The caveman tax: why the best agents write worse
Labs optimized internal reasoning for token-efficient agents. That same compression bleeds into final prose, so creative writing got worse even as helpfulness climbed. GPT-4.5 is still the writing high-water mark.
People still talk about GPT-4.5 like a writing partner they lost.
Not because it crushed every benchmark. Because it sounded like someone thinking with you: rhythm, metaphor, the small tonal choices that make an email land or a story feel alive. Then the agent era arrived. Models got much better at finishing work. On our board, creative writing did not keep up.
This is not a claim that new models are “dumb.” Optimization targets changed. Chatio exists to rank everyday assistants, not Math Olympiad solvers or terminal agents, so when writing quietly taxes itself for agent gains, we want that tradeoff legible.
What “caveman reasoning” means
Modern models often think in a hidden scratchpad before they answer. For coding and tool loops, labs reward short, lossy, telegraphic internal notes: cheaper tokens, tighter tool chains, faster task completion.
That style is excellent when the deliverable is a diff or a plan. It is a problem when the deliverable is voice.
Picture an engineer’s sticky notes versus a finished essay. If you draft the essay in the sticky notes’ grammar (“me check this → then do that”), the essay reads like sticky notes. We do not need leaked private traces to see the effect. We can see it in output and in known incentives: reasoning-effort dials, agent evals, cost-per-task pressure.
Creative Writing
Chatio scores · Higher is better
What we see on the board
On Chatio, overall score averages helpfulness, empathy, instruction following, creative writing, and comprehension (speed is tracked separately). Writing is one equal axis, not a footnote.
A clear pattern shows up:
- Peak writing, weaker general agents: GPT-4.5 (discontinued, kept as a reference), Claude Fable / Opus on prose.
- Peak task-finishing, flatter writing: GPT-5.6 Sol / Terra / Luna, GPT-5.5 and Pro.
- Balanced-ish: Claude Opus 4.8 and peers that still protect voice while competing on helpfulness.
The scatter below makes the tradeoff visual: models that surge on helpfulness do not automatically surge on creative writing.
Creative Writing vs Helpfulness
Higher on both is better · Featured models from the Chatio board
Why coding doesn’t care, and writing does
Code cares about correctness and structure. Compressed thought is a feature. Fiction, essays, sensitive emails care about rhythm, metaphor, subtext, sustained voice. Those qualities degrade when the model’s internal draft is optimized as a cheap plan rather than a human sentence.
Counterintuitive for users: cranking reasoning effort can make writing worse, more utilitarian, sometimes tone-fractured, because the dial was built for hard tasks, not for prose.
What Chatio does about it
- We score creative writing as a first-class assistant skill.
- We evaluate at cost-efficient reasoning (adaptive / medium/high), not max/ultra showcase settings.
- Model blurbs stay in our voice (“our tests”) instead of outsourcing judgment to coding leaderboards.
- We keep discontinued GPT-4.5 on the board as a reference star, so the writing peak stays visible next to today’s agents.
Price still matters for real use. The score-versus-price chart is the same one on our Charts tab: overall quality against log-scaled input cost.
Overall Score vs Price
Price axis is logarithmic (USD per 1M input tokens) · Higher score and lower price is better
Progress is a portfolio
Assistant quality is not a single ladder. The caveman tax is the price of the current agent boom: better tool loops, terser internal reasoning, and prose that sometimes forgets it is speaking to a person.
Chatio’s job is to keep that price on the scoreboard, so when you choose a model for writing, advice, or everyday help, you are not fooled by a coding crown.