ResearchUpdated 5 October 2026Archived
LLM Accuracy Comparison, June 2026 (Archived)
Archived page
Archived. The models on this page (GPT-5.4, Claude Sonnet 4.6 and Opus 4.7, Gemini 3.1 Pro) have been superseded. The benchmark figures were vendor claims, not aixjo tests.
Short answer
In June 2026 Anthropic reported Claude Opus 4.7 at 64.3% on SWE-bench Pro and 78.0% on OSWorld; these were vendor claims about models that have since been replaced. For current models, see our October 2026 comparison.
Benchmark claims as published in June 2026
Evidence: Claimed by provider| Model (superseded) | Claim | Source type |
|---|---|---|
| Claude Opus 4.7 | SWE-bench Pro 64.3%; OSWorld 78.0% | Vendor claim |
| GPT-5.4 | No comparable figure recorded | — |
| Gemini 3.1 Pro | Knowledge cutoff early 2026 (vendor-stated) | Vendor claim |
How to test models yourself
Evidence: Editorial assessmentPublic benchmarks rarely match your own data. Use 20–50 real prompts from your work, score outputs against a written rubric, and re-run after each model upgrade. Keep a person reviewing anything customer-facing.
What changed in this update
- Removed the phrase 'aixjo editorial testing protocols': no aixjo tests were run.
- Labelled all benchmark figures as vendor claims about superseded models.
Common questions
Are these results still valid?
No. They describe superseded models and were vendor claims, not independent tests.
Sources
- Anthropic: Introducing Claude Opus 5.5 · checked 5 October 2026
- Anthropic: Introducing Claude Sonnet 5.5 · checked 5 October 2026