AI Models · Tutorial 05
Frontier vs Open-Weight Models: A Decision Guide
Combine L01 through L04 into five questions that choose an AI model for a task and show when to switch tiers mid-project.
0 of 5 complete
Last tested and updated: June 2026
Pick the model by the loop it will run in, not by its tier label. Five questions narrow the choice for any workload.
The task changes the answer
The choice comes down to your task type.
Take a $5/mo Kimi K2.7 and write a thank-you email. It works. The output matches a $200/mo Claude Max.
Now take Kimi on a harder job: architect a multi-tenant Postgres schema with row-level security. It can do it. You’ll spend three rounds catching edge cases. Opus catches them in one.
That gap is exactly what tier choice captures.
Choose by loop
The channel’s guidance across L02, L03, and L04 is consistent: pick by loop. A capable model inside the wrong loop is still the wrong tool. The same principle used to choose a harness in Hermes L01 also applies to the model tier.
Three tiers, three jobs:
- Tier 1: Frontier (closed, expensive, Mythos-class). Opus 4.8, Fable 5, GPT 5.5, Gemini 3.1 Pro. Hard one-shots: novel creative work, ambiguous specs, architecture decisions. $80 to $300/month.
- Tier 2: Cost-efficient (open-weight or cheap closed). Kimi K2.7, DeepSeek V4 Pro, Qwen 3.7 Max/Plus, MiniMax M3, GLM 5.2, Sonnet 4.5. Eighty percent of daily work: UI, vision, code review, summarisation, agent workflows. $5 to $40/month.
- Tier 3: Niche. Sakana Fugu, Mistral Codestral, DeepSeek Coder. One workload, done exceptionally well. Don’t generalise out of a Tier 3 model: pick the niche on purpose, not by accident.
Tier 2 by default. Tier 1 for hard work. Tier 3 only when the niche is your workload. That move works for almost everyone, almost every day.
Five questions before you route
Use this 5-question framework to pick the right tier for any workload. Answer in order.
Q1: What loop does the workload run in? Five loops account for most work in 2026:
| Loop | Default tier | Default model |
|---|---|---|
| One-shot creative / novel spec | Tier 1 | Opus 4.8 or Fable 5 |
| Daily Q&A, summarisation | Tier 2 | Sonnet 4.5 or Kimi K2.7 |
| Coding all day (production) | Tiered | Sonnet default, Opus on hard tasks |
| Long-horizon agent (hours) | Tier 2 + agent harness | Qwen 3.7 Max or MiniMax M3 |
| Multimodal (vision / UI / images) | Tier 2 specialist | MiniMax M3 or Kimi K2.7 |
The single biggest beginner mistake: treating loop as a synonym for task name.
A coding session is a loop (interactive, verified, multi-turn). A one-shot creative prompt is a different loop (single turn, no test feedback).
Match the loop first. The model falls out of that.
Q2: What’s your monthly budget? Under $20/mo: Tier 2 only. $20 to $50/mo: Tier 2 with occasional Tier 1 escalations. $50 to $200/mo: Tier 2 default plus Tier 1 routing on the hardest 10 to 20% of tasks.
Over $200/mo: Tier 1 is fine for daily use. Audit whether you consume that capability.
Q3: How reproducible is the workload? Daily research summaries, recurring cron jobs, repeated Q&A: Tier 2, full stop.
Variable inputs push you toward Tier 1 or tiered routing. One-off creative work for an actual deadline is the one case where Tier 1 wins.
Q4: How time-sensitive is it? Days to weeks: Tier 2. Hours: Tier 2 if Qwen Max or DeepSeek Flash is fast enough. Minutes (a one-shot deadline): Tier 1.
Mythos-class intelligence matters when you don’t get a second attempt.
Q5: How much do you trust the harness? This is the question people forget.
A trusted harness (Claude Code for code, Hermes for agentic loops) catches Tier 2’s mistakes through verification.
A new or experimental harness cannot catch them: so you pay the Tier 1 premium as insurance. Tiered routing is in Hermes L06.
When to switch tiers mid-project
Tier 2 → Tier 1: three or more retries with the same prompt. The task escalates from routine to novel. Or production-critical work where reliability outweighs cost.
Tier 1 → Tier 2: you realise Tier 1 isn’t doing anything Tier 2 can’t. Workload shifts from novel to routine. Or you’re spending $200+/mo on marginal tasks.
Within Tier 2: workload changes shape (coding → UI → vision). A new release ships in another family clearly better for your workload. Or a cost-tier special lands (Qwen Plus 40% off, Kimi flash promo).
Worked examples
Example 1: Daily research cron (Hermes, 7am). Cron summarises three AI news stories and posts to Discord. Loop: reproducible, batch, low-stakes. Default: Tier 2 (DeepSeek V4 Flash or Kimi K2.7). Escalate only if summaries miss important stories three days running. Cost: $3 to $8/month.
Example 2: Production feature (Claude Code). Multi-tenant billing flow with row-level security. Loop: interactive coding, test-verified. Default: tiered routing: Sonnet 4.5 for routine 80%, Opus 4.8 for the hard sub-tasks. Cost: $30 to $80/month.
Example 3: One-shot creative brief. Pitch deck due tomorrow. Loop: single-turn, high-stakes. Default: Tier 1, Opus 4.8 or Fable 5. That one attempt needs the best model you can buy.
Example 4: Long-horizon agent (Hermes over Qwen Max). Research project running four hours overnight, reading 50 PDFs. Loop: long-running, verification-by-inference. Default: Tier 2 + agent harness, Qwen 3.7 Max or MiniMax M3 wrapped in Hermes. Cost: $10 to $25/month.
Write your routing rule
The exercise
Open a text file on your computer. Write down the answer to this in your own words, no copy-paste:
“Pick three AI workloads you did this week. For each: name the loop (one-shot / daily Q&A / coding / long-horizon / multimodal), name the default tier (1 / 2 / tiered), and name one specific model. Then write down the trigger that should make you switch tiers.”
A good answer covers three workloads, three tiers, three models, and three triggers.
To fork the framework for your own workloads, replace the five loops in Q1 with the five you actually run. Then replace the budgets in Q2 with your real monthly spend.
Print the result and tape it above your monitor.
What’s next
Hermes L06: Model Routing and Cost: tiered routing in a real harness.
Hermes L09: Hermes vs Alternatives: Claude Code vs Hermes vs Mavis.