options, which allow up to three requests to the same model by default. With fallback_model set, the primary is tried once and the fallback gets that budget. Only the failed request is repeated, so tools that already ran keep their results.
Per token in EUR, from Project credit. The connic model catalog in the docs lists the current price of every model. On prepaid Projects, turn on Auto-refill before you send production traffic, because new usage pauses when the credit runs out.
It covers the model call: every connic/* request runs in the EU. The Project as a whole stays in the EU when everything else you configure does too, including BYOK models, tools, guardrails, judges and external data destinations.
Yes. Add a custom provider with a prefix such as vllm and the server's OpenAI-compatible base URL, then pick a bridge under Route via Bridge. Agents reference it like any other model, for example vllm/your-model, and requests reach the server through the bridge's outbound connection.