Artificial Analysis is often the first stop when you compare LLM APIs, because it times real requests to each provider. But it is one lens. Other sites rank by human votes, by benchmark scores, by usage or by price, and each answers a different question. This guide sets out what each one measures, how, what data you can take away, and when to use it. Every fact below was read from the site’s own pages on 15 November 2026.
A disclosure first: BroadRouter is BroadNet’s own website. It sends developers to BroadNet Inference and publishes price data. It does not benchmark models, and it appears in this list only for prices.
What does Artificial Analysis do well?
Artificial Analysis describes itself as “Independent analysis of AI”. Its methodology page explains three things it does that few others do:
- Speed per provider endpoint. Output speed is “output tokens per second over each request’s full duration”, including reasoning tokens, averaged “over the workloads we measured in the past 5 days”. Because one model “may have multiple endpoints across different providers”, you can see the same model run faster on one host than another.
- Latency you can compare. It reports time to first token, time to first answer token (which includes reasoning, so reasoning and non-reasoning models compare fairly) and end-to-end time.
- Its own evaluations. Intelligence Index v5.0.1 combines nine evaluations in four equally weighted groups (knowledge work, coding, science, general). Several are its own benchmarks or its own runs of others’ tasks, with stated repeat counts.
Its provider leaderboard covers “over 500 AI Model endpoints” with cost per task, time per task, first-answer-token latency, error rate, context window and licence. Prices are “the current prices listed by providers”, and its blended price uses a 7:2:1 ratio of cache-hit, input and output tokens. Change that mix in your head if your workload is output-heavy.
If you need measured speed and error rates by provider, it is hard to replace. People look for alternatives when they want a different signal: human preference, a specific domain, downloadable raw data, or a price list with a source for every row.
Which Artificial Analysis alternatives are there?
| Site | What it measures | Method | Data you can take away | Best for |
|---|---|---|---|---|
| Artificial Analysis | Intelligence index, output speed, latency, error rate, price per endpoint | Runs its own evaluations and timed requests; 5-day rolling speed window | Free API, attribution required | Speed and latency by provider |
| Arena (formerly LMArena) | Human preference | Two anonymous models answer, a person votes, names revealed after | Open preference datasets on Hugging Face | How answers feel to people |
| Epoch AI Benchmarking Hub | Benchmark scores over time | Runs some benchmarks itself with Inspect; collects others from official leaderboards | CSV and a Python client; its data is CC BY | Trends and raw scores |
| Vals AI | Domain accuracy, cost and latency | 42 benchmarks, some private or held out | Not stated on the benchmarks page | Legal, finance, tax and other domains |
| OpenRouter rankings | Usage | Tokens processed through OpenRouter, in daily UTC buckets | Public models API with prices | What developers actually run |
| Open LLM Leaderboard (Hugging Face) | Open-weight model scores | Six benchmarks in the Eleuther AI harness | Results and requests datasets | Comparing open checkpoints |
| BroadRouter price data (BroadNet’s site) | Published prices only | Read from each provider’s own pricing page, with URL, date and quote | CSV and JSON, CC BY 4.0 for the compilation | Checking a price at its source |
How is Arena different?
Arena (lmarena.ai now redirects to arena.ai) asks real people to compare two anonymous answers and pick one. Model names appear only after the vote. That measures preference, which mixes accuracy with tone, length and formatting. Its leaderboard has separate boards, including Text, WebDev, Vision, Document and Search, and its Pareto view shows a cost per task. Arena says it has open-sourced its preference data on Hugging Face, which is useful if you want to study votes rather than read a rank.
Use it to see how a chat answer lands with people. Do not use it to choose a provider for an API: it ranks models, not hosts, and says nothing about speed on a given endpoint.
When are Epoch AI and Vals AI the better fit?
Epoch AI keeps a database of benchmark results, tracking 99 benchmarks and 428 models when we checked. Some results it runs itself with the Inspect framework; the rest come from official leaderboards kept by benchmark creators. You can download the data as CSV or use its Python client. Its own data is free to reuse with credit under CC BY; externally sourced results keep their original licences. Pick Epoch when you want raw scores or a trend line, not a single index.
Vals AI runs 42 benchmarks across legal, finance, healthcare, coding, maths and more, and some sets are private or held out, which reduces the risk that a model was trained on the test. Its leaderboards compare accuracy, cost and latency, and its Inference Provider Index compares providers on accuracy, cost and speed. Its composite Vals Index weights sectors by their share of US GDP. Pick Vals when your task looks like one of its domains.
What do OpenRouter rankings actually show?
Usage. Each model is ranked by “tokens it processed through the OpenRouter API”, prompt and completion together, over trailing one, seven or thirty days. The page states plainly that the rankings “do not rank models by accuracy, reasoning ability, or benchmark performance”. A model that writes long answers or runs in long agent loops climbs the table. Read it as a popularity signal and a list of models worth testing. OpenRouter also has a public models endpoint with per-model prices, which covers what OpenRouter charges, not what each host charges directly.
Is the Hugging Face Open LLM Leaderboard still worth using?
For open-weight checkpoints, yes, with care. Hugging Face’s documentation says the first version was archived in June 2024 and replaced. The current version scores models on IFEval, BBH, MATH level 5, GPQA, MuSR and MMLU-Pro in the Eleuther AI evaluation harness, and publishes results as datasets. It evaluates model files, not hosted APIs, so it cannot tell you about price, speed or how a provider serves the model.
Where do you compare Arabic models?
General leaderboards say little about Arabic, and Arena’s visible filters do not include a language. Two Hugging Face Spaces fill part of the gap:
- The Open Arabic LLM Leaderboard, run by the OALL organisation, for Arabic text.
- The Open Universal Arabic ASR Leaderboard, described as “a benchmark for open-source multi-dialect Arabic ASR models”, for speech recognition. BroadNet reports its Transcribe Arabic model at #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026). That is BroadNet’s figure, and the board changes as models are added.
Read speech results by dialect, not just the average. Arabic ASR word error rate explained shows how.
Where does BroadRouter’s price data fit?
BroadRouter publishes three price resources: the LLM API pricing pages per model, the price index with medians and spreads, and the raw provider price database. Every row is read from the provider’s own pricing page and stored with the source URL, the date and a verbatim quote, and a weekly script re-fetches the sources. No aggregator feed is used.
What it does not do matters as much:
- It measures no speed, latency, uptime or quality. For those, use Artificial Analysis, Arena, Epoch AI or Vals AI.
- It holds third-party prices only. BroadNet’s own prices are in its pricing.json and on the pricing page.
- It is BroadNet’s site, so treat it as a vendor’s compilation that you can check row by row, not as a neutral judge.
How should you combine these sites?
- Shortlist on quality. Use Arena for chat feel and Epoch AI or Vals AI for task scores in your domain. Keep three to five models.
- Check speed on the host you will use. Artificial Analysis shows per-endpoint speed and latency. LLM latency explained covers which number matters for your app.
- Check the price at the source. Use a price database to find the providers, then open each provider’s own page. Watch cached-input and output prices, which move the bill more than input for many workloads.
- Test on your own data. Run 50 to 100 real inputs through each finalist and score them. A benchmark predicts your results only as far as its tasks look like yours. How to evaluate an LLM before switching gives the method, and the 20-point provider checklist covers limits, privacy and support.
Where BroadNet Inference fits
BroadNet Inference is an OpenAI-compatible API, not a leaderboard. It serves a small set of models: DeepSeek V4.1 Flash (partner-hosted), GLM 5.3, Qwen3.8 27B, Gemma 4 VL 26B and two speech-to-text models. Tool calling and structured output are documented for GLM 5.3 and Qwen3.8 27B only. Self-serve accounts have no SLA, and there is no batch API or embeddings. If one of these models makes your shortlist, the $5 starting credit covers Gemma 4 VL 26B, both Transcribe models and Gemma hourly keys, so you can run step 4 on your own inputs before paying; DeepSeek, GLM and Qwen need a top-up. Prompts and answers are not retained. Compare the models or create an account.
