Leading Serverless Inference Providers in 2026, Compared
The short answer
“Serverless inference” means you call a model by API, pay per token, and never see a GPU. The leading serverless inference providers in October 2026 split into three groups:
- Broad catalog of frontier open models: Together AI and Fireworks AI. Both serve Kimi K3, GLM-5.3, DeepSeek V4.1 Flash, Qwen 3.8 and gpt-oss-120b at near-identical list prices, and both let you move to committed capacity without changing APIs.
- Raw speed: Groq and Cerebras. A small shared catalog, but gpt-oss-120b runs at about 500 tokens per second on Groq and about 3,000 on Cerebras.
- Lowest price: DeepInfra. Often a few percent to half below the others on the same model.
- Edge and Cloudflare-native apps: Cloudflare Workers AI. Billed in Neurons ($0.011 per 1,000), with 10,000 free Neurons per day on every plan.
If you want one bill across all of them, OpenRouter passes provider prices through without markup and charges a 5.5% fee when you buy credits.
This page is the serverless spoke of our best LLM inference providers ranking.
The comparison
Prices are USD per million tokens (input / cached input / output), read from each vendor’s own pricing page on October 12, 2026.
| Provider | gpt-oss-120b | Kimi K3 | GLM-5.3 | DeepSeek V4.1 Flash | Capacity upgrade path |
|---|---|---|---|---|---|
| Together AI | $0.15 / — / $0.60 | $3.00 / $0.27 / $15 | $1.40 / $0.26 / $4.40 | $0.30 / $0.006 / $1.20 | Provisioned throughput units; dedicated H100 $5.49/h, B200 $8.99/h |
| Fireworks AI | $0.15 / $0.015 / $0.60 | $3.00 / $0.30 / $15 (Fast $4.50 / $22.50) | $1.40 / $0.26 / $4.40 | $0.30 / $0.006 / $1.20 | Priority tier (~25% more), Fast variants, batch at 50% |
| Groq | $0.15 / $0.075 / $0.60, ~500 tok/s | — | — | — | Developer plan: higher limits, Batch and Flex |
| Cerebras | $0.35 / — / $0.75, ~3,000 tok/s | — | — | — | Dedicated Inference with production SLAs |
| DeepInfra | — | $2.85 / $0.285 / $14.25 | — | — | Auto-scaling; 200 concurrent requests per account by default |
| Cloudflare Workers AI | $0.35 / — / $0.75 | — | $1.40 / $0.26 / $4.40 | — | Workers Paid plan or prepaid AI Gateway credits |
DeepSeek V4.1 Flash’s own API charges $0.15/$0.60 off-peak and twice that at peak; the $0.30/$1.20 above is the reseller rate, which is the same at all hours. Closed-model prices are on our current API prices page.
Picks by situation
Most production apps on open models: Together AI or Fireworks AI. They carry the same frontier checkpoints within days of release and price them identically, so choose on the upgrade path. Fireworks’ Priority tier keeps the per-token model and simply costs more; Together’s provisioned throughput reserves fixed capacity measured in PTUs. Both have batch pricing for offline jobs.
Voice agents, autocomplete, live UI: Groq or Cerebras. When the reply has to start and finish inside a second, output speed matters more than catalog. Cerebras is the fastest; Groq is cheaper on the same model ($0.15/$0.60 against $0.35/$0.75) and charges half price for cached input. Both serve only a few shared models — gpt-oss-120b and Qwen 3.8 27B on Cerebras’ shared tier — so larger models need a dedicated contract.
Lowest bill on big open models: DeepInfra. Kimi K3 at $2.85/$14.25 is 5% under Together and Fireworks; on some older checkpoints the gap is much wider. Check quantization on the model page before you switch: the cheapest listing is not always the same precision.
Apps already on Cloudflare: Workers AI. Inference runs next to your Worker, the first 10,000 Neurons each day are free, and GLM-5.3 costs the same as on Together. Frontier models (GLM-5.3, Kimi K2.6, DeepSeek V4) require a paid billing method.
Switching between providers: OpenRouter. One key, provider fallbacks and pooled uptime; no inference markup, 5.5% on credit purchases (minimum $0.80), and bring-your-own-key free up to $25,000 of monthly list-price usage.
What to check before you commit
- Rate limits are per organization, not per key — Groq documents RPM, tokens-per-minute and daily caps per model, and a 429 with
retry-afterwhen you hit one. - Cached-input pricing decides agent costs. Fireworks charges $0.006 for cached DeepSeek V4.1 Flash input, 2% of the uncached price; providers without a cached rate cost far more on long, repeated prompts.
- Regional variants cost more. Fireworks’ US-hosted Kimi K3 and GLM-5.3 are 50% above the default endpoints.
- Speed claims are per model. Cerebras’ context window is 65K on the free tier and 131K on paid for gpt-oss-120b.
Related: best dedicated inference endpoints, cheapest inference provider for open-weight models and best inference providers for side projects.
Last verified: October 12, 2026.