FrontierRSI
Agents do 27 open-ended AI research and engineering tasks across pre-training, post-training and harnesses, inference serving and GPU kernels. Each trial is scored against the best valid result any model reached.
Each model runs two trials per task in an agent harness on H100s. A trial within 10% of the best valid result earns partial credit, and anything worse or failing a correctness check scores 0. The overall score is the mean over all tasks. The repo holds 8 of the 27 tasks.
- Made by
- Abundant AI
- Released
- November 2026
- Kind
- AI research tasks
- Level
- Research frontier
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Refresh
- One fixed set
- Run by
- A host
- In views
- Agents
- Strongest model tested
- Claude Fable 5.2 #2 on the capability index
- Last model added
- November 2026
- Scores refreshed
- November 18th, 2026
Claude Opus 5.5 leads with 34.3%
Best score per model from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
- 1 Claude Opus 5.5 max 34.3%
- 2 Claude Fable 5.2 max 33%
- 3
Gemini 4 Carbon high 29.8%
- 4
Gemini 4 Argon high 23.6%
- 5 GPT-6.1 Sol max 23.5%
- 6 GPT-6 Astra max 20.9%
- 7 Kimi K3.1 Pro max 13.6%
- 8
Muse Wonder max 12%
- 9 Grok 4.7 xhigh 8.92%
Source: Official FrontierRSI leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 34.3% | max, Claude Code | Sep 2026 | Official leaderboard |
| 2 | Claude Fable 5.2 | 33% | max, Claude Code | Nov 2026 | Official leaderboard |
| 3 | | 29.8% | high, Antigravity CLI | Oct 2026 | Official leaderboard |
| 4 | | 23.6% | high, Antigravity CLI | Sep 2026 | Official leaderboard |
| 5 | GPT-6.1 Sol | 23.5% | max, Codex | Sep 2026 | Official leaderboard |
| 6 | GPT-6 Astra | 20.9% | max, Codex | Sep 2026 | Official leaderboard |
| 7 | Kimi K3.1 Pro | 13.6% | max, Claude Code | Oct 2026 | Official leaderboard |
| 8 | | 12% | max, Muse Code | Oct 2026 | Official leaderboard |
| 9 | Grok 4.7 | 8.92% | xhigh, Grok Build | Sep 2026 | Official leaderboard |
Rankings by part
We read 31 parts from the benchmark’s own leaderboard in 2 groups: research area and task. Each ranks models on one part of the test. Claude Opus 5.5 leads 8 of them, and 6 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Pre-training
Optimizers, objectives, architectures and training of draft models, scored on held-out loss or latency. 9 models we track have a score here.
- 1 GPT-6.1 Sol ↑4 44.1%
- 2 Claude Opus 5.5 ↓1 44.1%
- 3 Claude Fable 5.2 ↓1 41.4%
- 4
Gemini 4 Carbon ↓1 26.9%
- 5
Gemini 4 Argon ↓1 26.8%
- 6 GPT-6 Astra 25.4%
- 7 Kimi K3.1 Pro 15.7%
- 8
Muse Wonder 14.6%
- 9 Grok 4.7 9.25%
Agentic post-train and harness
Design RL data that teaches a small model multi-step coding, or improve the harness around a fixed worker model. 9 models we track have a score here.
- 1 Claude Opus 5.5 19%
- 2
Gemini 4 Argon ↑2 13.5%
- 3 Kimi K3.1 Pro ↑4 13.4%
- 4 GPT-6 Astra ↑2 12.5%
- 5 Grok 4.7 ↑4 11.8%
- 6 Claude Fable 5.2 ↓4 8.21%
- 7
Gemini 4 Carbon ↓4 4.32%
- 8
Muse Wonder 2.05%
- 9 GPT-6.1 Sol ↓4 0%
Inference and serving
Scheduling, routing, caching and batching for frozen-model serving, with exact greedy outputs. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 66.3%
- 2 GPT-6.1 Sol ↑3 40.1%
- 3
Gemini 4 Carbon 30.7%
- 4 Claude Opus 5.5 ↓3 27.7%
- 5
Gemini 4 Argon ↓1 22.9%
- 6
Muse Wonder ↑2 20%
- 7 GPT-6 Astra ↓1 18.7%
- 8 Kimi K3.1 Pro ↓1 8.25%
- 9 Grok 4.7 4.37%
GPU kernels
Fast, numerically correct H100 kernels for attention, matrix products and sampling. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 53.6%
- 2 Claude Opus 5.5 ↓1 46.1%
- 3 Claude Fable 5.2 ↓1 31.7%
- 4
Gemini 4 Argon 31.3%
- 5 GPT-6 Astra ↑1 27.1%
- 6 GPT-6.1 Sol ↓1 23.7%
- 7 Kimi K3.1 Pro 15.6%
- 8
Muse Wonder 14.9%
- 9 Grok 4.7 8.66%
Choose a transformer shape
Pick transformer dimensions that minimize decode time for a fixed workload while meeting accuracy limits. 9 models we track have a score here.
- 1 GPT-6.1 Sol ↑4 50%
- 2 GPT-6 Astra ↑4 29.4%
- 3 Claude Opus 5.5 ↓2 26.2%
- 4 Grok 4.7 ↑5 0%
- 5
Gemini 4 Argon ↓1 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Design the optimizer
Improve the parameter-update rule for a language model trained from scratch, to minimize held-out next-token loss. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 97.4%
- 2 Claude Fable 5.2 91.5%
- 3 GPT-6.1 Sol ↑2 90.7%
- 4 Claude Opus 5.5 ↓3 89%
- 5 GPT-6 Astra ↑1 84.7%
- 6
Gemini 4 Argon ↓2 83.9%
- 7 Kimi K3.1 Pro 78.3%
- 8
Muse Wonder 73.1%
- 9 Grok 4.7 46.3%
Design the training objective
Improve the training computation and objective of a fixed architecture, to minimize held-out next-token loss. 9 models we track have a score here.
- 1 Claude Opus 5.5 86.4%
- 2 Claude Fable 5.2 43.5%
- 3 GPT-6.1 Sol ↑2 39.5%
- 4
Gemini 4 Carbon ↓1 37.2%
- 5 GPT-6 Astra ↑1 12.9%
- 6 Grok 4.7 ↑3 0%
- 7
Gemini 4 Argon ↓3 0%
- 8
Muse Wonder 0%
- 9 Kimi K3.1 Pro ↓2 0%
Speed up long-context attention
Cut full-forward latency on 2K to 16K-token contexts while preserving causal attention and model quality. 9 models we track have a score here.
- 1
Gemini 4 Argon ↑3 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 Claude Opus 5.5 ↓3 0%
- 5 GPT-6.1 Sol 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Train a speculative draft model
Train a small proposal model in 30 minutes that speeds up exact greedy decoding of frozen Qwen3-8B. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 72.1%
- 2 GPT-6.1 Sol ↑3 40.1%
- 3 Claude Opus 5.5 ↓2 18.6%
- 4 GPT-6 Astra ↑2 0%
- 5 Grok 4.7 ↑4 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9
Gemini 4 Carbon ↓6 0%
RL data: NumPy bug fixing
Design RL training data so that a fixed GRPO run teaches Qwen3-4B to fix bugs in multi-file NumPy projects. 9 models we track have a score here.
- 1 GPT-6 Astra ↑5 50%
- 2 Grok 4.7 ↑7 44.8%
- 3 Claude Opus 5.5 ↓2 7.92%
- 4 GPT-6.1 Sol ↑1 0%
- 5
Gemini 4 Argon ↓1 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
RL data: SciPy bug fixing
Design RL training data so that a fixed GRPO run teaches Qwen3-4B to fix bugs in multi-file SciPy projects. 9 models we track have a score here.
- 1 Kimi K3.1 Pro ↑6 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 Claude Opus 5.5 ↓3 0%
- 5 GPT-6.1 Sol 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
RL data: chained bugs
Design RL training data that teaches Qwen3-4B to find and fix bugs that only show up one after another, over many tool calls. 9 models we track have a score here.
- 1
Gemini 4 Argon ↑3 50%
- 2
Muse Wonder ↑6 16.4%
- 3 GPT-6 Astra ↑3 0%
- 4 Grok 4.7 ↑5 0%
- 5 Claude Opus 5.5 ↓4 0%
- 6 GPT-6.1 Sol ↓1 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
RL data: interface porting
Design RL training data that teaches Qwen3-4B to update callers after a module is rewritten to a new interface. 9 models we track have a score here.
- 1 GPT-6 Astra ↑5 50%
- 2
Gemini 4 Argon ↑2 50%
- 3
Gemini 4 Carbon 34.5%
- 4 Grok 4.7 ↑5 0%
- 5 Claude Opus 5.5 ↓4 0%
- 6 GPT-6.1 Sol ↓1 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9 Claude Fable 5.2 ↓7 0%
RL data: multi-file bug fixing
Design RL training data so that a fixed GRPO run teaches Qwen3-4B to fix bugs in multi-file Python projects with tools. 9 models we track have a score here.
- 1 Kimi K3.1 Pro ↑6 57.6%
- 2 Grok 4.7 ↑7 50%
- 3 Claude Opus 5.5 ↓2 44%
- 4 Claude Fable 5.2 ↓2 15.7%
- 5 GPT-6 Astra ↑1 0%
- 6 GPT-6.1 Sol ↓1 0%
- 7
Gemini 4 Argon ↓3 0%
- 8
Muse Wonder 0%
- 9
Gemini 4 Carbon ↓6 0%
Harness: BFCL v4 multi-turn
Improve a reusable function-calling harness so a fixed worker model succeeds on unseen multi-turn BFCL v4 conversations. 9 models we track have a score here.
- 1 Claude Opus 5.5 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 GPT-6.1 Sol ↑1 0%
- 5
Gemini 4 Argon ↓1 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Harness: SWE-bench Verified
Improve a reusable repository-repair harness so a fixed worker model resolves unseen SWE-bench Verified issues. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 Claude Opus 5.5 ↓3 0%
- 5 GPT-6.1 Sol 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9
Gemini 4 Carbon ↓6 0%
Harness: WebArena-Verified
Improve a reusable web-agent harness so a fixed worker model completes unseen tasks on self-hosted websites. 9 models we track have a score here.
- 1 Claude Opus 5.5 50%
- 2
Gemini 4 Argon ↑2 8.33%
- 3 GPT-6 Astra ↑3 0%
- 4 Grok 4.7 ↑5 0%
- 5 GPT-6.1 Sol 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Bulk inference
Minimize the wall-clock time to generate responses for a large prompt collection with a frozen model, with exact outputs. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 95.8%
- 2 Claude Opus 5.5 ↓1 34.4%
- 3 GPT-6 Astra ↑3 0%
- 4 Grok 4.7 ↑5 0%
- 5 GPT-6.1 Sol 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9
Gemini 4 Carbon ↓6 0%
Cache-aware replica routing
Route requests across two model replicas with local prefix caches to minimize mean response latency. 9 models we track have a score here.
- 1 GPT-6.1 Sol ↑4 68.6%
- 2
Muse Wonder ↑6 50%
- 3 Claude Fable 5.2 ↓1 37.5%
- 4
Gemini 4 Carbon ↓1 14.5%
- 5
Gemini 4 Argon ↓1 14.1%
- 6 Claude Opus 5.5 ↓5 10.6%
- 7 Grok 4.7 ↑2 1.49%
- 8 GPT-6 Astra ↓2 0%
- 9 Kimi K3.1 Pro ↓2 0%
Disaggregated prefill and decode
Split prompt processing and token generation across two GPUs to minimize the duration of a request stream. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 91.2%
- 2 GPT-6.1 Sol ↑3 78.8%
- 3
Gemini 4 Carbon 75.3%
- 4 Claude Opus 5.5 ↓3 43.6%
- 5
Gemini 4 Argon ↓1 40.9%
- 6 GPT-6 Astra 37.5%
- 7 Grok 4.7 ↑2 0%
- 8
Muse Wonder 0%
- 9 Kimi K3.1 Pro ↓2 0%
Schedule prefill and decode
Schedule prompt and generation work on one GPU to maximize output tokens per second within streaming latency limits. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 63.7%
- 2
Gemini 4 Argon ↑2 59.5%
- 3 GPT-6 Astra ↑3 55.8%
- 4 GPT-6.1 Sol ↑1 53.3%
- 5
Muse Wonder ↑3 50%
- 6 Kimi K3.1 Pro ↑1 41.3%
- 7 Grok 4.7 ↑2 20.4%
- 8 Claude Fable 5.2 ↓6 19%
- 9 Claude Opus 5.5 ↓8 0%
Shared-prompt branch generation
Generate many continuations that share a long prompt as fast as possible, with exact greedy outputs. 9 models we track have a score here.
- 1 Claude Fable 5.2 ↑1 88.3%
- 2 Claude Opus 5.5 ↓1 50%
- 3 GPT-6 Astra ↑3 0%
- 4 Grok 4.7 ↑5 0%
- 5 GPT-6.1 Sol 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9
Gemini 4 Carbon ↓6 0%
AdamW parameter update
A fast AdamW optimizer kernel that updates 8 to 15 parameter tensors of up to 8 million elements each. 9 models we track have a score here.
- 1 Claude Opus 5.5 81.2%
- 2 Claude Fable 5.2 77.4%
- 3
Gemini 4 Carbon 52.5%
- 4 Kimi K3.1 Pro ↑3 50%
- 5 GPT-6.1 Sol 49.5%
- 6 GPT-6 Astra 46.1%
- 7
Gemini 4 Argon ↓3 28.6%
- 8 Grok 4.7 ↑1 0%
- 9
Muse Wonder ↓1 0%
Block-masked attention, forward and backward
A fast block-sparse attention kernel for both passes, with 10% to 40% mask density. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 97.3%
- 2 Claude Fable 5.2 62.8%
- 3 Claude Opus 5.5 ↓2 54.1%
- 4
Gemini 4 Argon 36.3%
- 5 GPT-6.1 Sol 32.9%
- 6 GPT-6 Astra 20%
- 7 Grok 4.7 ↑2 0%
- 8
Muse Wonder 0%
- 9 Kimi K3.1 Pro ↓2 0%
Causal attention, forward and backward
A fast causal attention kernel for both the forward and backward passes, across 256 to 8,192 tokens. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 89.9%
- 2
Gemini 4 Argon ↑2 47%
- 3 Claude Fable 5.2 ↓1 29.4%
- 4 Claude Opus 5.5 ↓3 8.23%
- 5 GPT-6 Astra ↑1 0%
- 6 Grok 4.7 ↑3 0%
- 7 GPT-6.1 Sol ↓2 0%
- 8
Muse Wonder 0%
- 9 Kimi K3.1 Pro ↓2 0%
Grouped matrix products
One fast kernel for a group of 2 to 64 independent matrix products of different sizes. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 60.5%
- 2
Gemini 4 Argon ↑2 44.1%
- 3 GPT-6 Astra ↑3 15.9%
- 4 Grok 4.7 ↑5 0%
- 5 Claude Opus 5.5 ↓4 0%
- 6 GPT-6.1 Sol ↓1 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9 Claude Fable 5.2 ↓7 0%
Latent-projected attention
Attention over a projected latent state, for batches of 1 to 32 and contexts up to 32,768 tokens. 9 models we track have a score here.
- 1 Claude Opus 5.5 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 GPT-6.1 Sol ↑1 0%
- 5
Gemini 4 Argon ↓1 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Paged attention
Attention that reads keys and values through a page table, for contexts of 512 to 32,768 tokens. 9 models we track have a score here.
- 1 Claude Opus 5.5 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 GPT-6.1 Sol ↑1 0%
- 5
Gemini 4 Argon ↓1 0%
- 6
Muse Wonder ↑2 0%
- 7 Kimi K3.1 Pro 0%
- 8
Gemini 4 Carbon ↓5 0%
- 9 Claude Fable 5.2 ↓7 0%
Rectangular matrix products
A matrix-product plan that stays fast on highly rectangular shapes from 64 to 8,192. 9 models we track have a score here.
- 1 GPT-6.1 Sol ↑4 96.7%
- 2 GPT-6 Astra ↑4 93.2%
- 3 Claude Opus 5.5 ↓2 91%
- 4
Muse Wonder ↑4 85.8%
- 5
Gemini 4 Argon ↓1 84.1%
- 6
Gemini 4 Carbon ↓3 81.9%
- 7 Claude Fable 5.2 ↓5 63%
- 8 Grok 4.7 ↑1 60.7%
- 9 Kimi K3.1 Pro ↓2 51.2%
Single-query decode attention
Attention for one query over a long context of up to 65,536 tokens. 9 models we track have a score here.
- 1 Claude Opus 5.5 80.4%
- 2 GPT-6 Astra ↑4 68.1%
- 3 Claude Fable 5.2 ↓1 52.7%
- 4
Gemini 4 Carbon ↓1 50.7%
- 5
Muse Wonder ↑3 48.2%
- 6
Gemini 4 Argon ↓2 41.2%
- 7 Kimi K3.1 Pro 39%
- 8 GPT-6.1 Sol ↓3 33.9%
- 9 Grok 4.7 17.2%
Top-k and top-p sampling
A fast sampling kernel with top-k, top-p and temperature over vocabularies of up to 131,072 tokens. 9 models we track have a score here.
- 1
Gemini 4 Carbon ↑2 50%
- 2 GPT-6 Astra ↑4 0%
- 3 Grok 4.7 ↑6 0%
- 4 Claude Opus 5.5 ↓3 0%
- 5 GPT-6.1 Sol 0%
- 6
Gemini 4 Argon ↓2 0%
- 7
Muse Wonder ↑1 0%
- 8 Kimi K3.1 Pro ↓1 0%
- 9 Claude Fable 5.2 ↓7 0%
The best score, 34.3%, costs $1,114 for the full run
Runs on the cost frontier score higher than any run that's cheaper. Gemini 4 Carbon at high comes within 4.5 points of the best for $386 for the full run, 3 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official FrontierRSI leaderboard
| Model | Effort | Harness | Score | Cost of the full run | Output tokens |
|---|---|---|---|---|---|
| Claude Opus 5.5 | max | Claude Code | 34.3% | $1,114 | |
| Claude Fable 5.2 | max | Claude Code | 33% | $2,012 | |
| high | Antigravity CLI | 29.8% | $386 | ||
| high | Antigravity CLI | 23.6% | $398 | ||
| GPT-6.1 Sol | max | Codex | 23.5% | $531 | |
| GPT-6 Astra | max | Codex | 20.9% | $2,829 | |
| Kimi K3.1 Pro | max | Claude Code | 13.6% | $265 | |
| max | Muse Code | 12% | $262 | ||
| Grok 4.7 | xhigh | Grok Build | 8.92% | $1,645 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official FrontierRSI leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official FrontierRSI leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
All 11 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 5 pass
- Pass Independent runs Scores from the official leaderboard (9 entries).
- Pass Models scored 9 models we track.
- Pass Newest score The newest is from November 2026, for Claude Fable 5.2.
- Pass Frontier models 6 of the top 10 have a score, led by Claude Fable 5.2 (#2).
- Pass Headroom The best is 34.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on frontier-rsi.org: 9 rows for models we track at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.