The Minimum Viable AI Tool

I got access to Claude at work and watched 7% of my monthly allocation disappear in one hour.

At home, the same behavior would barely register. My personal AI subscriptions encourage exploration. I can run several agents, try the same problem three ways, throw away weak outputs, and learn by spending inference.

The meter changed the moment the work moved into an enterprise account.

That is not just a pricing difference. It is an operating-model difference.

Personal AI plans teach fluency by making experimentation feel abundant. Enterprise AI usage has budget, governance, data policy, auditability, reliability, and shared capacity attached to every call. The output has to belong in a real workflow, not just an interesting conversation.

The next AI skill is not prompting.

It is routing.

The strongest model is not the default answer

For the last few years, power users learned a simple reflex: choose the most capable model available and let it think.

That made sense while the main constraint was capability. It makes less sense when models can plan, search, write files, use a browser, call tools, run terminals, and keep working for long stretches.

An agentic task does not spend inference once. It spends inference across a sequence.

The cost of choosing the largest model is multiplied by every search, retry, tool result, verification pass, and handoff in the workflow. The cost of choosing a model that is too small shows up somewhere else: bad routing, missed context, incomplete work, and human cleanup.

The goal is not cheap AI.

The goal is the minimum viable AI tool: the smallest system that can complete the job reliably without handing the savings back as review debt.

Model tiers are becoming product decisions

OpenAI made this logic unusually explicit with the GPT-5.6 family. Sol is the flagship tier for complex professional work. Terra balances capability, speed, and cost. Luna is the fastest and least expensive tier.

The names matter less than the structure. Capability is no longer a single ladder where everyone should stand on the highest rung. It is a routing surface.

The same shift is visible at Anthropic. Claude Sonnet 5 is positioned as an agentic execution model that can plan, use browsers and terminals, and finish multi-step work at a lower cost than the top Opus tier.

This is where model selection starts to look like product architecture.

Shape of the work	Starting point	Why
Ambiguous strategy, difficult trade-offs, high review cost	Sol or another frontier reasoning tier	Judgment and synthesis justify the spend
Multi-step professional work with tools	Terra, Sonnet, or an equivalent balanced tier	Capability, follow-through, and cost all matter
Extraction, formatting, classification, and fast fan-out	Luna or another fast tier	The task is bounded and easy to verify
Sustained codebase work	Codex, Claude Code, or a repo-aware coding agent	Workflow context and tool control beat chat alone
Sensitive internal material	An approved enterprise workspace	Policy and access boundaries outrank convenience
Open-ended personal exploration	A personal general-purpose plan	Cheap experimentation builds judgment through reps

This is not a permanent ranking. Models change too quickly for that.

It is a durable way to ask the question.

Personal subscriptions are part of the adoption funnel

The generous feel of a personal subscription is useful to the AI companies. It creates habit, product feedback, and people who know what the tools can do before a procurement process begins.

I see this as part of the enterprise adoption funnel.

The power user is not only a consumer customer. They become the person inside a company who can recognize a good use case, explain the limits, and show colleagues a workflow that already works.

That fluency comes from waste.

You try prompts that go nowhere. You compare tools. You leave an agent running and discover that your instructions were underspecified. You learn which tasks need the strongest reasoning and which ones only looked difficult because the process was a mess.

An enterprise cannot remove all of that experimentation and still expect mature adoption. But it also cannot fund every production workflow as if it were a personal sandbox.

The answer is cost literacy, not austerity.

Route by the cost of being wrong

Raw complexity is only one input. The more useful routing question is: what happens if this result is wrong?

A quick rewrite is cheap to inspect. A strategic recommendation built from six systems is not. A formatting mistake is visible. A plausible but incorrect synthesis can survive several meetings.

That changes the model choice.

Use more capability when:

The problem is ambiguous.
Several sources disagree.
The output will drive a meaningful decision.
Tool use requires judgment between steps.
Human review would be slow or difficult.

Use a lighter tier when:

The input and output are tightly bounded.
Success can be checked automatically.
The task repeats at high volume.
A failure is obvious and cheap to rerun.
The model is transforming information rather than deciding what it means.

The minimum viable tool is not always the smallest model. Sometimes a stronger model is cheaper because it finishes in one pass and checks its own work. Sometimes a fast model is the honest choice because the task never deserved deep reasoning.

Tools changed the shape of the stack

The old AI stack was a pile of point solutions. One tool for search. One for writing. One for automation. One for coding. Another layer to connect them.

That pile is shrinking.

General-purpose systems now reach into files, browsers, terminals, apps, and structured workflows. OpenAI describes current agents as a combination of a trigger, a process, tools, and guardrails. That framing is more useful than a directory of hundreds of hypothetical use cases.

The durable unit is the workflow:

What starts the work?
What process should the agent follow?
Which tools and context can it use?
Where must it stop or ask for approval?
How will the result be checked?

The model is one component inside that system.

This is why I am using fewer standalone AI tools than I did a year ago. ChatGPT and Claude absorbed many of the gaps those products once filled. A specialized product still earns a place when it owns a workflow the general systems do not. A thin wrapper around a capability they already have is living on borrowed time.

Teach routing, not restraint

The wrong enterprise response to visible inference cost is to make people afraid to use AI.

Timid users do not become efficient users. They become shallow users. They save the allowance by avoiding the experiments that would teach them where the value is.

Teams need a simple routing language:

Match model strength to the value and ambiguity of the task.
Keep sensitive work inside approved systems.
Prefer tools that already sit near the source context.
Reserve frontier inference for work where a weak answer creates expensive review.
Use fast tiers for bounded steps that can be checked.
Measure the finished workflow, not the cost of one prompt.

The best model is the one that gets the job to a trustworthy end state at the lowest total cost.

AI literacy without cost literacy is unfinished literacy.