connic/claude-opus-5Anthropic · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €6.50
- Output / 1M
- €26.00
Use your keys.
Or use ours.
Every agent names its own model: an EU-hosted connic/* model with no account to create, your own provider key, or an endpoint you host. Retries and cross-provider failover come with it.
Read the model configuration docsModel selection
Connic-managedEU-hosted
connic/glm-5.2
Z.ai · 256k context
Bring your own key
anthropic/claude-sonnet-4-5
Uses your provider account
Provider choice
Three ways to run a model. One YAML field.
Pick one per agent, and any of them can serve as that agent's fallback. Moving to a different provider is a one-line change, not a rewrite.
Connic-managed
No account to create
model: connic/glm-5.2- Enabled on every Project by default, with no key to configure
- EU-only inference at published EUR rates
- Paid from Project credit, and Basic starts you with a one-time 25 EUR
Bring your own key
Your provider account
model: anthropic/claude-sonnet-4-5- OpenAI, Azure OpenAI, Anthropic, Gemini, OpenRouter, Bedrock, Vertex AI
- Your provider contract, your rates, your rate limits
- Credentials encrypted per Project and injected at runtime
Your own endpoint
Any OpenAI-compatible server
model: vllm/mistral-7b- Pick a prefix, point it at a base URL, and use it like any provider
- vLLM, Ollama, an inference proxy, or a fine-tune you host yourself
- Reach an endpoint inside your own network over a Bridge, no inbound ports
A single Project can mix all three. See the provider prefixes and required credentials.
Current catalog
Exact models. Published prices.
A selection from the managed catalog, priced in net EUR per million tokens and billed from Project credit. Every connic/* ID points at one named model and keeps pointing at it, so a deploy never swaps out the model you tested.
| Model | Context | Inputs & strengths | Input / 1M | Output / 1M |
|---|---|---|---|---|
connic/claude-opus-5Anthropic · Provider terms | 1M | TextVisionReasoning | €6.50 | €26.00 |
connic/claude-sonnet-5Anthropic · Provider terms | 1M | TextVisionReasoning | €2.60 | €10.40 |
connic/gpt-5.6-solOpenAI · Provider terms | 1M | TextVisionReasoning | €5.20 | €31.20 |
connic/gpt-5.6-terraOpenAI · Provider terms | 1M | TextVisionReasoning | €2.60 | €15.60 |
connic/gpt-5.6-lunaOpenAI · Provider terms | 1M | TextVisionReasoning | €1.10 | €6.30 |
connic/gemini-3.5-flashGoogle · Provider terms | 1M | TextVisionAudioReasoning | €1.40 | €8.40 |
connic/gemini-3.1-flash-liteGoogle · Provider terms | 1M | TextVisionAudioReasoning | €0.30 | €1.60 |
connic/glm-5.2Z.ai · MIT | 256k | TextReasoning | €1.90 | €5.80 |
connic/qwen3.5-397b-a17bQwen · Apache-2.0 | 250k | TextVisionReasoning | €0.65 | €3.80 |
connic/qwen3.6-35b-a3bQwen · Apache-2.0 | 256k | TextVisionReasoning | €0.30 | €1.60 |
connic/mistral-large-2512Mistral · Provider terms | 256k | TextVision | €0.55 | €1.60 |
connic/mistral-small-2603Mistral · Provider terms | 256k | TextVisionReasoning | €0.25 | €0.65 |
connic/llama-3.3-70b-instructMeta · Llama 3.3 Community | 100k | Text | €0.95 | €0.95 |
connic/minimax-m2.7MiniMax · Provider terms | 197k | TextReasoning | €0.65 | €2.60 |
connic/claude-sonnet-5Anthropic · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €2.60
- Output / 1M
- €10.40
connic/gpt-5.6-solOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €5.20
- Output / 1M
- €31.20
connic/gpt-5.6-terraOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €2.60
- Output / 1M
- €15.60
connic/gpt-5.6-lunaOpenAI · Provider terms
TextVisionReasoning
- Context
- 1M
- Input / 1M
- €1.10
- Output / 1M
- €6.30
connic/gemini-3.5-flashGoogle · Provider terms
TextVisionAudioReasoning
- Context
- 1M
- Input / 1M
- €1.40
- Output / 1M
- €8.40
connic/gemini-3.1-flash-liteGoogle · Provider terms
TextVisionAudioReasoning
- Context
- 1M
- Input / 1M
- €0.30
- Output / 1M
- €1.60
connic/glm-5.2Z.ai · MIT
TextReasoning
- Context
- 256k
- Input / 1M
- €1.90
- Output / 1M
- €5.80
connic/qwen3.5-397b-a17bQwen · Apache-2.0
TextVisionReasoning
- Context
- 250k
- Input / 1M
- €0.65
- Output / 1M
- €3.80
connic/qwen3.6-35b-a3bQwen · Apache-2.0
TextVisionReasoning
- Context
- 256k
- Input / 1M
- €0.30
- Output / 1M
- €1.60
connic/mistral-large-2512Mistral · Provider terms
TextVision
- Context
- 256k
- Input / 1M
- €0.55
- Output / 1M
- €1.60
connic/mistral-small-2603Mistral · Provider terms
TextVisionReasoning
- Context
- 256k
- Input / 1M
- €0.25
- Output / 1M
- €0.65
connic/llama-3.3-70b-instructMeta · Llama 3.3 Community
Text
- Context
- 100k
- Input / 1M
- €0.95
- Output / 1M
- €0.95
connic/minimax-m2.7MiniMax · Provider terms
TextReasoning
- Context
- 197k
- Input / 1M
- €0.65
- Output / 1M
- €2.60
Catalog updated July 31, 2026See the full model catalog
Privacy boundary
EU inference built into connic/*
No regional flag to remember and no global fallback hidden behind the model ID.
EU-only by contract and routing
Every connic/* request is sent only to EU inference capacity. A model is removed or fails closed if that boundary cannot be maintained.
No prompt training
Connic-managed inference is not used to train models. Connic records model and token counts for billing, not prompt or response content; narrow security and error-retention exceptions are described in our Privacy Policy.
A clear Project boundary
connic/* covers the model call. Your Project is fully EU-resident when its deployment region and every configured provider, tool, guardrail, judge, and data destination are EU-resident too.
Configuration
Model settings live in the agent file
Your model choice sits in Git next to the prompt and the tools, so switching model is a pull request with a reviewer, not a console setting nobody can trace.
agents/refund-agent.yaml
version: "1.0"
name: refund-agent
description: Reviews refund requests and drafts a decision
system_prompt: |
Review the refund request and draft a decision.
model: connic/qwen3.5-397b-a17b
fallback_model: anthropic/claude-sonnet-4-5
temperature: 0.3
reasoning_effort: high
retry_options:
attempts: 5
max_delay: 60What each field does
model- A provider prefix plus a model ID. Required for LLM agents, and managed, BYOK, and custom endpoints all use this one field.
fallback_model- A second model, usually at a different provider, that takes over when the primary provider fails.
temperature- 0.0 to 2.0, within what the model accepts. Lower is more deterministic.
reasoning_effort- auto, off, minimal, low, medium, high, or xhigh. The level is translated to each provider's own parameter, so the file survives a model swap.
retry_options- Attempts per model, up to 10, plus the initial and maximum delay between them.
One line changes a whole fleet
A _defaults.yaml file sets the model for every agent in that directory and below, and an individual agent still wins on conflict. See cascading defaults
Tool calling and structured output
Managed models accept tool definitions and response-format constraints, so agent tools and output_schema behave the same on connic/* as they do on a BYOK model.
Failover
A provider outage stops being an incident
Rate limits, timeouts, and revoked keys are everyday events once agents run in production. Connic handles them at the model call, not in your application code.
LLM call, one run
- connic/qwen3.5-397b-a17battempt 1 of 4429rate limited, retry in 7s
- connic/qwen3.5-397b-a17battempt 2 of 4503provider unavailable, switching model
- anthropic/claude-sonnet-4-5fallback, attempt 1 of 42001.9s, run continues
Run finished successfully and is tagged
fallback_model_usedRetries at the failure point
Connic repeats the single failing model call instead of replaying the run. Tools that already executed keep their results, and the conversation is not rebuilt from scratch.
Failover across providers
fallback_model takes over on timeouts, rate limits, unavailable providers, and on authentication, quota, or unknown-model errors. An eligible error switches model straight away instead of burning the remaining attempts. Managed primary with BYOK fallback works, and so does the reverse.
Visible after the fact
A run that switched is flagged in its trace. Filter the run list with the expression context.fallback_model_used to see how often a provider let you down last week.
Context limits recover
Switch on context_compression and a context-window error triggers one pass that summarizes older turns and oversized tool results, then retries. The run keeps going instead of dying on a provider error.
Attempts, backoff, and delays are configurable per agent. See execution limits and retries
Model changes
Switch models on evidence
New models ship every few weeks. Run a candidate beside the model you already trust, on real traffic, and compare cost and quality before you commit.
agents/refund-agent-test-cheaper.yaml
# Copy of refund-agent.yaml, one line changed
version: "1.0"
name: refund-agent-test-cheaper
description: Reviews refund requests and drafts a decision
system_prompt: |
Review the refund request and draft a decision.
model: connic/qwen3.6-35b-a3brefund-agent, cheaper test
MetricControlVariantDelta
- Token cost per run€0.021€0.006-71%
- Success rate97.4%97.1%-0.3pp
- Judge score17/2017/20even
- Duration P952.4s3.1s+0.7s
Confidence mode plans the sample size and issues a recommendation at each analysis checkpoint. Guardrails pause a test whose failure rate or judge quality slips.
Prove it before live traffic
Point a test suite at the variant agent and run the same cases with the same assertions. Exit codes plug into any pipeline, and the deploy gate can run the suite before the variant goes live. See testing
Cost broken out per model
Token Usage reports spend by agent and by model, with EUR and USD kept in separate columns so a managed model and a BYOK model are never blended into one number. See usage docs
Limits that stop spend
Set alerts or hard limits on token cost, scoped globally, per environment, or per agent, daily or monthly. Anomaly detection flags a run that costs far more than that agent's 30-day average. See observability
The model call is the easy part
What has to exist around it in production, compared with the usual ways to run inference
| Feature | Connic | Provider SDK | Model router | Self-hosted |
|---|---|---|---|---|
| Ship without creating a provider account | Included | Not included | Not included | Not included |
| EU-only inference option | Included | Partial | Partial | Included |
| Per-model prices published in EUR | Included | Not included | Not included | Not included |
| Retries and cross-provider failover | Included | Not included | Included | Not included |
| One config for managed, BYOK, and self-hosted models | Included | Not included | Partial | Not included |
| Per-run trace with tool calls and token cost | Included | Not included | Partial | Not included |
| Spend alerts and hard limits | Included | Partial | Partial | Not included |
| A/B a model swap on live traffic | Included | Not included | Not included | Not included |
| Test suites and judges gating a model change | Included | Not included | Not included | Not included |
| No GPU capacity to operate | Included | Included | Included | Not included |
Frequently Asked Questions
No. Connic-managed models are enabled on every Project by default and need no provider key. The free Basic plan starts with a one-time 25 EUR of Project credit, the same balance that pays for runs, compute, and storage. Provider credentials are only needed for BYOK or custom providers.
The catalog includes both open-weight and hosted proprietary models. Each entry lists its creator and applicable model terms. You can also run an open model yourself and point Connic at it through a custom OpenAI-compatible endpoint.
Yes. Each agent YAML names its own model, so a cheap model can handle classification while a stronger model handles the reasoning step. A _defaults.yaml file sets a shared model for every agent in a directory and below, and an individual agent still overrides it.
A connic/* ID stays on the named model while Connic switches eligible capacity behind it when a route is unavailable and retrying is safe. If the call still fails definitively, the agent's fallback_model can take over. Requests with an uncertain outcome are not replayed.
Yes. Managed models accept tool definitions and response-format constraints, so agent tools, MCP servers, and output_schema behave the same way they do on a BYOK model.
Two ways. Run a test suite with the same cases and assertions against the variant agent in CI, or deploy the candidate as an A/B variant and split live traffic. The A/B comparison reports success rate, duration, token cost, and judge scores for both sides, and pauses a test that breaches a failure-rate or quality guardrail. Judges score deployed runs, so they apply to the A/B comparison rather than to the test suite.
No. BYOK and Connic-managed models are equal options. Either can be the primary or fallback model for an agent.
No. Managed models are part of the deployed Connic runtime, not a public pass-through model API.
Standard Projects can add prepaid credit manually or enable auto-refill with a threshold, amount, and monthly cap. Calls pause when a prepaid Project has no credit. Enterprise Projects are invoiced monthly in arrears.
Purchased credits stay available while the Project remains active. If the Project or contract closes, unused purchased credit is refunded after any outstanding amount is settled. Monthly credit expires at cycle end; promotional credit follows its disclosed expiry.
It guarantees the managed model call stays in the EU. The whole Project is EU-only when every other component you configure, such as BYOK models, tools, guardrails, and external data destinations, also stays in the EU.