Stop lighting cigars with a flamethrower

A developer asks the most powerful frontier model to rewrite a button label.

That is the AI equivalent of lighting a cigar with a flamethrower. It works. The cigar is lit. The table is also on fire, the curtains are nervous, and nobody can remember why this needed military equipment.

The same thing happens with document classification, small text edits, simple extraction, and the kind of code change where the answer is visible before the model starts thinking. An enormous model handles those tasks. So does a smaller one. The difference is that the enormous model drags along more latency, more cost, more pressure on rate limits and sometimes more compliance machinery than the task has any use for.

The sharp rule is simple: use the smallest model that handles the task reliably.

Not the smallest model in theory. Not the cheapest model on a price list. The smallest model that passes the actual test in front of you.

The Fable lesson is not about fear

Fable 5 and Mythos 5 made this visible because they were unusually powerful and unusually gated. Anthropic launched both on 9 June 2026. On 12 June the company said an export control directive from the US government forced it to cut off access to both models for every customer, while other Anthropic models stayed available.

That is a dramatic trigger, but the daily lesson is quieter.

Fable was not just one more model on a tidy ladder. Anthropic described it as a Mythos class model, above Opus, built for long and complex work. They said Fable could work more independently, with fewer interruptions, and that Mythos 5 was the same underlying model with some guardrails removed for selected users.

That matters. It means the top tier is no longer just a smarter autocomplete. It is capability plus access rules, retention rules, safety gates and possible interruptions from someone outside the vendor. The list starts to look less like a tool menu and more like airport security.

Using that tier to polish marketing copy is not brave. It is careless.

Most work has a ceiling

Anthropic's own guide to the Claude models says Haiku is for fast answers, extraction, categorisation and simple summaries. Sonnet is positioned as the workhorse for coding, writing, analysis, research and multi step flows. Opus is reserved for deep research and complex reasoning, especially when Sonnet gets stuck. Fable is described as the heaviest option for the biggest and most important long jobs.

That is not a slogan about saving money. It is a control system.

Keep an escalation ladder

A sensible ladder starts low.

For bounded, repeatable and privacy sensitive work, start with a local or small model if it passes the evaluation. Not because local AI is morally superior. Because some jobs are small, private, repetitive and best kept near the machine.

For routine cloud work, use Haiku or Sonnet. Plenty of routine cloud work stays there. Fast enough, capable enough, less drama.

Escalate when the task earns it. A vague product refactor across a large codebase may earn it. A long running agent that has to plan, edit, run tests, recover from errors and hold the whole goal in view earns it. Security work in a messy repository may earn it. A migration that would take a team weeks may earn it.

The examples from Anthropic's Fable launch point exactly there: long independent software development, large changes across codebases, deeper document reasoning, hard scientific work and specialised use in cyber or biology. Those are not button labels. Those are the jobs where more capability can change the outcome rather than decorate the answer.

The mistake runs both ways.

Using the top tier for everything is wasteful and brittle. Throwing it away because most work does not need it is exactly as wrong. The top model should stay as an escalation tier, with clear rules for when it gets called.

You do not light a cigar with a flamethrower. You also do not throw away the flamethrower because you mostly smoke cigars.