feat(proxy): configuring smart-router for improved costs handling

[fab-siciliano]

Add new models and refactors the smart-router model to improve costs performances and provide the best options for different tasks

Details

Main recommendations are:

1. Keep SIMPLE cheap
2. Make MEDIUM your real default workhorse
3. Keep COMPLEX and REASONING meaningfully different
4. Do not make the fallback/default your cheapest model
5. Only use the most expensive model in REASONING

Cost picture from the repo metadata

From model_prices_and_context_window.json, these are the most relevant candidates:

Model	Provider	Input cost / 1M tokens	Output cost / 1M tokens	Notes
gpt-5.4-mini	OpenAI	$0.75	$4.50	cheap simple-tier candidate
gpt-5.4	OpenAI	$2.50	$15.00	solid, but not the best value middle tier
gpt-5.5	OpenAI	$5.00	$30.00	premium OpenAI tier
gpt-5.3-codex	OpenAI	$1.75	$14.00	older coding-focused option
o4-mini	OpenAI	$1.10	$4.40	cheap reasoning-capable option
o3	OpenAI	$2.00	$8.00	strong value reasoning option
o3-mini	OpenAI	$1.10	$4.40	cheaper reasoning option
claude-haiku-4.5	Bedrock	$1.00	$5.00	cheap Bedrock Anthropic option
claude-sonnet-5	Bedrock	$2.00	$10.00	best balanced middle-tier candidate
claude-opus-4.8	Bedrock	$5.00	$25.00	premium Bedrock tier
claude-fable-5	Bedrock	$10.00	$50.00	top-end quality, highest spend
qwen3-coder-next	Bedrock	$0.60	$1.44	very cheap code-focused option

EU Bedrock Anthropic costs

Model	Provider	Input cost / 1M tokens	Output cost / 1M tokens
eu.anthropic.claude-sonnet-5	Bedrock EU	$2.20	$11.00
eu.anthropic.claude-opus-4.8	Bedrock EU	$5.50	$27.50
eu.anthropic.claude-fable-5	Bedrock EU	$11.00	$55.00

Possible cleanest setup:

• SIMPLE = gpt-5.4-mini
• MEDIUM = bedrock-anthropic-sonnet-5
• COMPLEX = gpt-5.5
• REASONING = bedrock-anthropic-opus-4.8
• complexity_router_default_model = bedrock-anthropic-sonnet-5

Why this works:

• gpt-5.4-mini is a good cheap front line for greetings, basic questions, short rewrites, simple requests
• Claude Sonnet 5 is the strongest value candidate here for the default tier. It is cheaper than gpt-5.4 and a bettet real” work
• gpt-5.5 gives you a premium OpenAI option for complex tasks
• Claude Opus 4.8 keeps REASONING truly premium without jumping all the way to Fable pricing
• Defaulting to Sonnet 5 is safer than defaulting to Mini

Best quality-first setup

If we care more about quality than spend:

• SIMPLE = gpt-5.4-mini
• MEDIUM = bedrock-anthropic-sonnet-5
• COMPLEX = bedrock-anthropic-opus-4.8
• REASONING = bedrock-anthropic-fable-5
• complexity_router_default_model = bedrock-anthropic-sonnet-5

This is the config that best matches the “good practices” reading of the Artificial Analysis page. The page clearly points to Claude Fable 5 and top OpenAI premium models as top-end intelligence choices. So if you want a real “last tier”, REASONING is where
Fable belongs, not COMPLEX

Best cost-optimized code-heavy setup

If most traffic is engineering/coding and you want to save aggressively:

• SIMPLE = bedrock-qwen3-coder-next
• MEDIUM = bedrock-anthropic-sonnet-5
• COMPLEX = gpt-5.5
• REASONING = bedrock-anthropic-opus-4.8
• complexity_router_default_model = bedrock-anthropic-sonnet-5

I would only do this if your proxy is mostly used for coding tasks. I would not use qwen3-coder-next as SIMPLE if thgeneric business chat, summaries, or non-technical prompts

Type

🆕 New Feature