Live scoreboard for inference economics
Inference economics for teams shipping at the edge of latency budgets
InferenceIndex is a benchmark front-end for evaluating provider tradeoffs before they become incident review material. The leaderboard reads from Postgres, refreshes hourly, and marks speed metrics pending unless they came from authenticated inference runs.
Fastest p95
Benchmark pending
N/A — real benchmark pending
Lowest cost
OpenRouter
$0.00021 blended / 1K tokens
Mean throughput
N/A
Real inference throughput pending; catalog probes are not counted
API access
Pro API access to real-time LLM benchmark data - prices, latency, quality for every major provider.
Provider leaderboard
Public catalog probes only confirm provider docs are reachable. Latency, throughput, and composite score are published only from authenticated streaming inference runs; otherwise they stay marked as pending.
Feed status
Feed refreshed Sep 26, 8:00 PM UTC
Rank #1
meta-llama/llama-3.3-70b-instruct
Source: provider docs/catalog only; latency and throughput pending
Score
Pending
Latency
N/A — real benchmark pending
Cost
$0.00021
Throughput
N/A — real benchmark pending
Updated
Sep 26, 8:00 PM UTC
Rank #2
llama-3.3-70b-versatile
Source: provider docs/catalog only; latency and throughput pending
Score
Pending
Latency
N/A — real benchmark pending
Cost
$0.00069
Throughput
N/A — real benchmark pending
Updated
Sep 26, 8:00 PM UTC
Rank #3
meta-llama/Llama-3.3-70B-Instruct-Turbo
Source: provider docs/catalog only; latency and throughput pending
Score
Pending
Latency
N/A — real benchmark pending
Cost
$0.00088
Throughput
N/A — real benchmark pending
Updated
Sep 26, 8:00 PM UTC
| Rank | Provider | Latency | Cost / 1K | Throughput | Updated | Score |
|---|---|---|---|---|---|---|
1 | OpenRouter meta-llama/llama-3.3-70b-instruct OpenRouter is currently represented by a public public models API freshness check. Speed metrics are pending until an API key enables real inference. Source: provider docs/catalog only; latency and throughput pending | N/A — real benchmark pending | $0.00021 | N/A — real benchmark pending | Sep 26, 8:00 PM UTC | Pending |
2 | Groq llama-3.3-70b-versatile Groq is currently represented by a public public model docs freshness check. Speed metrics are pending until an API key enables real inference. Source: provider docs/catalog only; latency and throughput pending | N/A — real benchmark pending | $0.00069 | N/A — real benchmark pending | Sep 26, 8:00 PM UTC | Pending |
3 | Together AI meta-llama/Llama-3.3-70B-Instruct-Turbo Together AI is currently represented by a public public serverless model docs freshness check. Speed metrics are pending until an API key enables real inference. Source: provider docs/catalog only; latency and throughput pending | N/A — real benchmark pending | $0.00088 | N/A — real benchmark pending | Sep 26, 8:00 PM UTC | Pending |
Embed this leaderboard
Drop this snippet into any page to show live rankings.
Latency
Teams reach for the fastest endpoints when user-visible interactions start to stack.
Cost
Cost curves matter most when volume spikes and routing policies need immediate tradeoffs.
Throughput
Burst handling decides whether a provider belongs in the hot path or the fallback pool.
FAQ
These are the questions developers search when they need a live LLM benchmark leaderboard, clearer API pricing comparisons, and faster signal on which provider belongs in production.
A fastest-provider claim is pending until authenticated inference runs are available. InferenceIndex still shows documented pricing, but marks latency and throughput as N/A rather than using catalog-response timings as fake benchmarks.
OpenRouter is currently the lowest-cost provider in the InferenceIndex dataset, but the cheapest LLM API in 2026 still depends on your prompt-output mix, model choice, and routing strategy. Benchmarking price alongside latency and throughput is the safest way to avoid optimizing for a misleading headline rate.
There is no single best provider across every workload. Teams usually compare Groq, Together AI, and OpenRouter on latency, model access, reliability, and blended token pricing, which is why InferenceIndex benchmarks OpenRouter, Groq, Together AI side by side instead of declaring one permanent winner.
Developers should compare AI inference providers on measured latency, tokens per second, normalized token pricing, available models, and consistency under load. If real inference measurements are unavailable, a trustworthy leaderboard should say so clearly instead of inventing speed numbers.
Yes. InferenceIndex is a live LLM provider leaderboard built for developers who want a current AI inference comparison across speed, throughput, and price. The homepage is backed by rows stored in Postgres and refreshed on a schedule, with speed fields marked pending until real inference runs exist.