Top 50 AI Models Leaderboard
Compare the top 50 Large Language Models using published reasoning, coding, and agent benchmarks alongside reported throughput (tps) and API pricing.
Daily metadata checks; updates publish after review•85 models tracked•How we verify scores
| Rank | Model & Provider | License | RankLLMs | Reason | Code | Agent | Arena | Params | Context | Speed | Price |
|---|---|---|---|---|---|---|---|---|---|---|---|
| #1 | GPT-6 AstraNEW OpenAI | Proprietary | 68.5 | 96.0 | 74.1 | 72.6 | 2,250 | Proprietary | 1.1M | 59 tps | $10.00 |
| #2 | Anthropic | Proprietary | 66.8 | 88.5 | 86.2 | 78.4 | 2,750 | Proprietary | 1M | 42 tps | $5.00 |
| #3 | Anthropic | Proprietary | 64.2 | 93.7 | 67.4 | 75.0 | 2,045 | Proprietary | 1M | 65 tps | $10.00 |
| #4 | Anthropic | Proprietary | 62.0 | 67.5 | 84.5 | 86.0 | 1,932 | Proprietary | 1M | 110 tps | $10.00 |
| #5 | GPT-6 SolNEW OpenAI | Proprietary | 60.5 | 78.5 | 84.8 | 65.2 | 1,950 | Proprietary | 1.1M | 126 tps | $2.00 |
| #6 | Claude Mythos PreviewUNRELEASED Anthropic | Proprietary | 58.2 | 65.8 | 82.2 | 85.4 | 1,743 | Proprietary | 1M | 50 tps | $0.00 |
| #7 | Anthropic | Proprietary | 57.5 | 64.5 | 81.8 | 65.0 | 2,668 | Proprietary | 1M | 68 tps | $5.00 |
| #8 | Meta | Proprietary | 57.4 | 63.5 | 75.4 | 66.9 | 1,754 | Proprietary | 1M | 221 tps | $1.25 |
| #9 | OpenAI | Proprietary | 57.2 | 65.5 | 81.2 | 58.5 | 1,730 | Proprietary | 1.1M | 89 tps | $4.00 |
| #10 | Google | Proprietary | 56.8 | 62.5 | 73.7 | 59.0 | 1,545 | Proprietary | 1M | 297 tps | $0.75 |
| #11 | Xiaomi | Proprietary | 56.5 | 65.2 | 78.4 | 42.5 | 1,745 | Proprietary | 1M | 55 tps | $0.44 |
| #12 | Moonshot AI | Open Source | 56.0 | 64.2 | 80.5 | 56.4 | 1,682 | 2.8T | 1M | 52 tps | Free/Open Source |
| #13 | Qwen / Alibaba | Proprietary | 55.8 | 62.8 | 79.4 | 56.2 | 1,724 | 2.4T | 1M | 39 tps | $2.00 |
| #14 | Zhipu AI | Open Source | 55.8 | 60.5 | 78.2 | 52.5 | 1,773 | 320B | 1M | 61 tps | Free/Open Source |
| #15 | xAI | Proprietary | 55.6 | 63.2 | 78.5 | 58.0 | 1,753 | Proprietary | 500K | 59 tps | $2.00 |
| #16 | GLM-5.3NEW Zhipu AI | Open Source | 55.4 | 62.5 | 77.8 | 51.0 | 1,769 | 753B | 1M | 57 tps | $1.40 |
| #17 | Hy4 PreviewNEW Tencent | Open Source | 55.2 | 58.5 | 77.0 | 49.0 | - | 770B | 1M | 60 tps | Free/Open Source |
| #18 | OpenAI | Proprietary | 55.0 | 59.2 | 77.4 | 52.0 | 1,523 | Proprietary | 1.1M | 117 tps | $2.00 |
| #19 | Anthropic | Proprietary | 54.8 | 59.5 | 76.5 | 83.4 | 1,890 | Proprietary | 1M | 57 tps | $5.00 |
| #20 | Anthropic | Proprietary | 54.8 | 58.6 | 75.8 | 81.2 | 1,618 | Proprietary | 1M | 77 tps | $2.00 |
| #21 | DeepSeek | Open Source | 54.4 | 62.5 | 81.0 | 55.0 | - | 1.6T (49B active) | 1M | 176 tps | Free/Open Source |
| #22 | DeepSeek | Open Source | 54.2 | 58.4 | 80.6 | 51.8 | 1,554 | 1.6T (49B active) | 1M | 50 tps | $1.74 |
| #23 | Google | Proprietary | 54.2 | 61.5 | 78.6 | 55.0 | 1,588 | Proprietary | 1M | 169 tps | $0.75 |
| #24 | Anthropic | Proprietary | 54.0 | 57.2 | 72.4 | 47.2 | 1,619 | Proprietary | 1M | 35 tps | $5.00 |
| #25 | OpenAI | Proprietary | 53.8 | 58.2 | 72.5 | 54.6 | 1,674 | Proprietary | 1.1M | 14 tps | $2.50 |
| #26 | Meta | Proprietary | 53.5 | 57.0 | 59.3 | 48.5 | 1,631 | Proprietary | 1M | 143 tps | $1.25 |
| #27 | Meta | Proprietary | 53.2 | 62.1 | 61.5 | 54.7 | 1,371 | Proprietary | 1M | 140 tps | $1.25 |
| #28 | Google | Proprietary | 52.4 | 58.6 | 74.2 | 48.8 | 1,314 | Proprietary | 1M | 151 tps | $1.25 |
| #29 | Xiaomi | Open Source | 52.4 | 54.8 | 67.2 | 32.5 | 1,560 | Open Weights | 1M | 185 tps | Free/Open Source |
| #30 | Moonshot AI | Open Source | 52.2 | 54.0 | 73.2 | 50.0 | 1,482 | 1.0T | 262K | 165 tps | $0.95 |
| #31 | OpenAI | Proprietary | 52.2 | 56.4 | 75.6 | 78.7 | 1,769 | Proprietary | 1.1M | 33 tps | $5.00 |
| #32 | DeepSeek | Open Source | 52.0 | 54.2 | 76.8 | 47.8 | 1,395 | 284B (13B active) | 1M | 54 tps | $0.14 |
| #33 | DeepSeek | Open Source | 51.8 | 55.1 | 72.5 | 44.0 | 1,675 | Proprietary | 1M | 50 tps | $0.15 |
| #34 | GPT-6 LunaNEW OpenAI | Proprietary | 51.2 | 62.0 | 72.4 | 42.1 | 1,350 | Proprietary | 1.1M | 157 tps | $0.10 |
| #35 | xAI | Proprietary | 51.2 | 56.0 | 68.2 | 47.1 | 1,526 | Proprietary | 500K | 39 tps | $2.00 |
| #36 | Qwen / Alibaba | Open Source | 50.5 | 52.4 | 68.2 | 42.0 | - | 125B | 262K | 50 tps | Free/Open Source |
| #37 | Anthropic | Proprietary | 49.8 | 48.8 | 64.5 | 78.5 | 1,395 | Proprietary | 1M | 82 tps | $3.00 |
| #38 | OpenAI | Proprietary | 49.8 | 52.5 | 68.5 | 46.0 | 1,166 | Proprietary | 1.1M | 157 tps | $0.20 |
| #39 | Zhipu AI | Open Source | 49.8 | 52.0 | 69.5 | 42.0 | 1,535 | 754B | 200K | 50 tps | Free/Open Source |
| #40 | Google | Proprietary | 48.5 | 52.0 | 66.8 | 38.5 | 1,538 | Proprietary | 1M | 165 tps | $0.75 |
| #41 | Qwen / Alibaba | Proprietary | 48.4 | 50.2 | 65.4 | 38.0 | 1,538 | Proprietary | 1M | 32 tps | $2.50 |
| #42 | Zhipu AI | Open Source | 48.2 | 54.7 | 46.2 | 26.2 | 1,508 | 753B | 1M | 126 tps | Free/Open Source |
| #43 | ByteDance | Proprietary | 47.0 | 47.9 | 36.0 | 33.1 | - | Proprietary | 256K | 50 tps | $0.89 |
| #44 | DeepSeek | Open Source | 46.1 | 43.3 | 36.2 | 33.2 | 2,407 | 284B (13B active) | 1M | 227 tps | Free/Open Source |
| #45 | Qwen3.8-27BNEW Qwen / Alibaba | Open Source | 46.1 | 45.6 | 34.5 | 32.9 | - | 27.8B | 262K | 41 tps | Free/Open Source |
| #46 | OpenAI | Proprietary | 45.5 | 45.4 | 48.6 | 28.5 | - | Proprietary | 1.1M | 85 tps | $30.00 |
| #47 | Meta | Proprietary | 45.2 | 50.4 | 55.0 | 34.8 | - | Proprietary | 1M | 160 tps | $0.10 |
| #48 | Anthropic | Proprietary | 44.8 | 46.5 | 39.1 | 33.1 | 1,926 | Proprietary | 1M | 125 tps | $5.00 |
| #49 | ByteDance | Proprietary | 44.0 | 45.4 | 33.2 | 30.4 | - | Proprietary | 256K | 50 tps | $0.45 |
| #50 | Google | Proprietary | 43.6 | 44.4 | 33.8 | 31.5 | 1,708 | Proprietary | 1M | 376 tps | $1.50 |
No models match your filter criteria
Try adjusting your search query or selecting "All Models".
How to Read This LLM Leaderboard
Every column in the table, explained in plain English. Full sources and scoring weights live in our methodology.
- Overall
- The composite 0-100 score blending reasoning, coding, agentic, and arena performance into one comparable number.
- Reasoning
- Formal logic and science questions: GPQA Diamond (PhD-level) and MATH-500 competition mathematics.
- Coding
- SWE-bench Verified: resolving real GitHub issues end to end - the closest proxy for day-to-day engineering work.
- Agent
- Autonomous computer and tool use: OSWorld desktop tasks and multi-step tool-invocation pipelines.
- Code Arena Elo
- Blind head-to-head coding battles rated by human voters - the community verdict on practical quality.
- Params & Context
- Parameter count where disclosed, and the maximum input window (K = thousand, M = million tokens) per session.
- Speed (tps)
- Reported throughput and first-token latency are upstream model specifications, not controlled RankLLMs tests. Provider, region, plan, and inference settings affect results.
- Price $/M
- Blended cost per 1M tokens at a 3:1 input-to-output workload ratio, sourced from official provider pricing pages.
- Verification
- Benchmark results come from published sources when available; RankLLMs does not rerun every benchmark. See our methodology for source notes and limitations.
Not sure which model to pick?
Start with our data-driven picks in Best LLM for Coding, compare the open-weights field in the Open-Source LLM Guide, or run any two models side by side in the comparison engine.
Latest Benchmark Guides & Model Deep Dives
Recent independent evaluations, provider architecture reviews, and pricing breakdowns.
AI Leaderboard Evaluation Methodology & FAQ
How RankLLMs benchmarks, evaluates, and verifies Large Language Models.
- 01
What does the Overall LLM Score measure?
The Overall LLM score synthesizes normalized performance across formal reasoning (GPQA Diamond, MATH-500), SWE-bench coding benchmarks, agentic tool invocation, and human preference Code Arena Elo rankings.
- 02
How is Coding & Code Arena measured?
Coding performance evaluates real-world repository bug fixing and automated feature generation via SWE-bench Verified, combined with crowdsourced blind pairwise head-to-head coding battles in Code Arena.
- 03
How are Speed & Tokens Per Second measured?
Throughput and first-token latency are reported model specifications from upstream sources, not a controlled RankLLMs benchmark. Provider, region, plan, and inference settings affect results, so treat them as directional.
- 04
What is the Pricing $/M Tokens Metric?
Pricing $/M represents the blended API inference expenditure per 1 Million tokens (calculated with standard 3:1 input-to-output token workload ratios) directly sourced from official provider API pricing tiers.
- 05
How often is the leaderboard updated?
A scheduled workflow checks model specifications and pricing daily; changes go live after review and merge. Benchmark results change as published source results are reviewed, so their update cadence varies. See our methodology for details. Material formula changes are announced on the changelog.
- 06
SWE-bench Verified vs SWE-bench Pro - which matters?
Verified is the human-validated public subset; Pro adds harder, privately held tasks that resist training-data contamination. Only compare scores within the same version - our SWE-bench Pro guide explains the distinction in depth.
- 07
Which is the best open-source LLM right now?
Kimi K3 is the top open-weights model overall, and DeepSeek-V4-Pro-0813 leads open-weights coding (81.0 SWE-bench) at 176 tokens per second. The full landscape, licenses, and self-hosting math are in the Open-Source LLM Guide.
- 08
What is the best LLM for coding?
Claude Fable 5 leads on raw capability (84.5 SWE-bench, 88.0 Terminal Bench); Gemini 3.7 Flash and DeepSeek-V4-Pro win on cost per solved task. The full breakdown with routing strategy is in Best LLM for Coding.
- 09
Can I use this data in my own project?
You can browse every included model in the full model directory and see how scoring works in the methodology.