The Joke About AI Model Names Lands Because the Names Stopped Making Sense

A short satirical post by Sam Wilkinson has been circulating in developer circles, imagining Anthropic's product roadmap as a literary escalation. It starts with "Aphorism" ("One sentence, but it always feels right") and climbs through "Diatribe" ("Sonnet, but angry"), "Mythos" ("Opus, but scary"), and eventually "Overwhelmingly Large Narrative Unit" ("Requires viewing a 'previously on' segment prior to usage"). The closing entry, "Omnibus," promises that "fine-tuning will continue until morale improves."

It reads as throwaway comedy. But the reason it spread is more interesting than the jokes themselves. The bit lands because the real naming situation has become hard to track, and the people best positioned to notice are the engineers who pick a model from a dropdown several times a day.

The actual confusion the joke is pointing at

Anthropic ships Claude in tiers named after literary forms: Haiku for the small, fast, cheap model; Sonnet for the mid-range; Opus for the largest. That part is coherent. Each name signals roughly where the model sits on the cost-and-capability curve, which is the kind of metaphor that actually helps. A haiku is short and a sonnet is longer, so the ordering carries information.

The trouble starts when you add version numbers, dates, and qualifiers on top of the poetic names. A developer reaching for the API has to reason about strings like claude-opus-4-8, claude-sonnet-4-6, and claude-haiku-4-5-20251001 at the same time as the friendly product labels. The satire's "Cinematic Universe (Director's Cut)," defined as "Same answer, 42% more tokens," is funny precisely because token-count inflation across model generations is a real cost concern that teams budget around.

Anthropic is not alone here. OpenAI spent much of the past two years fielding complaints about its own model naming, where GPT-4, GPT-4 Turbo, GPT-4o, and a parade of o-prefixed reasoning models left users unsure which one was newest or strongest. Sam Altman publicly acknowledged the mess. Google's Gemini line has shuffled through Pro, Flash, Ultra, and assorted version suffixes. The pattern repeats across the industry: a clean naming idea at launch, then entropy as releases pile up faster than anyone planned a taxonomy for.

Why the literary frame invites this kind of mockery

There is a structural reason Anthropic's scheme in particular gets satirized this way. When you name products after literary forms, you invite people to keep going. Poems, essays, and epics form an obvious size ladder, so it costs a reader nothing to imagine "Treatise" sitting above "Opus" or "Saga" sprawling past "Fable." The joke writes itself, which is both a compliment to the original metaphor and a sign of its limits.

The arrival of Claude Fable, which sits outside the original Haiku-Sonnet-Opus poetry ladder, is what gave the post its hook. Once a vendor steps off its own established metaphor, the door opens for everyone to speculate about what comes next, and speculation in a developer audience usually arrives as comedy. The "Fable (xhigh)" entry, captioned "Bankruptcy speedrun," is a direct jab at the reasoning-effort and verbosity settings that can multiply a query's cost without an obvious ceiling.

The counter-view: naming is a genuinely hard problem

It would be easy to read all this as pure vendor failure, but that framing is too tidy. Naming a rapidly evolving product line is legitimately difficult, and the critics rarely propose something better.

Consider the alternatives. Pure version numbers (Model 5.2.1) are precise but tell a non-expert nothing about size or purpose. Capability descriptors (Fast, Smart, Smartest) age badly the moment a new tier ships, because today's "Smartest" becomes tomorrow's midrange. Marketing names disconnected from any scheme (think GPU code names) require memorization with no mental model attached. Every option trades clarity in one dimension for confusion in another.

The literary metaphor, whatever its eventual breaking point, does one thing well: it gives non-technical users an intuitive ordering without exposing them to benchmark numbers they cannot interpret. A product manager who has never read a model card still understands that a sonnet is bigger than a haiku. That is real communicative value, and the satire, by extending the ladder to absurdity, actually demonstrates how legible the underlying system is. You cannot parody a scheme nobody understands.

What the reaction signals

The sentiment underneath the laughter is mild affection mixed with fatigue. Developers are not abandoning these tools over their names; they are poking fun the way people tease a product they use constantly. That is a healthier signal for Anthropic than indifference would be. Jokes require familiarity, and the granular references to token counts, reasoning settings, and version sprawl show an audience that knows the products intimately.

Still, there is a practical message buried in the comedy worth taking seriously. As model families grow, the gap between the friendly marketing name and the exact API identifier keeps widening, and that gap is where confusion and surprise billing live. Clearer documentation about which underlying model a product name maps to, and what a given effort or verbosity setting actually costs, would defuse most of the anxiety the satire plays on. The literary names can stay. What users want is a reliable map from the poem to the bill.

Wilkinson's post will fade from timelines in a week, as these things do. The condition that made it resonate will not. Frontier labs are shipping faster than their naming conventions can absorb, and until that stabilizes, the people who depend on these tools will keep doing what good engineers always do with systems that confuse them. They will document the quirks, build mental models that work, and write jokes about the parts that still do not make sense.