Trends
One line per model creator, stepped at each release that raised that creator's best Known Good Index score. It holds flat until a better model ships, so a step is a real improvement rather than a rolling average. Click a legend entry to hide a creator.
Frontier quality over time
Known Good Index by creator, stepped at each release that raised their best score
Frontier Model Quality, Over Time
Known Good Index combines seven published evaluations: GPQA Diamond, Mock AIME 2024–25, MATH Level 5, SWE-bench Verified, LiveBench, Humanity's Last Exam and Terminal-Bench, each normalised 0–100. Scores are ingested from their publishers — Epoch AI (CC-BY) and LiveBench — and we do not run these evaluations. Each step is a new release that raised that creator's best score.
All 7 of 7 model creators⇩
OpenAIGoogleDeepSeekAlibabaAnthropicZ.aixAI
The Known Good
What this chart can and cannot tell you. A creator only appears once we hold at least two dated releases
with an index score, so newcomers are missing rather than flat, and a creator with undated releases cannot be
plotted at all — 7 of 7 tracked creators are drawn here. The line is a
running maximum of that creator's published scores, so it can never fall: a cheaper or smaller follow-up model does
not pull it down. Index scores themselves shift as new evaluations are ingested, so a point reflects what the
published data says today, not what was known on the release date.
Plotted creators — 7 of 7
| Creator | Steps plotted | First plotted release | Latest step | Best index |
|---|---|---|---|---|
| OpenAI | 7 | OpenAI: GPT-4o · May 2024 | gpt-5.5-pre-release · April 2026 | 97 |
| 5 | gemma-2-9b-it · June 2024 | Google: Gemini 3.1 Pro Preview · February 2026 | 97 | |
| DeepSeek | 3 | DeepSeek-R1-Distill-Qwen-32B · January 2025 | DeepSeek: DeepSeek V4 Pro 0423 · April 2026 | 93 |
| Alibaba | 4 | QwQ-32B · March 2025 | Qwen: Qwen3.7 Max · May 2026 | 93 |
| Anthropic | 5 | claude-3-opus-20240229 · February 2024 | Anthropic: Claude Opus 4.7 · April 2026 | 92 |
| Z.ai | 4 | Z.ai: GLM 4.7 · December 2025 | Z.ai: GLM 5.2 · June 2026 | 91 |
| xAI | 2 | grok-2-1212 · December 2024 | grok-4-0709 · July 2025 | 68 |
The Known Good
Source. Release dates come from provider metadata via models.dev and OpenRouter; index scores are computed
from evaluations published by Epoch AI (CC-BY) and LiveBench. We aggregate published results and do not run our own
evaluations. Method: methodology ·
Licences: attribution ·
Raw figures: download.