multi-model-gateway

Multi-Model Gateway Orchestrator

Route tasks to the best AI model (or multiple models) through AI Gateway v2.

When to use

• "compare models", "ask GPT", "ask Gemini", "cross-model", "consensus"
• Tasks that benefit from a specific model's strengths
• Validation: run same analysis through 2-3 models, compare results
• When one model is rate-limited, route to another

Available Models

Claude (native in Claude Code, also via gateway)

Model	ID	Best for
Opus 4.6	claude-opus-4-6	Deep reasoning, architecture, complex analysis
Sonnet 4.5	claude-sonnet-4-5-20250929	Most tasks, code gen, balanced
Haiku 4.5	claude-haiku-4-5-20251001	Fast classification, simple tasks

OpenAI (via gateway)

Model	ID	Best for
GPT-5.4	gpt-5.4	Latest flagship, best quality
GPT-5.4 Pro	gpt-5.4-pro	Pro tier, higher limits
GPT-5.1 Codex	gpt-5.1-codex	Code generation
GPT-4.1	gpt-4.1	Balanced quality/speed
GPT-4.1 Mini	gpt-4.1-mini	Fast, cheap
o4-mini	o4-mini	Reasoning, math (latest)
o3-pro	o3-pro	Deep reasoning (premium)

Gemini (via gateway)

Model	ID	Best for
Gemini 3.1 Pro	gemini-3.1-pro-preview	Latest flagship
Gemini 3 Flash	gemini-3-flash-preview	Fast, good quality
Gemini 2.5 Pro	gemini-2.5-pro	Stable, 2M context
Gemini 2.5 Flash	gemini-2.5-flash	Fast, long context
Deep Research	deep-research-pro-preview	In-depth research

Gateway Access

# Local gateway (start once):
cd ${WORKSPACE}/projects/ai-gateway && GATEWAY_CONFIG=./config.local.yaml uvicorn app.main:app --port YOUR_PORT &

# Call any model:
curl -s http://localhost:GATEWAY_PORT/v1/messages \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL_NAME",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "PROMPT"}]
  }'

Orchestration Patterns

Pattern 1: Best Model Selection

Analyze the task and pick the optimal model:

Task type	Recommended model
Complex architecture	Claude Opus 4.6 (native)
Code generation	Claude Sonnet 4.5 (native)
Quick classification	Claude Haiku 4.5 (native)
Alternative perspective	GPT-5.4 (via gateway)
Code gen (OpenAI)	gpt-5.1-codex (via gateway)
Large document analysis	Gemini 3.1 Pro (via gateway)
Deep research	deep-research-pro-preview (via gateway)
Math/logic problems	o4-mini (via gateway)
Deep reasoning	o3-pro (via gateway)

Pattern 2: Cross-Model Consensus

Run the same prompt through 2-3 models, then synthesize:

1. Send to Claude (native) → result_claude
2. Send to GPT-5.4 (gateway) → result_gpt
3. Send to Gemini 3.1 Pro (gateway) → result_gemini
4. Compare and synthesize best answer

Use dispatching-parallel-agents skill to run agents in parallel.

Pattern 3: Chain of Models

Each model does what it's best at:

1. Gemini 3.1 Pro → summarize large input (2M context)
2. Claude Opus 4.6 → deep analysis of summary
3. GPT-5.4 → format as structured JSON output

Agents

Agent	Description
gpt-agent	Calls GPT models via gateway
gemini-agent	Calls Gemini models via gateway

Claude models are called natively (no gateway needed for Claude Code).

Dispatching Example

# Ask GPT for a second opinion:
Agent(subagent_type="gpt-agent", prompt="Analyze this architecture: ...")

# Ask Gemini to process a large doc:
Agent(subagent_type="gemini-agent", prompt="Summarize this 100-page doc: ...")

# Parallel consensus (both at once):
Agent(subagent_type="gpt-agent", prompt="...", run_in_background=true)
Agent(subagent_type="gemini-agent", prompt="...", run_in_background=true)

Response Format

When presenting multi-model results:

## Cross-Model Analysis

### Claude Opus 4.6 (native)
[result]

### GPT-5.4 (via Gateway)
[result]

### Gemini 3.1 Pro (via Gateway)
[result]

### Synthesis
[combined best answer with reasoning]

Admin & Monitoring

• Dashboard: https://gateway.your-monitoring-domain.com/admin (admin / YOUR_GATEWAY_PASSWORD)
• Stats API: curl -u admin:YOUR_GATEWAY_PASSWORD https://gateway.your-monitoring-domain.com/admin/stats?hours=24
• Logs API: curl -u admin:YOUR_GATEWAY_PASSWORD https://gateway.your-monitoring-domain.com/admin/logs?limit=50