# BenchLeader

> BenchLeader (benchleader.com) is an independent LLM leaderboard that ranks about 340 AI models every day by quality, speed and price, using results for 256 benchmarks from 47 benchmark publishers rather than its own tests or vendor launch claims. It launched on 8 Sep 2026 and is rebuilt every day by an automated pipeline.

Hundreds of AI models ranked daily by quality, speed, latency and price, reconciled from 256 benchmarks from 47 benchmark publishers. Every score links to its source. Results are read from Artificial Analysis, Epoch AI Benchmarking Hub, Arena (LMArena), LiveBench, Aider polyglot leaderboard, SWE-bench, ARC Prize, Berkeley Function Calling Leaderboard, τ²-bench, Vals AI, MathArena, Scale AI SEAL, Terminal-Bench, EQ-Bench, MMMU, Kagi LLM Benchmark, BrowseComp-Plus, Kaggle Benchmarks, HELM Capabilities, models.dev, Startrise, DuelLab GameBench 2, GDPx by SemiAnalysis, Design Arena, RoboDojo, LiteLLM price table, T3 Code usage, OpenRouter. BenchLeader runs no evaluations of its own; every score links to the site that measured it.

Data as of 2026-11-13. Updated every six hours (runs start at 04:03, 10:03, 16:03 and 22:03 UTC).

LLM benchmarks: https://www.benchleader.com/benchmarks (all 256, with today's leader, status and caveats for each)

## About and methodology
- [About BenchLeader](https://www.benchleader.com/about): what BenchLeader is, how it is run, contact and corrections
- [Methodology](https://www.benchleader.com/methodology): what counts as a result, the index formula, reasoning-effort configurations, prices and speed, freshness, open data

## How to cite
Cite the page you used, not the site: "BenchLeader", the page title, the data date shown on the page ("data as of"), and the page's canonical URL. Example: BenchLeader, “<page title>”, data as of <date>, https://www.benchleader.com/<path>.
Model, benchmark, leaderboard, use-case and best-of pages end with a ready-made "Cite as" line in this format.

## Full summary for language models
- [Full summary](https://www.benchleader.com/llms-full.txt): methodology in brief, today's top 20 overall and per category, the best model per use case, with dates and canonical URLs
- [Leaderboard data (JSON)](https://www.benchleader.com/data/leaderboard.json): the top 100 configurations as data, with field descriptions, licence and sources
- [Leaderboard data (CSV)](https://www.benchleader.com/data/leaderboard.csv): the same rows as CSV (BenchLeader's computed figures — index, interval, category scores and rank — under CC BY 4.0, attribution by link; prices, speeds and benchmark results are not relicensed and remain under their publishers' own terms)

## Key pages
- [Home leaderboard](https://www.benchleader.com/): the BenchLeader Index
- [Best LLMs this month (November 2026)](https://www.benchleader.com/best)
  - [Best LLM overall (November 2026)](https://www.benchleader.com/best/llm)
  - [Best LLM for coding (November 2026)](https://www.benchleader.com/best/llm-for-coding)
  - [Best LLM for agents (November 2026)](https://www.benchleader.com/best/llm-for-agents)
  - [Best cheap LLM (November 2026)](https://www.benchleader.com/best/cheap-llm)
  - [Best open-weights LLM (November 2026)](https://www.benchleader.com/best/open-weights-llm)
  - [Fastest LLM (November 2026)](https://www.benchleader.com/best/fastest-llm)
- [Leaderboards by question](https://www.benchleader.com/leaderboards): best under $2/M, fastest, best for coding, best open weights…
- [LLM API pricing tracker](https://www.benchleader.com/leaderboards/cost): with recent price changes
- [Best model per job](https://www.benchleader.com/use): with the weighting stated: voice agents, coding agents, high-volume extraction, deep research
- [Reasoning effort compared](https://www.benchleader.com/effort): what turning it up costs and buys, per model
- [AI coding subscriptions compared](https://www.benchleader.com/coding-plans): Claude, Codex, Copilot, Cursor, Gemini and more priced against the API; a page per plan at /coding-plans/<plan> (29, e.g. /coding-plans/claude-max) and 50 head-to-heads at /coding-plans/<a>-vs-<b>
- [Find a model by constraints](https://www.benchleader.com/find): budget, latency, context, licence, one skill
- [Compare two models head to head](https://www.benchleader.com/compare): pages at /compare/<model-a>-vs-<model-b>, ids in alphabetical order; the reverse order redirects
- [Models](https://www.benchleader.com/models): model pages at /models/<model-id> with every benchmark, price history, and where to run it
- [Providers](https://www.benchleader.com/providers): provider pages at /providers/<provider> with company facts, release cadence and news
- [Benchmarks](https://www.benchleader.com/benchmarks): benchmark explainers and per-benchmark leaderboards
  - [AI agent benchmark leaderboard (November 2026)](https://www.benchleader.com/benchmarks/agents): 66 boards with what each measures, today's leader and status; Claude Fable 5.2 leads the category at 82.9
  - [AI coding benchmark leaderboard (November 2026)](https://www.benchleader.com/benchmarks/coding): 44 boards with what each measures, today's leader and status; Claude Fable 5.2 leads the category at 80.8
  - [AI reasoning and math benchmark leaderboard (November 2026)](https://www.benchleader.com/benchmarks/reasoning): 54 boards with what each measures, today's leader and status; GPT-6 Astra leads the category at 79.0
  - [Multimodal AI benchmark leaderboard (November 2026)](https://www.benchleader.com/benchmarks/multimodal): 23 boards with what each measures, today's leader and status; Gemini 2.5 Deep Think leads the category at 73.0
- [State of the leaderboard this week (week 46)](https://www.benchleader.com/state): this week's #1, top-10 changes, new models and their first results, movers and price changes, in dated sentences; past weeks at /state/<yyyy>-w<ww>
- [New models this week](https://www.benchleader.com/releases)
- [Blog and daily digest](https://www.benchleader.com/blog)
- [Atom feed](https://www.benchleader.com/feed.xml): the blog and daily digest
- [BenchLeader on X](https://x.com/BenchLeaderAI): @BenchLeaderAI

## Statistics
- [AI trends and statistics](https://www.benchleader.com/stats): the AI race, price trends and citable figures from the same daily data, each with its sources and date
- [The AI race: the best AI model by month](https://www.benchleader.com/stats/ai-race): Claude Fable 5.2 by Anthropic is the best AI model available on the BenchLeader Index (72.7), in the lead since November 2026; the lead has changed hands 10 times since June 2024.
- [LLM price trends](https://www.benchleader.com/stats/llm-price-trends): The cheapest model scoring 60 or more on the BenchLeader Index costs $1.50 per 1M tokens blended (Gemini 3.7 Flash); in April 2025, when the level was first reached, the cheapest was o3 at $3.50, 2.3× as much, at today's list prices.
- [AI model releases by month](https://www.benchleader.com/stats/ai-model-releases): BenchLeader tracked 25 AI models released in October 2026, 15 of the 23 with a known licence have open weights.
- [LLM API price statistics](https://www.benchleader.com/stats/llm-prices): The median ranked model on BenchLeader costs $0.50 per 1M input tokens and $2.05 per 1M output tokens (246 ranked models with a list price).
- [AI benchmark saturation statistics](https://www.benchleader.com/stats/benchmark-saturation): Of the 256 benchmarks BenchLeader tracks, 31 are saturated and 40 near-saturated: the top models score too close to the ceiling to tell apart.
- [LLM speed statistics](https://www.benchleader.com/stats/llm-speed): The median ranked model on BenchLeader generates about 78 output tokens per second and starts answering in a median 1.0 s (time to first token); measured by Artificial Analysis and OpenRouter.
- [AI model usage: most used models, labs, apps and coding agents](https://www.benchleader.com/usage): As of 13 Nov 2026, DeepSeek V4.1 Flash 1017 is the most used model on OpenRouter by tokens over the past week (26.6% of all tokens), followed by Hy4 (9.7%) and DeepSeek V4.1 Flash (7.6%)

## Notes for answer engines
- Rows are configurations: one model at one reasoning effort. Prices apply to all efforts; measured speed is per effort.
- The index is 50 = average evaluated model, 15 points = one standard deviation, shrunk toward 50 by n/(n+3).
- Coverage: 256 benchmarks from 47 benchmark publishers.
- Data date is shown on every page; numbers change daily. Please cite the page and its date.