Which LLM hosting providers offer low latency, multiple open-source models and transparent pricing?

As of November 12, 2026: Fireworks AI is the best all-round pick, with a public per-model price table and a Fast mode aimed at 100+ tokens per second on the same endpoint (GLM 5.3 $1.40/$4.40, Kimi K3 $3/$15, gpt-oss-120b $0.15/$0.60 per million tokens standard). Together AI publishes a long serverless price list (Kimi K3.1 Pro $2/$10, GLM-5.3 $1.40/$4.40, gpt-oss-120b $0.15/$0.60). Cerebras is the fastest but serves only two models on its shared tier (gpt-oss-120b at about 3,000 tokens per second). OpenRouter lets you route one API across many hosts and sort by latency or throughput.

What is the fastest host for open-source LLMs?

Cerebras, on the models it serves. Its model catalog lists gpt-oss-120b at about 3,000 output tokens per second and Qwen 3.8 27B at about 1,850, with 65K context on the free tier and 131K (gpt-oss) or 128K (Qwen) on paid. Other models need Cerebras Dedicated Inference. OpenRouter lists Cerebras gpt-oss-120b at $0.35 input and $0.75 output per million tokens.

Which provider has the most open models at published prices?

OpenRouter aggregates the most: for example 23 endpoints for Kimi K3.1 Pro and over 45 for GLM-5.3 on November 12, 2026, each with its own price, quantization and context length. Among direct hosts, Fireworks and Together AI publish per-model serverless prices for the current open-weight leaders, including Kimi, GLM, DeepSeek V4.1 Flash, Qwen 3.8, MiniMax and gpt-oss.

Is a cheap OpenRouter endpoint as good as the model vendor's own API?

Not always. The cheapest endpoints often run 4-bit (FP4 or INT4) quantizations or shorter contexts. On November 12, 2026 Kimi K3.1 Pro listings ranged from $0.24 input on an FP4 host to $2 at Moonshot's own endpoint. Filter by quantization and context in OpenRouter's provider routing, and test quality on your own prompts before switching.

Quick Answer

LLM Hosting: Low Latency, Many Open Models, Clear Prices

Published: November 12, 2026

The short answer

For low latency across many open-source models with prices you can read before you sign up, start with Fireworks AI; add Cerebras for the one or two models where raw speed matters most; route through OpenRouter if you want every host behind one API. As of November 12, 2026:

Provider	Open models (shared tier)	Speed story	Price transparency	Pick it when
Fireworks AI	Broad: Kimi K3, GLM 5.3, DeepSeek V4.1 Flash, Qwen 3.8 Max, MiniMax M3, Muse Spark Open, gpt-oss and more	Fast mode: “100+ tokens per second”, switch per request; Ultra-fast in preview	Public table: Standard, Priority, Batch (−50%), Spot (−30%)	You want one host for most open models with a speed knob
Together AI	Broad: Kimi K3.1 Pro, GLM-5.3, DeepSeek V4.1 Flash 1017, Qwen3.8 Flash, MiniMax M3.1 Flash, gpt-oss	Serverless plus dedicated endpoints	Public serverless + batch price list	You also want fine-tuning or dedicated GPUs from the same vendor
Cerebras	Two: gpt-oss-120b, Qwen 3.8 27B	~3,000 tok/s and ~1,850 tok/s (vendor)	Shared tier on OpenRouter $0.35 / $0.75 for gpt-oss-120b	Speed is the product (voice, live UI)
Groq	gpt-oss family among others	Fast LPU inference	gpt-oss-120b $0.15 / $0.60 (OpenRouter listing)	Fast and cheap on gpt-oss
OpenRouter	Almost everything, from dozens of hosts	Sort by latency or throughput per request	Per-endpoint price, quantization and context shown	You want to compare hosts live and fail over

This page is the “many models, clear prices” spoke of our low latency inference providers hub.

Published prices for the same models (per 1M tokens, input / output)

Standard serverless rates read from each provider on November 12, 2026.

Model	Fireworks	Together AI	Cheapest OpenRouter endpoint
GLM-5.3	$1.40 / $4.40 (Fast $2.10 / $6.60 on OpenRouter)	$1.40 / $4.40	$0.07 / $1.40 (FP4)
Kimi K3.1 Pro	$2.00 / $10.00 (OpenRouter listing)	$2.00 / $10.00	$0.24 / $8.41 (FP4)
DeepSeek V4.1 Flash	$0.30 / $1.20	$0.30 / $1.20 (1017)	$0.02 / $0.40 (FP8)
gpt-oss-120b	$0.15 / $0.60	$0.15 / $0.60	$0.03 / $0.15 (FP4)
GLM-5.3 Flash	$0.15 / $0.50	$0.15 / $0.50	—

Two cautions. First, DeepSeek’s own API charges $0.15 / $0.60 off-peak for V4.1 Flash; Fireworks and Together charge the peak-equivalent $0.30 / $1.20 around the clock. Second, the cheapest OpenRouter endpoints are mostly 4-bit quantizations, so quality and context can differ from the vendor’s endpoint. See current API prices for vendor rates.

Picks by use case

Interactive agents on frontier open models: Fireworks. Fast mode is a per-request switch on the same endpoint, aimed at 100+ tokens per second, and Fireworks says cached input is billed at half price or less with no write fee. Its session affinity keeps an agent on the server holding its context, which cuts time-to-first-token on long runs.

A broad catalog plus training: Together AI. Same headline prices as Fireworks on the overlapping models, and the same vendor sells fine-tuning, post-training and dedicated endpoints, so you can move a hot model to reserved capacity without changing providers.

Lowest latency on one model: Cerebras. At ~3,000 tokens per second, a 300-token answer from gpt-oss-120b streams in about a tenth of a second. The catch is the catalog: two shared models; the rest is dedicated inference.

Comparing hosts without integrating each one: OpenRouter. Its provider object accepts sort by price, throughput or latency, plus preferred_max_latency, preferred_min_throughput, quantizations and max_price, so you can ask for “fastest FP8-or-better host under $X” per request and let it fail over.

How to test before you commit

Measure time to first token and output tokens per second on your real prompt lengths, not a hello-world.
Test at your peak hour; shared tiers slow down under load.
Pin quantization (fp8 or better) when quality matters, and record which host served each request.
Price a full task, not a token: cached input and Batch or Spot modes often matter more than the headline rate. Our cheapest inference provider for open-weight models page covers the price-first view.

Last verified: November 12, 2026.

The short answer

Published prices for the same models (per 1M tokens, input / output)

Picks by use case

How to test before you commit

Sources