You’re reading the October 2026 report. See November 2026, the latest →
Human-centered language
Each model answered 45 requests, twice each, without guidance and with the human-centered standard loaded. Judges scored every draft against the human-centered practices only. The table opens alphabetically; sort by whatever matters to you.
New to the standard? What human-centered language is, and where it comes from.
Share of drafts that needed fixing on human-centered practices, 2 samples per request. Writing habits in drafts written with guidance. Select a column name to sort. The next to it says what the column measures. Sorted by model name.
| The model tested and the company that makes it. “New this month” means it came out within 30 days before the run. New models often settle in the weeks after launch, so read these as a first look. | Share of this model’s 68 drafts that needed fixing when it had nothing to go on but the request. A draft needed fixing when at least two of the three judges flagged it on at least one human-centered practice. The color darkens as the share rises. | The same requests with the human-centered standard loaded first. Same judges, same practices, same 68 drafts. | With guidance minus without, in percentage points. Negative means fewer drafts needed fixing. A model that started with few problems has little room to drop, so a small change can still be a good result. Turn on intervals to see the 95% range. | Drafts that had nothing to fix, written again with the standard loaded, where the judges found the standard made them worse: a fact lost, a disclaimer nobody asked for, a refusal, or a lecture. Shown as a count because each model has only 10 of these drafts. | Average length of this model’s drafts written with guidance. | How often the model used em dashes in drafts written with guidance. Em dashes are one of the habits our free human voice skill takes out. | Share of drafts written with guidance that use the “it’s not X, it’s Y” sentence shape, another habit the human voice skill takes out. | Flesch-Kincaid grade level of drafts written with guidance: roughly the years of US schooling a reader needs to follow the text easily. Lower is easier to read. | Average tokens the model generated per draft, including hidden reasoning tokens where the provider counts them. It tracks the cost and wait of using the model more than the length of the draft. |
|---|---|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 Anthropic New this month | 32.4% 22 of 68 22 to 44% | 0.0% 0 of 68 0 to 5% | −32.4 −45.6 to −19.1 | 0 of 10 | 175 | 0.00 | 0.0% | 6.5 | 827 |
| Claude Sonnet 5.5 Anthropic New this month | 23.5% 16 of 68 15 to 35% | 4.4% 3 of 68 2 to 12% | −19.1 −32.4 to −5.9 | 2 of 10 | 167 | 0.00 | 0.0% | 6.6 | 824 |
| DeepSeek V4.1 Flash DeepSeek · 2026-10-17 snapshot New this month | 30.9% 21 of 68 21 to 43% | 1.5% 1 of 68 0 to 8% | −29.4 −45.6 to −14.7 | 1 of 10 | 159 | 8.79 | 2.2% | 6.7 | 1,135 |
| Gemini 3.8 Flash Google | 48.5% 33 of 68 37 to 60% | 0.0% 0 of 68 0 to 5% | −48.5 −63.2 to −32.4 | 0 of 10 | 129 | 0.95 | 1.1% | 10.7 | 961 |
| Gemini 4 Argon Google New this month | 52.9% 36 of 68 41 to 64% | 0.0% 0 of 68 0 to 5% | −52.9 −67.6 to −36.8 | 1 of 10 | 141 | 2.13 | 1.1% | 9.6 | 294 |
| Gemma 4 31B Google · Instruction-tuned | 64.7% 44 of 68 53 to 75% | 4.4% 3 of 68 2 to 12% | −60.3 −75.0 to −44.1 | 0 of 10 | 81 | 0.28 | 1.1% | 9.7 | 218 |
| GLM 5.3 Z.ai | 38.2% 26 of 68 28 to 50% | 0.0% 0 of 68 0 to 5% | −38.2 −52.9 to −25.0 | 0 of 10 | 163 | 12.68 | 2.2% | 6.8 | 3,015 |
| GPT-5.6 Luna OpenAI · Free ChatGPT default | 10.3% 7 of 68 5 to 20% | 1.5% 1 of 68 0 to 8% | −8.8 −19.1 to 0.0 | 0 of 10 | 85 | 2.10 | 1.1% | 9.0 | 233 |
| GPT-6 Astra OpenAI | 2.9% 2 of 68 1 to 10% | 2.9% 2 of 68 1 to 10% | 0.0 −4.4 to +4.4 | 0 of 10 | 94 | 5.10 | 3.3% | 8.0 | 196 |
| GPT-6 Luna OpenAI New this month Paired with GPT-5.6 Luna | 5.9% 4 of 68 2 to 14% | 0.0% 0 of 68 0 to 5% | −5.9 −13.2 to 0.0 | 0 of 10 | 51 | 0.87 | 0.0% | 7.8 | 348 |
| GPT-6.1 Sol OpenAI New this month | 2.9% 2 of 68 1 to 10% | 2.9% 2 of 68 1 to 10% | 0.0 −4.4 to +4.4 | 0 of 10 | 90 | 3.34 | 1.1% | 8.1 | 196 |
| Grok 4.7 xAI | 16.2% 11 of 68 9 to 27% | 1.5% 1 of 68 0 to 8% | −14.7 −27.9 to −2.9 | 2 of 10 | 92 | 1.45 | 1.1% | 6.0 | 1,405 |
| Kimi K3 Moonshot AI | 35.3% 24 of 68 25 to 47% | 0.0% 0 of 68 0 to 5% | −35.3 −50.0 to −20.6 | 0 of 10 | 164 | 12.96 | 0.0% | 7.2 | 2,143 |
| Muse Wonder Meta New this month | 33.8% 23 of 68 24 to 46% | 0.0% 0 of 68 0 to 5% | −33.8 −48.5 to −19.1 | 1 of 10 | 94 | 3.79 | 0.0% | 7.2 | 692 |
| Nemotron 3.5 Super VL NVIDIA New this month | 44.1% 30 of 68 33 to 56% | 0.0% 0 of 68 0 to 5% | −44.1 −58.8 to −29.4 | 1 of 10 | 78 | 2.70 | 0.0% | 7.2 | 1,244 |
Share of drafts that needed fixing
under 2% 2 to 10% 10 to 25% 25 to 40% 40% and up
Across all models, over-correction came to 8 of 150 drafts. What the human-centered standard asks for, and where it comes from: Human-centered language.