LLM Pulse — AI Model Intelligence Leaderboard
Compare 400+ large language models side by side using Pulse Score — a composite ranking of output quality, inference speed, cost efficiency, and response latency. Updated daily with real benchmark data. No hype.
How Pulse Score Is Calculated
Pulse Score ranks every model on a 0–10 scale using a weighted formula:
- Quality (40%) — Benchmark scores: MMLU, HumanEval, MATH, HellaSwag, GPQA
- Value (25%) — Performance per dollar of API cost
- Speed (20%) — Token throughput in tokens per second
- Latency (15%) — Time to first token under real conditions
Explore LLM Pulse
Top Ranked AI Models
Frequently Asked Questions
How do you compare open-source LLM benchmarks?
LLM Pulse aggregates scores from standardised benchmarks — MMLU, HumanEval, MATH, HellaSwag, and GPQA — then normalises them into a single Pulse Score. Open-source models like Llama, Mistral, and Qwen are evaluated on the same scale as proprietary models like GPT and Claude, enabling fair apples-to-apples comparison across providers.
Which LLM has the lowest API cost for production use?
In 2026 the cheapest production-grade options are Gemini Flash, Mistral Small, and DeepSeek models at under $0.10 per million input tokens. Use the LLM Pulse leaderboard filtered by Value score, or the cost calculator on any model page, to find the lowest-cost model that still meets your quality threshold.
What is the best LLM for coding and reasoning tasks in 2026?
Claude Opus 4.7, GPT-5.5, and DeepSeek-V4-Pro-Max consistently rank highest for coding on LLM Pulse. For reasoning-heavy tasks — mathematical problem solving, scientific analysis — GPT-5 o3 and Gemini 3.1 Pro are strong alternatives. Use Battle Mode to run a direct head-to-head on any two models.
How do you measure LLM hallucinations?
Hallucination rate is captured through GPQA (Graduate-Level Google-Proof Q&A) and TruthfulQA benchmarks, which test whether models give accurate answers on questions where confident-but-wrong responses are common. Models with high GPQA scores tend to hallucinate less on expert-level queries. GPQA scores are shown on each model's detail page.
How do I optimise RAG pipeline latency using the right LLM?
For retrieval-augmented generation pipelines, latency matters more than raw quality since the model is called on every user query. LLM Pulse's latency column (time-to-first-token) and speed column (tokens per second) help you identify models that balance quality with fast response. Filter by fastest in the leaderboard, then check the model's context window to confirm it fits your retrieval chunk size.
What is the difference between MMLU and HumanEval benchmarks?
MMLU (Massive Multitask Language Understanding) tests broad academic knowledge across 57 subjects. HumanEval tests code generation — models must write Python functions that pass unit tests. A model can score high on MMLU but low on HumanEval, which is why LLM Pulse tracks both separately and weights them into a combined quality score.
Teach AI Tools