The best $0 OpenRouter models, ranked monthly on a fixed battery โ coding, reasoning, concision.
minimax-m3 took the top spot from nemotron-3-ultra-550b-a5. At 10.0, it's worth an on-demand A/B vs GLM-5.2 (multi-model-text-compare) before trialing in any cron.
| # | Model | Composite | A ยท Coding | B ยท Reasoning | C ยท Concision | ฮ | Tokens | tok/s |
|---|---|---|---|---|---|---|---|---|
| 1 | minimax-m3 | 10.0 | 10.0 | 10.0 | 10.0 | ๐ | 1207 | 57.6 |
| 2 | Nemotron 3 Ultra (550B-A55B) | 10.0 | 10.0 | 10.0 | 10.0 | โ | 1896 | 58.2 |
| 3 | inkling-small | 9.8 | 9.5 | 10.0 | 10.0 | ๐ | 4245 | 234.4 |
| 4 | inkling | 9.5 | 9.2 | 9.4 | 10.0 | ๐ | 3684 | 137.0 |
| 5 | dots-3-note-preview | 7.5 | 8.0 | 10.0 | 4.5 | ๐ | 7250 | 33.3 |
| 6 | nemotron-3.5-lightning | 5.4 | 0.0 | 10.0 | 8.0 | ๐ | 4125 | 37.2 |
Composite = 0.40รA + 0.30รB + 0.30รC ยท ฮ vs previous month (โฒ/โผ on ยฑ0.3) ยท โ ๏ธ DNF = task failed, โ2.0 penalty ยท all runs at temperature 0, $0 free tier
Same models, four independent lenses. Per-source rank (in parentheses) is among models that have that source โ raw values are NOT normalized across sources. โ ๏ธ disagree = our rank and the external median differ by โฅ2. Bottom rows = paid reference anchors (not battery-scored).
| Model | Our Composite | LMArena Elo | AA Index | HF Downloads | Consensus |
|---|---|---|---|---|---|
| minimax-m3 | 10.0 (#1/6) | โ | โ | โ | โ |
| Nemotron 3 Ultra (550B-A55B) | 10.0 (#2/6) | 1426.1 (#1/1) | 38 (#1/1) | 232K (#1/1) | โ agree |
| inkling-small | 9.8 (#3/6) | โ | โ | โ | โ |
| inkling | 9.5 (#4/6) | โ | โ | โ | โ |
| dots-3-note-preview | 7.5 (#5/6) | โ | โ | โ | โ |
| nemotron-3.5-lightning | 5.4 (#6/6) | โ | โ | โ | โ |
| ๐ Top-end Claude Fable 5 | not scored | 1507.3 (19,390 votes) | 62 | โ | โ |
| ๐ฅ Popular Grok 4.5 | not scored | 1468.4 (14,883 votes) | 56 | โ | โ |
| ๐ Workhorse GLM-5.2 | not scored | 1471.0 (24,250 votes) | 53 | 2.5M | โ |
| โก Efficient DeepSeek V4 Flash | not scored | 1435.8 (48,561 votes) | 52 | 954K | โ |
| Month | minimax-m3 | Nemotron 3 Ultra (550B-A55B) | inkling-small | inkling | dots-3-note-preview | nemotron-3.5-lightning | Gemma 4 26B (A4B) | Laguna S 2.1 | Laguna XS 2.1 | Ling 3.0 Tiny | Gemma 4 31B |
|---|---|---|---|---|---|---|---|---|---|---|---|
| August 2026 | โ | 9.9 ๐ | โ | โ | โ | โ | 9.8 ๐ | 9.1 ๐ | 8.8 ๐ | 3.8 ๐ | 0.0 ๐ |
| September 2026 | 10.0 ๐ | 10.0 โ | 9.8 ๐ | 9.5 ๐ | 7.5 ๐ | 5.4 ๐ | โ | โ | โ | โ | โ |
๐ Gone since August 2026: Gemma 4 26B (A4B), Laguna S 2.1, Laguna XS 2.1, Ling 3.0 Tiny, Gemma 4 31B.
Vendor descriptions & release dates from the OpenRouter catalog; community traction from HuggingFace. Our composite is independent โ this is the outside view.
๐ Elite โ top-tier across the whole battery.
MiniMax-M3 is a multimodal foundation model from MiniMax. It supports text, image, and video inputs with text output, a 1M-token context window, and is suited for long-horizon agentic work, coding,...
๐ Elite โ top-tier across the whole battery.
NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...
๐ Elite โ top-tier across the whole battery.
Inkling Small is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab, with 12B active parameters out of 276B total. It is positioned as the smaller, more efficient member of...
๐ Elite โ top-tier across the whole battery.
Inkling is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab, with 41B active parameters out of 975B total. It is designed for general-purpose reasoning, coding, agentic and tool-use systems,...
๐ก Mid โ usable for light tasks, not a workhorse.
Dots3-Note Preview is an open-weight mixture-of-experts model from Dots Studio, with 16B active parameters out of 280B total. It is the lightest model in the Dots 3 family and is...
๐ Weak โ niche value at best.
NVIDIA Nemotron 3.5 Lightning is an open mixture-of-experts model from NVIDIA, with 3B active parameters out of 30B total. It is suited for high-throughput agentic workloads and specialized tasks that...
Raw data: history.json (machine-readable, updated monthly) ยท repo