Choose the Right Model, Optimize Your AI Costs

Your AI agent just spent 1.7x credits on a simple status check. Meanwhile, a complex root cause analysis failed on the cheapest model. Both mistakes come from the same place: one model setting for every kind of work. This post shows how CloudThinker's three tiers let you match the model to the task, and how a blend of them saves 40% or more without losing quality.

Why "just use the best model" breaks

The most common mistake in AI operations isn't picking the wrong tool. It's running the right tool at the wrong power setting, and it usually plays out in two phases.

Phase one: everything on the top tier. Teams default to the most powerful model because it feels safer. A status check, a log lookup and an alert classification all run at the same price as an architecture review. Credits go fast.

Phase two: everything on the bottom tier. The budget tightens, someone downgrades the default, and the complex analyses start failing. The team concludes the cheap model "isn't good enough" and swings back.

Both phases share one structural flaw: a single setting for a workload that isn't uniform. Most operational work is pattern matching with clear instructions. A small share needs real reasoning. Pricing every task like the hardest one, or the easiest one, is wrong for most of them.

The three tiers

CloudThinker gives you three tiers: Light (0.3x), Pro (1.0x) and Ultra (1.7x). The point isn't to pick one; it's knowing when to use each.

01Light 0.3x

Fast and cheap

Lookups, alert triage, status checks, and running pre-built skills.

02Pro 1.0x

Daily driver

Code review, incident investigation, multi-step workflows.

03Ultra 1.7x

Deep reasoning

Complex RCA, architecture decisions, building new skills.

Three tiers, picked per message. Credits scale with the multiplier.

Light (0.3x) is fast and cost-effective for simple tasks: quick Q&A and lookups, running pre-built skills, alert triage and classification, simple script generation, status checks and summaries. It handles 55 to 70% of tasks, meaning anything with clear instructions or a pre-built skill.
Pro (1.0x) balances intelligence and speed. It's the daily driver for cloud operations, code review and generation, incident investigation, infrastructure analysis and multi-step workflows.
Ultra (1.7x) brings maximum reasoning for complex problems: hard root cause analysis, architecture decisions, multi-system debugging, strategic planning and reports, and building the skills that make Light more capable.

Light vs Ultra is a 5.7x cost difference per message, so running a task on Light instead of Ultra saves about 82%.

You can switch tiers mid-conversation. The multiplier applies per message, not per conversation, so you only pay the Ultra premium for the turns that need it.

The pattern that matters: build with Ultra, run on Light

The teams that spend the least don't just route well. They use Ultra to build skills, then run those skills on Light.

01Build

Ultra, once

Deep reasoning works out the procedure at 1.7x, one time.

02Encode

Workspace skill

The steps become a structured, reusable, versioned skill.

03Run

Light, every time

The skill guides the 0.3x model through the same steps on every run.

Build with Ultra, run on Light: about 5x cheaper per execution than reasoning from scratch.

Instead of paying 1.0x to 1.7x every time a model works out a complex workflow from scratch, you pay 1.7x once to encode it as a Workspace Skill, then run it on Light at 0.3x from then on. The structured skill does the guiding the bigger model used to do, so the output stays the same quality at about a fifth of the cost.

The break-even math is short. Each run on Light instead of Pro saves 0.7x. The Ultra build session costs 1.7x. So the build pays for itself on the third run, and every run after that is savings:

build cost:        1.7          (Ultra, once)
saving per run:    1.0 - 0.3 = 0.7   (Pro from scratch vs Light with a skill)
break-even:        1.7 / 0.7 ≈ 2.4  → after 3 runs

The best candidates for a skill are:

Repeatable: runs more than once a week.
Structured: the steps are predictable.
Tool-heavy: specific tools, in a specific order.

Security audits, incident runbooks, deployment checklists and compliance reports all fit. After 20 to 30 skills, the blended cost across all operations drops by 40 to 60%, with higher consistency, because skills encode your best practices.

A worked example: routing a month of work

Here's how routing plays out. Based on patterns across hundreds of CloudThinker workspaces, this is our recommended mapping from task to tier:

Task type	Recommended tier	Why	Share of workload
Alert triage	Light (0.3x)	Pattern matching, no deep reasoning needed	~30%
Run an existing skill	Light (0.3x)	The skill has structured steps; the model follows them	~25%
Code review	Pro (1.0x)	Needs contextual understanding of the codebase	~15%
Incident investigation	Pro (1.0x)	Multi-step reasoning with tool use	~15%
Architecture design	Ultra (1.7x)	Complex trade-off analysis, deep reasoning	~5%
Build a new skill	Ultra (1.7x)	Creates a reusable template for Light to run later	~5%
Root cause analysis	Ultra (1.7x)	Multi-system correlation, deep investigation	~5%

That's roughly 55% of a typical workload on Light, 30% on Pro and 15% on Ultra. Your distribution depends on your team's workflows.

Most teams don't use one tier exclusively, and the savings come from the blend. For 1,000 tasks a month:

Strategy	Credit cost	vs. all Pro
All Ultra (1.7x)	1,700 credits	70% more expensive
All Pro (1.0x)	1,000 credits	Baseline
Smart mix (55% Light, 30% Pro, 15% Ultra)	~720 credits	28% savings
Optimized (70% Light, 20% Pro, 10% Ultra)	~580 credits	42% savings

The formula is simple enough to run on your own numbers: (% Light × 0.3) + (% Pro × 1.0) + (% Ultra × 1.7) = your average credit multiplier. The move from the smart mix to the optimized mix is mostly skills: every workflow you encode shifts a slice of Pro work onto Light.

Why tiered routing beats a single default

	One model for everything	Tiered routing with skills
Simple tasks	Paid at the top-tier rate, or fine on the cheap tier by luck	Light by default, 0.3x
Hard tasks	Fail when the default is cheap	Escalate to Pro or Ultra for the turns that need it
Repeat workflows	Reasoned from scratch every time	Encoded once with Ultra, run on Light
Budget pressure	Forces a blanket downgrade	Shifts the mix, not the quality
Consistency	Varies with the prompt and the model	Skills fix the steps and the output format

The deeper difference is that cost stops being a model choice and becomes a workflow choice. Once a procedure lives in a skill, the intelligence that worked it out is reusable, and the per-run cost is set by the cheapest model that can follow the steps.

Three habits that cut costs

1. Default to Light, escalate when needed. Start on Light. If the task is clearly complex, switch to Pro or Ultra. Most teams find 50 to 70% of tasks never need to escalate.

2. Build skills for repeatable workflows. Every skill built with Ultra and run on Light is a permanent cost reduction. Start with the highest-frequency tasks.

3. Review your tier distribution monthly. If more than 30% of credits go to Ultra, you likely have skill candidates hiding in your usage.

You can do all three from chat:

"Switch to Ultra and turn the incident runbook we just followed into a workspace skill, with numbered steps and a fixed output format."

"Back to Light. Run the weekly security audit skill on the production account and summarize the new findings."

"Show me last month's credit use by tier, and which repeated Pro or Ultra tasks look like good skill candidates."