Agentic discipline
Structured outputs, tool-use boundaries, instruction following, grounded reasoning, and hallucination resistance.
Independent model evaluation · public artifacts · tactical evidence
AI model benchmarks with rules, receipts, and Arena replays.
Resyst Labs Benchmarks is a public evidence surface for comparing AI systems across agentic reasoning, software engineering, reliability, runtime economics, and Resyst Arena duels.
Current leader
DeepSeek V4 Flash
Overall 91.38 · DeepSeek direct API
11
ranked entrants
5
Arena matches
Jun 11, 2026
data refresh
Software execution
Practical implementation quality, final-answer usefulness, source handling, and architecture cleanliness.
Resyst Arena
Turn-based spatial duels where legal action discipline and tactical continuity are measured separately from runtime telemetry.
Unified ranking
One table, visible tradeoffs.
Local and API-backed systems share a single tournament view. Provider, runtime, cost, latency, reliability, and lane basis remain visible so comparisons stay honest.
DeepSeek V4 Flash
DeepSeek direct API
Open result → 91.4Qwen3.7 Max
OpenRouter · extra-high reasoning
Open result → 86.1MiniMax M3
OpenRouter · extra-high reasoning
Open result → 86.1| Rank | Model | Basis | Overall | Full | SWE | Cost | Reliability | Details |
|---|---|---|---|---|---|---|---|---|
| #1 | DeepSeek V4 Flash deepseek-v4-flash-direct | DeepSeek direct API | 91.38 | 95.80 | 86.95 | $0.086 | 100.0% | Result |
| #2 | Qwen3.7 Max qwen3.7-max-openrouter-xhigh | OpenRouter · extra-high reasoning | 86.11 | 83.24 | 88.99 | $0.851 | 100.0% | Result |
| #3 | MiniMax M3 minimax-m3-openrouter-xhigh | OpenRouter · extra-high reasoning | 86.11 | 85.33 | 86.88 | $0.157 | 100.0% | Result |
| #4 | Claude Opus 4.8 claude-opus-4.8-openrouter-xhigh | OpenRouter · extra-high reasoning | 86.00 | 83.32 | 88.67 | $5.981 | 100.0% | Result |
| #5 | DeepSeek V4 Pro deepseek-v4-pro-direct | DeepSeek direct API | 85.86 | 92.04 | 79.68 | $0.280 | 100.0% | Result |
| #6 | GPT‑5.5 gpt-5.5-openrouter-xhigh | OpenRouter · extra-high reasoning | 85.64 | 85.76 | 85.51 | $4.071 | 100.0% | Result |
| #7 | Gemini 3.5 Flash gemini-3.5-flash-openrouter | OpenRouter · google/gemini-3.5-flash | 81.19 | 88.79 | 73.59 | $1.502 | 100.0% | Result |
| #8 | Claude Fable 5 claude-fable-5-openrouter-xhigh | OpenRouter · extra-high reasoning | 79.22 | 77.02 | 81.42 | $11.815 | 100.0% | Result |
| #9 | MiniMax M3 Direct Plus minimax-m3-direct-anthropic | MiniMax Plus · direct API · SWE xhigh | 76.05 | 84.32 | 67.77 | $0.0037 | 100.0% | Result |
| #10 | Kimi K2.7 Code kimi-k2.7-code-openrouter-xhigh | OpenRouter · extra-high reasoning | 73.28 | 87.95 | 58.61 | $0.569 | 100.0% | Result |
| #11 | NVIDIA Nemotron 3 Ultra nemotron-3-ultra-openrouter-xhigh | OpenRouter · nvidia/nemotron-3-ultra-550b-a55b · xhigh reasoning | 69.08 | 79.54 | 58.63 | $0.369 | 100.0% | Result |
Resyst Arena
A tactical testbed, not a latency race.
Resyst Arena evaluates spatial strategy in deterministic turn-based games. Each public match summary links to a replay with board states, legal actions, events, and tactical telemetry.
- Side-swapped series before ranking-grade claims.
- Legal action rate and invalid actions are first-class evidence.
- Replay artifacts stay linked to tactical summaries.
Encounter 1 · 2 replays · latest 20260612 · xhigh text-actions
DeepSeek V4 Flash vs Kimi K2.7 Code
Kimi K2.7 Code leads 2–0. Replays stay grouped under this model-vs-model encounter, including side-swapped rounds.
Encounter 2 · 2 replays · latest 20260531 · 0.3
DeepSeek V4 Flash vs Step 3.7 Flash
DeepSeek V4 Flash leads 2–0. Replays stay grouped under this model-vs-model encounter, including side-swapped rounds.
Encounter 3 · 1 replay · latest 20260530 · 0.3
DeepSeek V4 Flash vs Gemini 3 Flash Preview
Gemini 3 Flash Preview won the replay. Replays stay grouped under this model-vs-model encounter, including side-swapped rounds.
Methodology
Scores are claims with receipts.
Separate lanes
Agentic and software-engineering results are preserved as distinct measurements before any publication formula combines them.
Runtime honesty
The same model can appear through different providers or runtimes. The table exposes basis metadata instead of hiding infrastructure differences.
Evidence thresholds
Single runs are artifacts. Stronger claims require repeated series, side swaps, seed variation, and comparable scoring settings.
Evidence contract
Public by design. Auditable by default.
This site publishes benchmark summaries as versioned data artifacts. The presentation layer is intentionally separate from the scoring harness so rankings can evolve without rewriting the public record.