10-Tier Multi-Dimensional LLM Hierarchy
Tharior Remedai routes tasks across a 10-tier hierarchy evaluated on Cost, Performance, Coding Benchmarks (HumanEval & SWE-bench), Knowledge vs Reasoning biases, and Functional Specializations.
1. Multi-Dimensional Specialization Matrix
| Tier # | Functional Specialization | Knowledge vs Reasoning Profile | Representative Models | Benchmarks | Prompt $/1M |
|---|---|---|---|---|---|
| Tier 1 | Documentation & Syntax Lint | Knowledge-Biased (Grammars & Docstrings) | gemini-1.5-flash-8b, qwen-2.5-coder-7b |
HumanEval: 74.2% Ctx: 1M tokens |
$0.0375 |
| Tier 2 | Ultra-Cheap Fast Remediator | Knowledge-Biased (Broad Context & Markdown) | deepseek-chat (V3), gemini-2.0-flash-lite |
HumanEval: 82.6% Ctx: 1M tokens |
$0.1400 |
| Tier 3 | Economical Code Specialist | Balanced (Code Syntax + Assertions) | gpt-4o-mini, qwen-2.5-coder-32b |
SWE-bench: 33.2% HumanEval: 86.8% |
$0.1500 |
| Tier 4 | Mid-Tier Generalist | Balanced (Fast Full-Stack Handlers) | claude-3-5-haiku, codestral-2501 |
SWE-bench: 40.6% HumanEval: 88.9% |
$0.8000 |
| Tier 5 | Fast Structured Reasoner | Reasoning-Biased (Chain-of-Thought / Logic) | o3-mini (low), deepseek-r1-70b |
SWE-bench: 44.8% HumanEval: 92.1% |
$1.1000 |
| Tier 6 | Core Engineering Workhorse | Balanced High Capability (Full Modules) | openai/gpt-4o, anthropic/claude-sonnet-4.5 |
SWE-bench: 52.4% HumanEval: 94.2% |
$2.5000 |
| Tier 7 | Deep Concurrency Reasoner | Ultra-Deep Reasoning (Deadlocks & Races) | openai/o3-mini (high), deepseek/deepseek-r1 |
SWE-bench: 56.1% HumanEval: 96.8% |
$4.8000 |
| Tier 8 | Senior Architect | Frontier Balanced (Extended Thinking + Repo AST) | anthropic/claude-sonnet-4.6, meta-llama/llama-3.3-70b-instruct |
SWE-bench: 71.8% HumanEval: 97.6% |
$9.0000 |
| Tier 9 | Frontier Cognitive Synthesis | Frontier Synthesis (Maximal Cognitive Depth) | anthropic/claude-sonnet-5, openai/o1, openai/o3 |
SWE-bench: 75.4% HumanEval: 98.6% |
$18.0000 |
| Tier 10 | Elite Multi-Model Consensus | Consensus Ensemble (Triple-Engine Verification) | claude-sonnet-5 + o3 + gemini-2.5-pro |
SWE-bench: 81.2% Consensus Voting |
$45.0000 |
2. Knowledge vs Reasoning: Real-World Use Cases
Different engineering challenges require different cognitive strengths:
- Knowledge & Context Heavy (Tiers 1, 2, 4): Ingesting massive API documentation, library changelogs, or generating comprehensive markdown release notes. Models like
gemini-1.5-flash-8bandgemini-2.0-flash-liteexcel here with 1M+ token context windows at fractional cents. - Syntax & Code Specialized (Tiers 3, 4): Fast regex translation, isolated function repair, and PyTest assertions. Models like
qwen-2.5-coder-32bandcodestral-2501provide high code accuracy without reasoning overhead. - Reasoning & Logic Heavy (Tiers 5, 7): Unwinding complex state machines, race conditions, async mutex deadlocks, and backoff algorithms. Reasoning models like
o3-minianddeepseek-r1leverage structured thought traces to solve non-obvious execution flaws. - Architecture & Consensus (Tiers 8, 10): Zero-downtime database migrations, microservice interface refactoring, and financial kernels.
claude-sonnet-5with hybrid extended thinking and multi-model consensus voting ensure absolute correctness.
3. Dynamic OpenRouter Ingestion, Live Pricing Sync & Custom Models
Tharior Remedai eliminates rigid, hardcoded model identifiers in favor of a dynamic, live model catalog connected directly to the
OpenRouter API (https://openrouter.ai/api/v1/models).
- No Hardcoded Model IDs: Models are pulled live from OpenRouter and evaluated dynamically on pricing, context length, benchmarks, and reasoning modalities.
- Dynamic Custom Model Management: Add custom models (e.g. fine-tuned internal checkpoints, regional endpoints) to any tier or remove obsolete models in real-time via the Model Catalog Desk or REST API (
POST /api/v1/models/catalog,DELETE /api/v1/models/catalog). - Configurable Pricing Sync: Automated catalog synchronization intervals are persisted in
app-config.yamlwith options for Hourly (Every 1 Hour), Every 6 Hours, Daily (Every 24 Hours), or Weekly (Every 7 Days). - Multi-Model Candidate Failover & Circuit Breaker: The router iterates across up to 5 candidate models in the selected tier. Non-infrastructure 4xx model issues (such as temporary upstream availability or unsupported parameters) are filtered from tripping the global circuit breaker, ensuring seamless automatic failover to the next candidate model.
4. Customer Tier Overrides (±1 to ±2 Tiers Shift)
While Tharior Remedai applies intelligent baseline tiering, customers can customize model assignments:
- Allowed Model Whitelist: Restrict which models your organization permits.
- Controlled Tier Shifts (±1 or ±2 Tiers): Shift individual models up or down by 1 or 2 tiers to balance cost vs capability preferences. Shifts exceeding ±2 tiers are safely clamped to maintain system stability.
- Per-Tier Model Selection: Explicitly choose representative models per tier or fallback to system defaults.
5. Multimodal Tiering (Audio, Video, Image & Assets)
Beyond text code remediation, Tharior Remedai provides distinct multimodal tiers for rich technical assets:
| Modality | Tier | Representative Engines | Unit Cost | Target Capabilities |
|---|---|---|---|---|
| Audio | Audio Tier 1-2 | whisper-large-v3, elevenlabs/multilingual-v2 |
$0.006 - $0.030 / min | Voice narration, real-time command streaming, meeting transcripts |
| Video | Video Tier 1-2 | gemini-2.0-flash, runway/gen-3-alpha, kling-v1.5 |
$0.002 - $0.080 / sec | UI video test regression audits, feature walkthrough videos |
| Image | Image Tier 1-2 | flux-1-schnell, imagen-3-fast, dall-e-3 |
$0.003 - $0.040 / img | System architecture diagrams, UI mockups, slide decks |
6. Semantic Caching & Internet Search Plugin
To prevent redundant LLM invocations and slash token expenditure:
- Semantic Caching: Uses term vector cosine similarity (> 0.92) to match incoming user queries with previously verified remediations, returning instant solutions with 0 upstream API cost.
- Internet Search Plugin: Modular, toggleable web search tool allowing agents to query live documentation, API specs, and CVE advisories with concise summaries.
7. Automation Playbooks & Sentinel PR Review Agent
Enterprise automation playbooks streamline the end-to-end bug remediation lifecycle:
- Issue & Story Listener: Ingests Jira, GitHub, or Linear webhook events for stories assigned directly to the agent.
- Auto-Remediation & Commenting: Fixes bugs, executes sandbox tests, and automatically comments structured resolution details on the story.
- PR Review Agent & Auto-Merge: Sentinel Review Agent audits PR diffs, assigns quality scores (out of 100), checks SAST security, and conditionally auto-merges when all criteria are satisfied.
- Token Budget & Output Controls: Hardened system prompts prevent verbose internal monologue dumping, keeping token usage strictly bounded.