benchr
Ask any of these models to "write the whole thing in one go" and you eventually hit a wall that has nothing to do with talent. It's the max output token limit: the hard cap on how much a model can emit in a single response. For a tweet it never matters. For a long report, a sample chapter, or a full draft you don't want to stitch together by hand, it's the first real constraint, and it splits these three before you've judged a single sentence.
| Model | Max output (tokens) | ≈ words, one pass | Context window |
|---|---|---|---|
| Claude Opus 5 | 128K | ~90,000 | 1M |
| Claude Sonnet 5 | 128K | ~90,000 | 1M |
| GPT-5.6 Sol | 128K | ~90,000 | 1.05M |
| Claude Opus 4.8 (May 2026) | 128K | ~90,000 | 1M |
| GPT-5.5 (May 2026) | 128K | ~90,000 | 1.05M |
| Claude Sonnet 4.6 (May 2026) | 64K | ~45,000 | 1M |
So for long-form in one shot, the current generation removed the old split: Opus 5, Sonnet 5 and GPT-5.6 Sol each cap at 128K output tokens, close to a short book before they run out of room. When this piece first ran, only Opus 4.8 and GPT-5.5 reached that ceiling; Sonnet 4.6 stopped at 64K per Anthropic's own docs, even though some sources listed it at 128K. If you're still on Sonnet 4.6, that cap still applies.
Voice: the part no one will score for you
Neither Anthropic nor OpenAI markets a model as the best creative writer, and there's no official benchmark that settles "better prose." So anyone who tells you one of these flatly writes better is selling you their taste as a fact. What the providers do say is narrower and more useful.
Claude's house style tends to be plainer and more human on the first try, which is why it's the writer's default in the Sonnet 4.6 review and why people reach for it on messages and emails. GPT-5.5 is tuned to be concise and is pitched at professional, document-heavy work, so it's strong on briefs, summaries, and clean structure. The practical takeaway: run the same prompt through both and keep the voice that sounds like you. That's a five-minute test that beats any third-party claim, this one included. Those descriptions were written for Sonnet 4.6 and GPT-5.5. This piece has no provider writing claim on record for Sonnet 5 or GPT-5.6 Sol, so don't assume the house style carried over unchanged; the test matters more after a model swap.
Instruction-following: where the brief lives or dies
This is the axis with an actual paper trail, and it favors Claude. Anthropic's stated improvement for Sonnet 4.6 is "superior instruction-following" and consistency, and it describes the model as less prone to overengineering. In writing terms, that means when your brief says "1,200 words, second person, no bullet points, skip the intro," Claude is more likely to hold all four constraints at once. GPT-5.5's instinct toward concise, reshaped output is great when you want it and a problem when you wanted exactly what you asked for.
For anything where the spec matters, a style guide, a word count, a structure you have to hit, Claude's discipline is the safer bet. The deeper version of this trade, across coding as well as prose, is in Opus 4.8 vs GPT-5.5, and the generalist case for GPT-5.5 is in the GPT-5 review.
Go with Claude when the brief is detailed and you need it followed to the letter. Sonnet 5, the Claude.ai default on Free and Pro, reaches the same 128K output ceiling as Opus 5, so you no longer need Opus for length alone. Go with ChatGPT for concise, professional drafting and clean structure — GPT-5.6 Sol on Plus and Pro, GPT-5.6 Luna on Free and Go — and rerun your voice test on that model rather than trusting what GPT-5.5 did. On pure voice, skip everyone's claims, including ours: write the same prompt twice and keep the one that reads like you.