This week's two chart-toppers: Intelligence chart – Claude Opus 5.5 (Max, Default Fallback) (57.6 points), Coding chart – Claude Fable 5.1 (Max, Default Fallback) (81.6 points). See table for up/down arrows.
Full score is 100, but don't be fooled by the number: the test suite includes hellish benchmarks like Humanity’s Last Exam, and 57.6 points is already the highest in human history.
Each lab's strongest: Current status + two yardsticks
3/6 months ago is the baseline; this week + 1 month ago is the current status, telling you who's stagnating and who's parachuting in.
Intelligence
| Lab | This week's strongest | 1 month ago | 3 months ago | 6 months ago |
|---|---|---|---|---|
| Overall Champion | None | Claude Fable 5.1，53.4 | GPT-5.5，38.4 | Claude Opus 4.5，29.1 |
| Anthropic | None | Claude Fable 5.1，53.4 | Claude Sonnet 5，38.2 | Claude Opus 4.5，29.1 |
| OpenAI | None | GPT-6 Astra，52.7 | GPT-5.5，38.4 | GPT-5.2 Codex，28.5 |
| None | Gemini 3.5 Flash，32.6 | Same | Gemini 2.5 Pro，16.1 | |
| xAI | None | Grok 4.6，42.8 | Grok Build 0.1 0616，27.2 | Grok 4.20 0309，25.2 |
| Qwen | None | Qwen3.5 27B，22.9 | Same | Same |
| Zhipu | None | GLM-5.3-Flash，41.8 | GLM-5，27.9 | Same |
Coding
| Lab | This week's strongest | 1 month ago | 3 months ago | 6 months ago |
|---|---|---|---|---|
| Overall Champion | None | Claude Fable 5.1，81.6 | GPT-5.5，74.9 | Qwen3.5 397B A17B，48.2 |
| Anthropic | None | Claude Fable 5.1，81.6 | Claude Sonnet 5，71.5 | Claude 3.7 Sonnet，36.4 |
| OpenAI | None | GPT-6 Astra，76.9 | GPT-5.5，74.9 | o1-preview，34.0 |
| None | Gemini 3.5 Flash，70.1 | Same | Gemini 2.5 Pro，33.3 | |
| xAI | None | Grok 4.6，74.4 | Grok Build 0.1 0616，51.5 | None |
| Qwen | None | Qwen3.5 397B A17B，48.2 | Same | Same |
| Zhipu | None | GLM-5.3-Flash，71.5 | None | Same |
Can't beat the 6-month-ago champion = falling behind the times; beating the 3-month-ago champion counts as catching up to the era.
Passing list: Who caught up to the era? Who's free to use?
★ = passed the 3-month strong line as well; ✅ = this version free; ◐ = free version in same series; — = none. Free access mostly limited-time, as of publication date.
Intelligence——Passed 6-month line (29.1 points) 36 models
Coding——Passed 6-month line (48.2 points) 27 models
| Model | Score | Strong line | OpenRouter free | Opencode free |
|---|---|---|---|---|
| Claude Fable 5.1 (Max, Default Fallback) (—) | 81.6 | ★ | — | — |
| Claude Fable 5.1 (Medium, Default Fallback) (—) | 77.1 | ★ | — | — |
| Claude Opus 5 (Xhigh) (—) | 77.0 | ★ | — | — |
| GPT-6 Astra (Max) (—) | 76.9 | ★ | — | — |
| GPT-6 Astra (Xhigh) (—) | 75.9 | ★ | — | — |
| Claude Fable 5.1 (Low, Default Fallback) (—) | 75.2 | ★ | — | — |
| GPT-5.5 (Xhigh) (—) | 74.9 | Strong line itself | — | — |
| Grok 4.6 (Medium) (—) | 74.4 | — | — | |
| Muse Spark 1.2 (Xhigh) (—) | 72.2 | ◐ (same series) | ◐ (same series) | |
| GLM-5.3-Flash (New) | 71.5 | — | — (itself US$0.15 very cheap) | |
| Claude Sonnet 5 (Max) | 71.5 | — | — | |
| GPT-5.6 Luna (Max) | 71.4 | — | — (itself US$0.2 very cheap) | |
| Gemini 3.5 Flash (High) | 70.1 | — | — | |
| GPT-5.6 Sol (Low) | 69.7 | — | — | |
| Claude Opus 5 (Low) | 66.9 | — | — | |
| GPT-5.6 Terra (Medium) | 64.7 | — | — | |
| Kimi K2.6 (Reasoning) | 61.8 | — | — | |
| Apodex 1.1 | 60.8 | ◐ (same series) | ◐ (same series) | |
| MiMo-V2.5-Pro (Reasoning) | 60.2 | ◐ (same series) | ◐ (same series) | |
| Nex-N2-Pro (Based on Qwen3.5-397B-A17B) | 59.1 | — | — | |
| MiniMax-M3 | 58.6 | — | — | |
| Ling-3.0-flash-VL | 57.0 | ◐ (same series) | ◐ (same series) | |
| MiMo-V2.5 | 56.8 | — | ✅ | |
| Solar Pro 4 | 52.7 | — | — | |
| Inkling (Xhigh) | 52.1 | ✅ | — | |
| Grok Build 0.1 0616 | 51.5 | — | — | |
| Gemini 3.5 Flash-Lite | 49.3 | — | — |
Cost-performance: 6-month version vs 3-month version
6-month version entry > adequate line, pass strong line ×1.5; 3-month version entry > strong line, same tier comparison without multiplier.
6-month version
| Intelligence (Model/Score/Input price/Per-dollar value) | Coding (Model/Score/Input price/Per-dollar value) | |
|---|---|---|
| 1 | Claude Haiku 5.5 ／41.2／US$0.1／618.0 | Ling-3.0-flash-VL／57／US$0.075／760.0 |
| 2 | GLM-5.3-Flash／41.8／US$0.15／418.0 | GLM-5.3-Flash／71.5／US$0.15／476.7 |
| 3 | Claude Haiku 5.5 ／37.8／US$0.1／378.0 | MiMo-V2.5／56.8／US$0.14／405.7 |
3-month version
| Intelligence (Model/Score/Input price/Per-dollar value) | Coding (Model/Score/Input price/Per-dollar value) | |
|---|---|---|
| 1 | Claude Haiku 5.5 ／41.2／US$0.1／412.0 | Claude Opus 5 ／77／US$5／15.4 |
| 2 | GLM-5.3-Flash／41.8／US$0.15／278.7 | Claude Fable 5.1 ／81.6／US$10／8.2 |
| 3 | MiMo-V2.6-Pro／46.3／US$0.435／106.4 | Claude Fable 5.1 ／77.1／US$10／7.7 |
Strongest free + best bang for buck
Free・6-month version (all passed adequate line)
| Rank | Model | Score (chart) | OpenRouter | Opencode | Both? |
|---|---|---|---|---|---|
| 1 | MiMo-V2.5 | 56.8 (Coding) | — | ✅ | No overlap |
| 2 | Inkling (Xhigh) | 52.1 (Coding) | ✅ | — | No overlap |
| 3 | MiMo-V2.6-Flash | 37.9 (Intelligence) | — | ✅ | No overlap |
Free・3-month version: None. All failed to pass the strong line, free can't catch up to the frontier, but beat the world champion from six months ago, enough for daily use.
Best bang for buck
| 6-month version pick | 3-month version pick | |
|---|---|---|
| Intelligence | Claude Haiku 5.5 (Xhigh) (41.2 points/US$0.1/618.0) | Claude Haiku 5.5 (Xhigh) (41.2 points/US$0.1/412.0) |
| Coding | Ling-3.0-flash-VL (57 points/US$0.075/760.0) | Claude Opus 5 (Xhigh) (77 points/US$5/15.4) |
Compute timeline: What chips were the models scoring now trained on?
| Chip/Architecture | Announced | Shipping status |
|---|---|---|
| B200 (Blackwell) | March 2024 GTC | Mature large-scale commercial use, mainstay of major clouds |
| B300 (Blackwell Ultra) | Late 2024/early 2025 spec freeze | Shipping ramping up, DGX B300 available for purchase, clouds rentable by the hour |
| Vera Rubin (R100+HBM4) | 2025/2026 GTC finalized | Mass production mid-2026, partner deliveries H2, not yet widely commercial |
Pattern: 9–12 months from start of shipping is the next AI compute upgrade peak. B300 only rolling out at scale this year, earliest year-end to see next-gen models trained on it; Rubin-trained stuff, 2027 before we see it.
Other quick updates
・Most expensive: The most expensive input on the whole chart, expensive and weak – named as not recommended; current flagship price US$10.
・Specialist king: Those that never passed Intelligence line but are world-class in Coding, code-writing specialists.
Data source: Artificial Analysis Intelligence Index and Coding Agent Index, free public data subsets; free model lists at OpenRouter and Opencode, limited-time free can end anytime, all as of publication date.