Frontier Models & BYOK | Synter AI Media Agent
Model Transparency
Frontier Models & BYOK for Synter's AI Media Agent
Synter's AI Media Agent is model-aware and provider-agnostic. Teams can select a default model or bring your own key (BYOK) per workspace, and we auto-route by task (planning, copy, entity extraction, optimization notes).
Supported Providers & Models
| Provider | Models | Tools / Features | Status | Notes |
|---|---|---|---|---|
| Synter AI | Angstrom-32B (Domain Post-Trained), Angstrom-Omni | Zero-slop MCP tool calling, live audience grounding, GAQL & CAPI engine | Native | Synter's native domain foundation model. SOTA on GTM-Bench (96.4%) with 4x-6x token efficiency. |
| OpenAI | GPT-5, GPT-4o, o3, o4-mini | Function calling, structured outputs | Native | Default for planning and campaign brief generation |
| Anthropic | Claude Opus 4.5, Claude 4 Sonnet | Tool use, long context (200k tokens) | BYOK | Strong safety guardrails, excellent for complex briefs |
| Gemini 3.0, Gemini 2.5 Pro | Function calling, multimodal | BYOK | Multimodal capabilities for creative analysis | |
| Meta | Llama 3.3 70B, Llama 3.1 405B | Tool use via hosted endpoints | BYOK | Open weights via providers like Together AI, Fireworks |
| Mistral | Mistral Large 2, Mixtral 8x22B | JSON mode, function calling | BYOK | Fast inference, excellent for structured extraction |
| xAI | Grok-2 | Under evaluation | Roadmap | Evaluating for future integration |
Selection & Routing
Default model per workspace: Set your preferred provider and model at the workspace level. Override per task or per run if needed.
Fallback policy: If a provider is rate-limited or unavailable, Synter automatically falls back to an alternative model. You define the fallback order.
No training on your data: Model providers do not train on your inputs. Synter accesses providers under their business API terms, where training on customer inputs is off by default and inputs are retained only for the provider's standard abuse-monitoring window.
PII redaction: Sensitive data (emails, credit cards, API keys) is automatically redacted before being sent to any model.
Data Handling
Warehouse-centric: First-party conversions remain in your data warehouse (Snowflake, BigQuery, Databricks). Only minimal, necessary fields are shared with LLMs for planning and optimization.
Processing region: Model inference and data processing run on US infrastructure. Transfers for EU and UK customers are covered by our DPA with Standard Contractual Clauses. An EU processing region is on the roadmap.
Retention: Configurable log retention (0, 30, or 90 days). You control how long Synter stores API request/response logs for debugging and auditing.
Task-Specific Model Routing
Planning & Strategy
GPT-4o, o-series, or Claude 3.5 Sonnet
Campaign brief → channel mix, budget allocation, audience strategy. Uses function calling for structured outputs.
Long-Context Research
Claude 3.5/3.7 Sonnet
Analyze website content, competitor research, industry reports. 200k token context window for deep analysis.
Entity Extraction & Summaries
Gemini 1.5/2.0
Extract keywords, audiences, themes from unstructured data. Multimodal capabilities for image/video creative analysis.
Cost-Effective Routing
Llama 3.3, Mistral Large
High-volume tasks like keyword expansion, ad copy variants, simple summaries. Lower cost per token.
Frequently Asked Questions
Does Synter provide native domain foundation models?
Yes. Synter Angstrom-32B is our native, domain post-trained foundation model built specifically for programmatic media buying, GAQL search waste mining, and live audience grounding, achieving 96.4% on GTM-Bench with 4x-6x token efficiency.
Does Synter also support third-party frontier LLMs like GPT-4o, Claude, and Gemini?
Yes. Synter is provider-aware and supports BYOK. In addition to native Angstrom-32B, you can route tasks to OpenAI (GPT-5, GPT-4o), Anthropic (Claude Opus 4.5, Claude 4 Sonnet), Google (Gemini 2.5/3.0), and open weights. See the providers table on this page.
Can I bring my own API key (BYOK)?
Yes. BYOK is supported per workspace/project with per-task routing and fallbacks. This gives you full control over model selection and usage while keeping costs transparent.
What data leaves my warehouse?
Only the minimal fields required for planning or generation are sent to the selected model. Conversions and raw PII remain in your warehouse by default. We redact sensitive data before any API calls.
How does model routing work?
Synter routes tasks to the best model for the job. High-stakes media mutations, GAQL waste mining, and CAPI deduplication route natively to Synter Angstrom-32B. General planning can use GPT-4o or Claude, long-context briefs use Claude, and multimodal creative analysis uses Gemini.
Ready to use frontier models for media planning?
See how Synter integrates with Google Ads, Microsoft Ads, LinkedIn, Meta, Reddit, and X.
View Integrations →