model-select

name	model-select
description	Select the best LLM model for a task from 33 available models routed through BitRouter. Use this skill when choosing which model to use for a specific task, optimizing cost vs quality, switching models mid-workflow, or creating task-specific routes. Triggers on: model selection, choose model, pick model, which model, cost optimization, task routing, model switching.
license	MIT
metadata	{"author":"BitRouterAI","version":"0.1.0"}

Model Selection — Task-Aware Routing via BitRouter

You have access to 33 LLM models through BitRouter's OpenRouter upstream. Rather than using the auto load-balancer for every task, actively select the right model based on the task at hand. This saves cost and improves quality.

---

Decision Framework

Before starting a task, assess it along two axes:

1. Complexity: How hard is the reasoning / how many steps?
2. Cost sensitivity: Is this a one-shot or a high-volume repeated task?

Then pick the lowest tier that can handle the job well.

---

Model Tiers

Tier 1 — Frontier (complex reasoning, architecture, research)

Use for: multi-file refactors, system design, novel algorithm design, complex debugging, long-horizon planning, tasks where correctness is critical.

Model	In $/M	Out $/M	Context	Max Out	Vision	Notes
anthropic/claude-opus-4.6	5.00	25.00	1M	128K	yes	Best for long agentic workflows, coding
anthropic/claude-opus-4.5	5.00	25.00	200K	64K	yes	Strong reasoning, file input
openai/gpt-5.4	2.50	15.00	1M	128K	yes	Unified Codex+GPT, massive context
google/gemini-3.1-pro-preview	2.00	12.00	1M	65K	yes	Best SWE-bench, audio/video input
google/gemini-3-pro-preview	2.00	12.00	1M	65K	yes	Multimodal frontier
google/gemini-2.5-pro	1.25	10.00	1M	65K	yes	Thinking model, math/science

Cost guidance: $2–25/M output. Use only when task demands it.

Tier 2 — Strong (general coding, analysis, most dev work)

Use for: standard feature implementation, code review, writing tests, data analysis, technical writing, moderate-complexity debugging.

Model	In $/M	Out $/M	Context	Max Out	Vision	Notes
anthropic/claude-sonnet-4.6	3.00	15.00	1M	128K	yes	Best all-round dev model
anthropic/claude-sonnet-4.5	3.00	15.00	1M	64K	yes	Strong coding, file input
qwen/qwen3-max-thinking	0.78	3.90	262K	32K	no	Deep chain-of-thought reasoning
z-ai/glm-5	0.72	2.30	80K	131K	no	Systems design, agent workflows
moonshotai/kimi-k2.5	0.45	2.20	262K	65K	yes	Visual coding, agent swarms
mistralai/mistral-large-2512	0.50	1.50	262K	—	yes	MoE 675B, good multilingual
qwen/qwen3.5-397b-a17b	0.39	2.34	262K	65K	yes	Largest Qwen MoE, vision+video

Cost guidance: $0.40–15/M output. Default tier for most work.

Tier 3 — Fast (simple tasks, high throughput, cost-sensitive)

Use for: formatting, translation, summarization, triage, simple Q&A, boilerplate generation, linting, commit messages, quick lookups.

Model	In $/M	Out $/M	Context	Max Out	Vision	Notes
anthropic/claude-haiku-4.5	1.00	5.00	200K	64K	yes	Fast + capable, great value
google/gemini-3-flash-preview	0.50	3.00	1M	65K	yes	Fast thinking model, multimodal
mistralai/devstral-2512	0.40	2.00	262K	—	no	123B dense, code specialist
google/gemini-2.5-flash	0.30	2.50	1M	65K	yes	Workhorse, audio/video support
openai/gpt-5-mini	0.25	2.00	400K	128K	yes	Compact GPT-5, good reasoning
qwen/qwen3.5-plus-02-15	0.26	1.56	1M	65K	yes	1M context, vision+video
qwen/qwen3.5-122b-a10b	0.26	2.08	262K	65K	yes	MoE, good all-round
minimax/minimax-m2.5	0.20	1.17	196K	65K	no	Real-world productivity
x-ai/grok-4.1-fast	0.20	0.50	2M	30K	yes	2M context, great tool calling
qwen/qwen3.5-27b	0.20	1.56	262K	65K	yes	Dense, fast, vision+video
qwen/qwen3.5-35b-a3b	0.16	1.30	262K	65K	yes	MoE, linear attention, fast
openai/gpt-4o-mini	0.15	0.60	128K	16K	yes	Cheap, reliable baseline
deepseek/deepseek-chat	0.32	0.89	163K	163K	no	DeepSeek V3, huge output window
minimax/minimax-m2.1	0.27	0.95

Cost guidance: $0.13–5/M output. Use for bulk work and simple tasks.

Tier 4 — Ultra-cheap (maximum throughput, minimal cost)

Use for: batch processing, classification, extraction, format conversion, simple parsing, test data generation, any task where speed >> quality.

Model	In $/M	Out $/M	Context	Max Out	Vision	Notes
openai/gpt-5-nano	0.05	0.40	400K	128K	yes	Smallest GPT-5, ultra-fast
xiaomi/mimo-v2-flash	0.09	0.29	262K	65K	no	MoE 309B/15B active, cheapest
stepfun/step-3.5-flash	0.10	0.30	256K	256K	no	Cheap MoE, massive output
qwen/qwen3-coder-next	0.12	0.75	262K	65K	no	Code-specialist, open-weight
deepseek/deepseek-v3.2	0.26	0.38	163K	—	no	Efficient reasoning, tool use

Cost guidance: $0.05–0.75/M output. 10–100x cheaper than frontier.

Specialist Models

These excel at specific tasks regardless of their tier placement:

Task	Best Models	Why
Code generation	qwen3-coder-next, devstral-2512, deepseek-v3.2	Purpose-built for code
Code review / refactor	claude-sonnet-4.6, devstral-2512, gpt-5.4	Strong at understanding intent
Multimodal (image)	gemini-3-pro-preview, claude-sonnet-4.6, kimi-k2.5	Native vision
Audio/Video	gemini-2.5-flash, gemini-3-flash-preview, gemini-3-pro-preview	Only Gemini supports A/V
Huge context (>500K)	grok-4.1-fast (2M), gpt-5.4 (1M), gemini-* (1M), sonnet-4.6 (1M)	Large context windows
Maximum output	stepfun/step-3.5-flash (256K), deepseek-chat (163K), gpt-5.4/nano (128K)	For long-form generation
Multilingual	mistral-large-2512, qwen3.5-*, kimi-k2.5	Strong non-English
Math/Science	gemini-2.5-pro, qwen3-max-thinking, claude-opus-4.6	Thinking/reasoning modes
Agentic tool use	grok-4.1-fast, claude-sonnet-4.6, glm-5, glm-4.5-air	Optimized for tool calling loops

---

How to Switch Models

Use the BitRouter CLI to create task-specific routes:

# Set a named route for the current task type
bitrouter route add my-task openrouter:anthropic/claude-sonnet-4.6

# Create a route with failover (tries in order)
bitrouter route add code-review \
  openrouter:qwen/qwen3-coder-next \
  openrouter:mistralai/devstral-2512 \
  --strategy priority

# Create a load-balanced pool for a task type
bitrouter route add batch-work \
  openrouter:openai/gpt-5-nano \
  openrouter:xiaomi/mimo-v2-flash \
  openrouter:stepfun/step-3.5-flash \
  --strategy load_balance

# Check current routes
bitrouter route list

# Remove a route when done
bitrouter route rm my-task

Then use the route name as the model in your API calls:

curl http://127.0.0.1:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "code-review", "messages": [...]}'

---

Quick Selection Cheatsheet

Task	Recommended Model	Cost
Architecture / system design	claude-opus-4.6	$$$$$
Complex multi-file refactor	gpt-5.4 or gemini-3.1-pro	$$$$
Standard feature / bug fix	claude-sonnet-4.6	$$$
Code review	devstral-2512 or qwen3-coder-next	$$
Write tests	claude-haiku-4.5 or gpt-5-mini	$$
Simple edit / formatting	gpt-5-nano or mimo-v2-flash	$
Summarize / translate	gpt-5-nano or step-3.5-flash	$
Batch classify / extract	gpt-5-nano or deepseek-v3.2	$
Image analysis	gemini-2.5-flash or claude-sonnet-4.6	$$-$$$
Audio/video processing	gemini-2.5-flash or gemini-3-flash	$$
Long document analysis (>500K)	grok-4.1-fast (2M ctx, $0.50/M)	$
Research / deep reasoning	qwen3-max-thinking or gemini-2.5-pro	$$-$$$$

---

Cost Comparison Examples

For a typical 2K input / 1K output task:

Tier	Model	Cost per call
Ultra-cheap	gpt-5-nano	$0.0005
Fast	gemini-2.5-flash	$0.003
Strong	claude-sonnet-4.6	$0.021
Frontier	claude-opus-4.6	$0.035

A 70x cost difference between cheapest and most expensive. Choose wisely.