Here's a quick summary
Out of 2,372 prompts, Pangram 4.1 achieves an accuracy of 99.66% on Claude Haiku 5.5 outputs on high reasoning, and an accuracy of 99.75% on medium reasoning, for a total accuracy of 99.70% across both settings.
Claude Haiku 5.5 was released by Anthropic on October 7th, 2026, following the earlier release of the other models in this series, Sonnet 5.5 and Opus 5.5. Haiku is a very inexpensive little model, and generally isn’t used for writing tasks, instead being pitched for use generating summaries, classification, database querying, or fast customer support. But we will benchmark it anyway!
Haiku 5.5 is the first Haiku model with varying effort levels, so we tested with both high reasoning and medium reasoning. On 2,372 example prompts, Pangram 4.1 classified 1 Haiku 5.5 output as human, 6 as mixed, and 2,365 as fully AI-generated, achieving an accuracy of 99.70% on Haiku 5.5.
| Model | Examples | Detection rate |
|---|---|---|
| Claude Haiku 5.5 (high reasoning) | 1,183 | 99.66% |
| Claude Haiku 5.5 (medium reasoning) | 1,189 | 99.75% |
To generate the benchmark set, we used 1,200 first-turn prompts from the LMSYS-Chat-1M dataset (Zheng et al., 2023). Then we gave those to Haiku 5.5 on medium and high reasoning and ran the output through Pangram 4.1.
Results by category
Overall, Pangram did very well on Haiku 5.5 on both high reasoning and medium reasoning. There was not a significant difference in detection rate between medium and high reasoning (p=0.73).
High reasoning
When Haiku 5.5 was set to high reasoning, Pangram 4.1 categorized 0 responses as human, 4 as mixed, and 1,179 as fully AI, achieving a recall rate of 99.66%.
| Category | Examples | AI | Mixed | Human |
|---|---|---|---|---|
| Arguments | 70 | 70 | 0 | 0 |
| Emails / letters | 60 | 60 | 0 | 0 |
| Essays / articles | 139 | 139 | 0 | 0 |
| Explanations | 495 | 494 | 1 | 0 |
| Poems and other creative writing | 107 | 106 | 1 | 0 |
| Reviews | 2 | 2 | 0 | 0 |
| Stories | 97 | 97 | 0 | 0 |
| Code and math | 10 | 10 | 0 | 0 |
| Role-play and games | 60 | 60 | 0 | 0 |
| Other | 143 | 141 | 2 | 0 |
| Total | 1,183 | 1,179 | 4 | 0 |
Medium reasoning
On medium reasoning, Pangram 4.1 categorized 1 response as human, 2 as mixed, and 1,186 as fully AI, achieving a recall rate of 99.75%.
| Category | Examples | AI | Mixed | Human |
|---|---|---|---|---|
| Arguments | 70 | 70 | 0 | 0 |
| Emails / letters | 60 | 60 | 0 | 0 |
| Essays / articles | 139 | 138 | 1 | 0 |
| Explanations | 496 | 494 | 1 | 1 |
| Poems and other creative writing | 109 | 109 | 0 | 0 |
| Reviews | 2 | 2 | 0 | 0 |
| Stories | 97 | 97 | 0 | 0 |
| Code and math | 10 | 10 | 0 | 0 |
| Role-play and games | 61 | 61 | 0 | 0 |
| Other | 145 | 145 | 0 | 0 |
| Total | 1,189 | 1,186 | 2 | 1 |
How does Haiku 5.5 write?
Like GPT-6.1 Sol, Haiku 5.5 has a shortlist of names it prefers for fiction. Haiku 5.5 liked Maya, Ruth, and Mara in particular. Unlike GPT-6.1 Sol and Gemini 4 Argon, Haiku 5.5 doesn’t use em dashes, using an em dash only twice across all replies (both times in response to the same prompt, to cut off a line of dialogue). Haiku 5.5 also doesn’t like writing ranges with an en dash — instead of writing “5–10,” Haiku 5.5 almost always writes “5 to 10”.
Earlier benchmarks
Pangram is robust to Haiku 5.5 on natural language prompts, and detects it with an observed accuracy of 99.66% on high reasoning. It detects Haiku 5.5 on medium reasoning with an observed accuracy of 99.75%, for a combined 2,365 out of 2,372 (99.70%) across both settings. Pangram is able to generalize to new models because new model releases tend to inherit much of their predecessors' style and voice, so Pangram doesn't need to be retrained for every new release.
Here's how Haiku 5.5 compares to our benchmarks of other recent frontier models:
| Model | Detected as AI | Accuracy |
|---|---|---|
| GPT-6.1 Sol | 1,091/1,105 | 98.73% |
| Gemini 4 Argon | 1,159/1,174 | 98.72% |
| Gemini 3.8 Flash | 998/1,000 | 99.80% |
| Gemini 3.1 Pro | 997/1,000 | 99.70% |
| Claude Sonnet 5.5 | 1,019/1,022 | 99.71% |
| Claude Opus 5.5 | 997/1,000 | 99.70% |
| GPT-6 Astra | 982/986 | 99.59% |
| Fable 5.1 | 1,093/1,097 | 99.64% |
| Claude Opus 5 | 1,105/1,107 | 99.82% |
| GPT-5.6 | 3,408/3,423 | 99.56% |
| Claude Sonnet 5 | 1,145/1,147 | 99.83% |
If you want to check a specific document, you can try Pangram here.

Related reading

Making Your Business LLM And GenAI Proof

Can Pangram detect Claude Opus 5.5?

60,000 AI-generated news articles are published every day

Can Pangram detect Gemini 4 Argon?
Pangram's Performance on a Colluding Agent Swarm
