Seminal AI
§6

Inference platforms for open models

This category covers the hosted services that run open-weight models (DeepSeek, Qwen, Llama, Kimi, GLM, Nemotron, gpt-oss) so you don't have to operate GPUs yourself. It splits into three genuinely different products that get lumped together: per-token serverless APIs where you pay for tokens and share capacity with everyone else; dedicated endpoints where you rent reserved GPU capacity but still get a managed serving stack; and raw GPU/serverless-container clouds where you bring your own vLLM or SGLang.

Data checked 2026-09-06

As of September 2026 nearly all of them are OpenAI-chat-completions-compatible, so the switching cost is a base URL and an API key — which is exactly why price, throughput, and catalog freshness are the only real differentiators. The market moved hard in the last year: NVIDIA acqui-hired most of Groq's team and licensed its LPU technology for ~$20B in December 2025, Groq re-raised at roughly half its prior valuation and repositioned as a neocloud, Nebius AI Studio was renamed Nebius Token Factory, and Lambda is winding down its per-token Inference API to focus on GPU instances. Meanwhile Fireworks ($17.5B valuation, >$1B ARR) and Baseten (~$13B) have become the two large managed-inference incumbents.

A How to choose

Start by deciding which of the three products you actually need, because picking the wrong shape wastes more money than picking the wrong vendor. If you have bursty traffic and a model that appears in a public catalog, use a per-token serverless API and stop there — Together, Fireworks, DeepInfra, Novita, or Nebius Token Factory will all serve DeepSeek V4 Flash or gpt-oss-120b within about 2x of each other on price. If you have steady traffic above roughly 5-10M tokens/hour, model the crossover: a dedicated H100 at $2.20/hr on DeepInfra or $6.50/hr on Baseten beats per-token rates surprisingly early, and it also removes the noisy-neighbour latency variance that kills p99 on shared serverless.

If you need a model nobody hosts, a custom fine-tune, or a non-transformer workload, you want a container platform (Modal, Runpod, Baseten) or raw GPUs, and you are accepting responsibility for the serving stack. The second axis is speed vs. catalog, and this is where the popular default is often wrong. Cerebras and SambaNova are 3-15x faster than GPU-based providers on the models they host — 1,849 tok/s vs. ~150-300 tok/s on gpt-oss-120b — but Cerebras publicly serves about four models and SambaNova about seven, so if your product needs the newest Qwen or Kimi release the week it drops, those platforms are simply not an option and you should not design around them.

Groq is now in an awkward middle: still very fast, still cheap, but with a catalog that has narrowed to roughly eight models and an organisation that sold its core engineering team to NVIDIA — treat it as a bet, not a default. Conversely, Together and Fireworks add new open weights within days but are middling on throughput (Together clocks ~94 tok/s on gpt-oss-120b, slower than DeepInfra's turbo tier), so don't choose them for latency-critical voice or agent loops. Third, watch the cheap tail.

DeepInfra, Novita, Parasail, and CoreWeave will quote you $0.03-0.07 per million blended tokens where Together or Baseten quote $0.09-0.14, and for batch summarisation or offline enrichment that 2-4x is real money. The catch is that the cheap tier is usually the slow tier — DeepInfra's base gpt-oss-120b runs at 46 tok/s vs. 163 on turbo — and the support, status page, and SLA quality drop correspondingly.

Fourth, treat compliance and data residency as a hard filter applied first, not a tiebreaker: Nebius Token Factory sells EU-region deployment with a 99.9% SLA, Baseten and Fireworks sell self-hosted/VPC deployment, and most of the cheap tail does not. Finally, if you genuinely cannot predict which model you'll need, front everything with OpenRouter for a quarter, read the usage data, then negotiate directly with the two providers you actually used — routing layers are excellent for discovery and bad as a permanent dependency, because you inherit an extra hop of latency, an extra billing relationship, and no leverage on committed-spend discounts.

B At a glance

Name Output speed, gpt-oss-120bModels hostedDedicated endpointsLlama/DeepSeek/Qwen-class priceOpenAI-compatible Pricing
Together AI ~94 tok/s (Artificial Analysis)200+Yes — dedicated model + container inference, provisioned throughputDeepSeek V4 Flash $0.14/$0.28 per 1M; Qwen3.8 Flash $0.15/$0.47Yes DeepSeek V4 Flash $0.14/$0.28 per 1M; Llama 3.3 70B $1.04/$1.04; Kimi K3 $3.00/$15.00; dedicated H100 $5.49/GPU/hr; clusters from $3.19/GPU/hr reserved
Fireworks AI Not published by Artificial Analysis; 273 tok/s on DeepSeek V4 FlashHundreds (LLM, VLM, embedding, reranker); no official countYes — on-demand and reserved GPU deploymentsDeepSeek V4 Flash ~$0.11/1M blended; Nemotron 3.5 Lightning $0.04/1MYes DeepSeek V4 Flash ~$0.11/1M blended, GLM-5.3 ~$1.35/1M, Kimi K3 ~$4.22/1M (Artificial Analysis); on-demand H100/H200 $8.00/hr, B200 $13.00/hr, B300 $15.00/hr, GB300 $20.00/hr; LoRA SFT from $0.50/1M tokens
Baseten 267 tok/s (Artificial Analysis)~22 on the Model API; unlimited via custom deploymentYes — the primary product, incl. self-hosted/VPCDeepSeek V4 Flash $0.13/$0.26 per 1M; GPT OSS 120B $0.10/$0.50Yes (Model APIs); custom endpoints are whatever you build GPT OSS 120B $0.10/$0.50 per 1M; DeepSeek V4 Flash $0.13/$0.26; Kimi K3 $3.00/$15.00; dedicated H100 $0.10833/min (~$6.50/hr), B200 $0.16633/min (~$9.98/hr), A100 ~$4.00/hr
Groq (GroqCloud) 475 tok/s (Artificial Analysis); 500 tok/s claimed~8 production + previewEnterprise/dedicated LPU capacity by contract onlyLlama 3.3 70B ~$0.59/$0.79 per 1M; no DeepSeek/Kimi flagshipYes GPT OSS 120B $0.15/$0.60 per 1M; GPT OSS 20B $0.075/$0.30; Llama 3.1 8B ~$0.05/1M blended; batch API roughly 25% of on-demand rate
Cerebras Inference 1,849 tok/s (Artificial Analysis) — fastest measured~4 on the public rate cardYes — reserved capacity, custom pricing, required for Llama/Qwen/DeepSeekNot on public card — dedicated endpoints onlyYes gpt-oss-120b $0.35/$0.75 per 1M; Gemma 4 31B $0.99/$1.49; ZAI-GLM-4.7 $2.25/$2.75; Developer tier from $10; Enterprise by quote
SambaNova Cloud 709 tok/s (Artificial Analysis)~7 on the public cloud rate cardYes — enterprise and on-prem RDU appliances, pricing by quoteLlama 3.3 70B $0.60/$1.20; DeepSeek V3.2 $3.00/$4.50 per 1MYes gpt-oss-120b $0.22/$0.59 per 1M; Llama 3.3 70B $0.60/$1.20; DeepSeek-V3.1/V3.2 $3.00/$4.50; gemma-4-31B $0.38/$1.15; MiniMax-M3 $0.60/$2.40
DeepInfra 163 tok/s Turbo / 46 tok/s base (Artificial Analysis)~100+ across text, image, embedding, audioYes — dedicated GPUs from $0.89/hr (A100)DeepSeek V4 Flash $0.08/$0.18; Llama 3.3 70B $0.10/$0.32 per 1MYes DeepSeek V4 Flash $0.08/$0.18 per 1M; Llama 3.3 70B $0.10/$0.32; Llama 3.1 8B $0.02/$0.04; Kimi K3 $2.85/$14.25; dedicated H100 $2.20/hr, B200 $3.69/hr
Novita AI 150 tok/s (Artificial Analysis)200+Yes — serverless and dedicated endpoints, plus GPU instancesLlama 3.3 70B $0.135/1M in; DeepSeek V4 Flash $0.14/1M in; Qwen3.8 Flash $0.15/1M inYes (also a Hugging Face Inference Provider) GPT OSS 120B $0.05/1M in; Llama 3.3 70B $0.135/1M in; DeepSeek V4 Flash $0.14/1M in; Qwen3 Coder 30B $0.07/1M in; Kimi K3 $3.00/1M in
Nebius Token Factory 307-313 tok/s (Artificial Analysis)60+Yes — 99.9% SLA, custom autoscaling, EU/US regionsQwen3 32B $0.12/1M blended; Qwen3 30B A3B $0.12/1M; Kimi K3 $4.20/1MYes $0.08/1M blended (Nemotron 3.5 Lightning BF16) to $4.20/1M (Kimi K3); Qwen3 32B $0.12/1M; gpt-oss-120b Base $0.12/1M; volume discounts 10% at 1M tok/mo rising to 50% at 1B/mo; batch API a further 20-30% off
Parasail 126 tok/s (Artificial Analysis)~40 curated endpoints; 2M+ HF checkpoints on demandYes — reserved GPUs billed per minute, rates on requestgpt-oss-20b $0.04/1M in; range $0.01-$15.00/1M across modelsYes Serverless $0.01-$15.00 per 1M depending on model; gpt-oss-20b $0.04/1M in; gpt-oss-120b ~$0.07/1M blended; batch from $0.02-$0.04/1M for sub-4B models at FP4; dedicated GPU rates on request
Hyperbolic Not measured by Artificial Analysis~25 open modelsYes — reserved GPU clusters, 3-12 month commits, up to 40% offLlama 3.3 70B ~$0.40/1M; DeepSeek R1 ~$3.00/1M (catalog appears dated)Yes H100 SXM $1.50-$3.19/hr, H200 $2.40-$3.99/hr, B200 $3.50-$5.99/hr, RTX 4090 $0.30/hr (rates moved sharply during 2026); serverless Llama 3.3 70B ~$0.40/1M, DeepSeek R1 ~$3.00/1M; reserved clusters up to 40% off
Replicate Not measured by Artificial AnalysisThousands of community models (image, video, audio, LLM)Yes — private/dedicated deployments, multi-GPU on committed spendMostly per GPU-second; DeepSeek-R1 $3.75/1M input, $0.01/1k outputPartial — native Replicate API is primary; OpenAI-compatible endpoint for some LLMs H100 $0.001525/s ($5.49/hr), A100 80GB $0.001400/s ($5.04/hr), L40S $0.000975/s ($3.51/hr), T4 $0.000225/s ($0.81/hr), 8x H100 $43.92/hr; DeepSeek-R1 $3.75/1M input
Modal N/A — depends entirely on the stack you deployNone — bring your own containerYes by construction — every deployment is your own containerNo per-token pricing; H100 $0.001097/s (~$3.95/hr)Only if you deploy an OpenAI-compatible server (vLLM/SGLang) H100 SXM5 $0.001097/s (~$3.95/hr), H200 $0.001261/s (~$4.54/hr), B200 $0.001736/s (~$6.25/hr), A100 80GB ~$2.50/hr, L40S ~$1.95/hr; Starter $0/mo + $30 credits, Team $250/mo + $100 credits
Runpod N/A — depends on the stack you deployA small set of public endpoints; primarily bring-your-own-containerYes — Pods (community and secure) plus reserved capacityNo competitive per-token card; H100 pod $2.69-$3.49/hrOnly via the vLLM template or an OpenAI-compatible worker you deploy Serverless H100 $4.79/hr, A100 $2.72/hr, RTX 4090 $1.10/hr; Pods H100 SXM $2.69 (community) / $3.49 (secure), A100 PCIe $1.19/$1.59, RTX 4090 $0.34/$0.74, A5000 $0.16/$0.27
fal N/A — no LLM servingHundreds of image/video/audio modelsYes — private serverless deployments and dedicated GPU computeN/A — media only; $0.02-$0.04/image, $0.05-$0.4/video-secondNo — native fal queue/REST API Video $0.05/s (Wan 2.5), $0.07/s (Kling 2.5 Turbo Pro), $0.4/s (Veo 3), $0.2/video (Ovi); images $0.02-$0.04 each; GPU compute from $1.89/hr H100 to $4.49/hr B300
CoreWeave Serverless Inference 74 tok/s (Artificial Analysis)Small curated set; count not publishedYes — this is the core business (reserved GPU clusters)Not published — contact sales; gpt-oss-120b ~$0.03/1M blendedYes gpt-oss-120b ~$0.03/1M blended (Artificial Analysis); GPU nodes HGX H100 $49.24/hr, HGX B200 $68.80/hr, GB200 NVL72 $42.00/hr, A100 $21.60/hr, L40S $18.00/hr; most per-token rates are contact-sales
Lambda N/A — inference API winding downEffectively none going forwardGPU instances and 1-Click Clusters onlyNo ongoing per-token card; H100 $3.99/hrWas yes; API deprecated H100 SXM $3.99-$4.29/hr, B200 SXM6 $6.69-$6.99/hr, GH200 $2.29/hr, A100 $1.99-$2.79/hr; 1-Click Clusters 16x H100 $6.16/GPU/hr, 256x H100 $5.54/GPU/hr, 16x B200 $9.86/GPU/hr; Inference API winding down
OpenRouter Varies by routed provider (46-1,849 tok/s across the pool)500+ models across 80+ providers (routed, not hosted)No — routes to providers' endpoints; supports provider pinning and BYOKProvider pass-through plus undisclosed platform feeYes — this is the entire point Pass-through of each provider's per-token rate plus an undisclosed platform fee (not published on the marketing site); credits purchased in advance, e.g. $10 or $99

C Entries

Together AI

Together runs a large serverless catalog (200+ text, vision, image, video and audio models) alongside dedicated endpoints and rentable GPU clusters, so you can start per-token and move the same model onto reserved hardware without changing the API. Pricing as of September 2026: DeepSeek V4 Flash 0731 $0.14/$0.28 per 1M, DeepSeek V4 Pro $1.32/$3.96, Qwen3.8 Flash $0.15/$0.47, Llama 3.3 70B $1.04/$1.04, Kimi K3 $3.00/$15.00, GLM-5.3-Flash $0.15/$0.50. Dedicated inference is $5.49/GPU/hr for H100 (promo $3.99) and $8.99 for B200; GPU clusters run $3.99/hr H100 on-demand down to $3.19/hr reserved at 181+ days. The differentiator against Fireworks is the cluster business — Together will sell you bare training capacity, not just inference.

Output speed, gpt-oss-120b~94 tok/s (Artificial Analysis)
Models hosted200+
Dedicated endpointsYes — dedicated model + container inference, provisioned throughput
Llama/DeepSeek/Qwen-class priceDeepSeek V4 Flash $0.14/$0.28 per 1M; Qwen3.8 Flash $0.15/$0.47
OpenAI-compatibleYes
Free tierSignup credits; one $0.00 model

Watch out: Throughput is unremarkable — Artificial Analysis clocks Together at ~94 tok/s on gpt-oss-120b, roughly 20x slower than Cerebras and slower than DeepInfra's turbo tier — so it is the wrong choice for latency-sensitive voice or tight agent loops. Serverless p99 latency varies with shared-capacity load, and the headline dedicated H100 rate ($5.49/hr) is well above DeepInfra ($2.20/hr) for the same silicon.

DeepSeek V4 Flash $0.14/$0.28 per 1M; Llama 3.3 70B $1.04/$1.04; Kimi K3 $3.00/$15.00; dedicated H100 $5.49/GPU/hr; clusters from $3.19/GPU/hr reserved

Fireworks AI

Fireworks pairs a fast-moving serverless catalog (hundreds of LLMs, VLMs, embedding and reranker models, with GLM 5.3 Flash, DeepSeek-V4-Pro at 1M context, Qwen3.8 Flash Next and Kimi K3 all live) with on-demand GPU deployments and a managed fine-tuning product. Serverless is tiered — Standard, Priority and Fast — so you buy latency rather than getting one shared queue; Artificial Analysis measures 273 tok/s on DeepSeek V4 Flash and 521 tok/s on Nemotron 3.5 Lightning at $0.04/1M. On-demand GPUs rose on 1 September 2026 to $8.00/hr for H100 and H200, $13.00 for B200, $15.00 for B300 and $20.00 for GB300. Managed training is priced per token — LoRA SFT from $0.50/1M on models up to 16B, $3.00/1M on 16-80B. It has become the largest pure-play inference vendor by revenue (>$1B ARR, $17.5B valuation as of July 2026).

Output speed, gpt-oss-120bNot published by Artificial Analysis; 273 tok/s on DeepSeek V4 Flash
Models hostedHundreds (LLM, VLM, embedding, reranker); no official count
Dedicated endpointsYes — on-demand and reserved GPU deployments
Llama/DeepSeek/Qwen-class priceDeepSeek V4 Flash ~$0.11/1M blended; Nemotron 3.5 Lightning $0.04/1M
OpenAI-compatibleYes
Free tier$1 credit

Watch out: The GPU price list jumped sharply on 1 September 2026 (H100 $7 to $8, B200 $10 to $13, B300 $12 to $15), so on-demand GPU economics are worse than DeepInfra, Lambda or Runpod for the same hardware and you should assume further moves. Per-model serverless rates are not published on the marketing pricing page — you have to dig into docs or the model cards — which makes cost modelling harder than at Together or DeepInfra. The three-tier serverless model also means the advertised low price is the slow tier.

DeepSeek V4 Flash ~$0.11/1M blended, GLM-5.3 ~$1.35/1M, Kimi K3 ~$4.22/1M (Artificial Analysis); on-demand H100/H200 $8.00/hr, B200 $13.00/hr, B300 $15.00/hr, GB300 $20.00/hr; LoRA SFT from $0.50/1M tokens

Baseten

Baseten's core product is Truss-based deployment of arbitrary models onto dedicated autoscaling GPUs, with a per-token Model API layered on top for the popular open weights so you can prototype without provisioning anything. Model API rates as of September 2026: GPT OSS 120B $0.10/$0.50 per 1M, DeepSeek V4 Flash $0.13/$0.26, GLM-5.3-Flash $0.15/$0.50, GLM-5.3 $1.40/$4.40, DeepSeek V4 Pro $1.74/$3.48, Kimi K3 $3.00/$15.00. Dedicated compute is billed per minute: H100 $0.10833/min ($6.50/hr), B200 $0.16633/min ($9.98/hr), A100 $0.06667/min ($4.00/hr), L4 $0.01414/min. Unlike Together or Fireworks it is genuinely a platform for your model, not just a catalog, and it sells self-hosted/VPC deployment for regulated buyers. It raised a $1.5B Series F in 2026 at roughly $13B on ~$600M ARR.

Output speed, gpt-oss-120b267 tok/s (Artificial Analysis)
Models hosted~22 on the Model API; unlimited via custom deployment
Dedicated endpointsYes — the primary product, incl. self-hosted/VPC
Llama/DeepSeek/Qwen-class priceDeepSeek V4 Flash $0.13/$0.26 per 1M; GPT OSS 120B $0.10/$0.50
OpenAI-compatibleYes (Model APIs); custom endpoints are whatever you build
Free tierSignup credits

Watch out: Dedicated GPU pricing is at the top of the market — $6.50/hr for an H100 is roughly 3x DeepInfra's $2.20/hr and above Runpod's $2.69-4.79/hr — so you are paying a large premium for the control plane. The per-token Model API catalog is much narrower than Together's or Novita's (about 22 models tracked), and throughput is mid-pack at ~267 tok/s on gpt-oss-120b. Wrong choice if you just want the cheapest tokens for a standard open model.

GPT OSS 120B $0.10/$0.50 per 1M; DeepSeek V4 Flash $0.13/$0.26; Kimi K3 $3.00/$15.00; dedicated H100 $0.10833/min (~$6.50/hr), B200 $0.16633/min (~$9.98/hr), A100 ~$4.00/hr

Groq (GroqCloud)

GroqCloud serves a deliberately small set of open models on custom LPU hardware at very low latency: GPT OSS 120B at 500 tok/s and $0.15/$0.60 per 1M, GPT OSS 20B at 1,000 tok/s and $0.075/$0.30, Llama 3.1 8B at 560 tok/s, Llama 3.3 70B at 280 tok/s, plus preview Qwen3.6/3.8-27B and the Groq Compound agentic system with built-in web search and code execution at ~450 tok/s. The corporate situation matters as much as the tech: in December 2025 NVIDIA paid roughly $20B to hire most of Groq's engineering team and license the LPU dataflow technology (now shipping as NVIDIA's Groq 3 LPX), and the remaining company raised $650M and then $350M at a $3.5B valuation — about half its pre-deal worth — to operate as a neocloud running both LPUs and NVIDIA GPUs.

Output speed, gpt-oss-120b475 tok/s (Artificial Analysis); 500 tok/s claimed
Models hosted~8 production + preview
Dedicated endpointsEnterprise/dedicated LPU capacity by contract only
Llama/DeepSeek/Qwen-class priceLlama 3.3 70B ~$0.59/$0.79 per 1M; no DeepSeek/Kimi flagship
OpenAI-compatibleYes
Free tierYes, rate-limited

Watch out: Serious platform risk: NVIDIA hired away most of the engineering team and licensed the core IP, and the company re-raised at roughly half its former valuation. The catalog has narrowed to about eight models with no current DeepSeek, Kimi or GLM flagship, so if your product depends on the newest open weights this is not a viable primary. Long-context and very large MoE models are not served, official per-model pricing is hard to find on the public site, and Artificial Analysis measures 475 tok/s on gpt-oss-120b — fast, but now well behind Cerebras.

GPT OSS 120B $0.15/$0.60 per 1M; GPT OSS 20B $0.075/$0.30; Llama 3.1 8B ~$0.05/1M blended; batch API roughly 25% of on-demand rate

Cerebras Inference

Cerebras runs models on wafer-scale CS-series systems and posts throughput no GPU or LPU provider matches: 1,849 tok/s on gpt-oss-120b (high reasoning) and 1,779 tok/s on Gemma 4 31B, versus 150-300 tok/s from mainstream GPU providers. The public pay-per-token rate card as of mid-2026 lists gpt-oss-120b at $0.35/$0.75 per 1M, Gemma 4 31B at $0.99/$1.49 and ZAI-GLM-4.7 at $2.25/$2.75, though only gpt-oss-120b is genuinely production-classified — the others are preview or slated for deprecation. Llama, Qwen3, Mistral and DeepSeek families are available only through dedicated endpoints on custom reserved-capacity pricing, not the public card. Entry is $5 in trial credits, then a self-serve Developer tier from $10.

Output speed, gpt-oss-120b1,849 tok/s (Artificial Analysis) — fastest measured
Models hosted~4 on the public rate card
Dedicated endpointsYes — reserved capacity, custom pricing, required for Llama/Qwen/DeepSeek
Llama/DeepSeek/Qwen-class priceNot on public card — dedicated endpoints only
OpenAI-compatibleYes
Free tier$5 credits

Watch out: The catalog is effectively four models, only one of which is production-grade on the public rate card; everything else (Llama, Qwen3, DeepSeek, Mistral) requires a reserved-capacity contract. That makes it unusable as a general-purpose backend and dangerous to build a product around, because a model you depend on can move to preview or deprecation. It is also not the cheapest — $0.35/$0.75 on gpt-oss-120b is several times DeepInfra's or CoreWeave's blended rate — so you are explicitly buying speed, and capacity has historically been the constraint.

gpt-oss-120b $0.35/$0.75 per 1M; Gemma 4 31B $0.99/$1.49; ZAI-GLM-4.7 $2.25/$2.75; Developer tier from $10; Enterprise by quote

SambaNova Cloud

SambaNova serves open models on its own RDU (reconfigurable dataflow unit) silicon, landing second to Cerebras on throughput at 709 tok/s on gpt-oss-120b while charging less than Cerebras for it. The published cloud rate card as of September 2026 covers gpt-oss-120b $0.22/$0.59, Meta-Llama-3.3-70B $0.60/$1.20, gemma-4-31B-it $0.38/$1.15, MiniMax-M2.7 $0.60/$2.40 with $0.06 cached input, MiniMax-M3 $0.60/$2.40, and DeepSeek-V3.1 and V3.2 both at $3.00/$4.50. The company's 2026 pitch is air-cooled inference hardware that drops into existing data centres, so the on-prem/appliance path is a real product rather than a slide.

Output speed, gpt-oss-120b709 tok/s (Artificial Analysis)
Models hosted~7 on the public cloud rate card
Dedicated endpointsYes — enterprise and on-prem RDU appliances, pricing by quote
Llama/DeepSeek/Qwen-class priceLlama 3.3 70B $0.60/$1.20; DeepSeek V3.2 $3.00/$4.50 per 1M
OpenAI-compatibleYes
Free tierYes, rate-limited developer tier

Watch out: Small catalog (about seven models on the public card) and it is visibly behind on model freshness — DeepSeek is still at V3.1/V3.2 while Together, DeepInfra and Fireworks all serve V4, and those V3.x endpoints are priced at $3.00/$4.50 per 1M, far above what V4 Flash costs elsewhere. No published dedicated-endpoint rate card or throughput SLA on the pricing page. Wrong choice if you need the current frontier open weights or a broad model menu.

gpt-oss-120b $0.22/$0.59 per 1M; Llama 3.3 70B $0.60/$1.20; DeepSeek-V3.1/V3.2 $3.00/$4.50; gemma-4-31B $0.38/$1.15; MiniMax-M3 $0.60/$2.40

DeepInfra

DeepInfra competes almost entirely on price and does it credibly: DeepSeek-V4-Flash-0731 at $0.08/$0.18 per 1M, Llama-3.3-70B-Turbo at $0.10/$0.32, Llama-3.1-8B-Turbo at $0.02/$0.04, Qwen3.5-9B at $0.10/$0.15, Kimi-K2.6 at $0.75/$3.50, Kimi-K3 at $2.85/$14.25, Mistral-Nemo at $0.019/$0.03. Dedicated GPUs are the cheapest published rates among managed inference vendors — A100 $0.89/hr, H100 $2.20/hr, H200 $2.69/hr, B200 $3.69/hr, B300 $4.89/hr — roughly a third of Baseten's or Fireworks' list. It offers a base tier and a faster Turbo tier on the same models, so you choose the latency/price point per request.

Output speed, gpt-oss-120b163 tok/s Turbo / 46 tok/s base (Artificial Analysis)
Models hosted~100+ across text, image, embedding, audio
Dedicated endpointsYes — dedicated GPUs from $0.89/hr (A100)
Llama/DeepSeek/Qwen-class priceDeepSeek V4 Flash $0.08/$0.18; Llama 3.3 70B $0.10/$0.32 per 1M
OpenAI-compatibleYes
Free tierSmall signup credit

Watch out: The cheap tier is genuinely slow — Artificial Analysis measures 46 tok/s on base gpt-oss-120b versus 163 on Turbo — so the headline price and the usable price are different numbers. Support, status transparency and enterprise paperwork (SOC 2 scope, DPAs, regional residency) are thinner than at Fireworks, Baseten or Nebius, and there is no meaningful SLA on the shared tier. Wrong choice for a regulated buyer or a latency-critical interactive product.

DeepSeek V4 Flash $0.08/$0.18 per 1M; Llama 3.3 70B $0.10/$0.32; Llama 3.1 8B $0.02/$0.04; Kimi K3 $2.85/$14.25; dedicated H100 $2.20/hr, B200 $3.69/hr

Novita AI

Novita runs a broad, aggressively priced serverless catalog (200+ models covering LLMs, image, video and audio) alongside GPU instances, serverless endpoints and dedicated endpoints, and is one of the Hugging Face Inference Providers so you can call it straight from the Hub. Input-side rates as of September 2026 include GPT OSS 120B $0.05/1M, GPT OSS 20B $0.04/1M, Llama 3.1 8B $0.02/1M, Llama 3.3 70B $0.135/1M, Qwen3 Coder 30B $0.07/1M, Qwen3.8 Flash $0.15/1M, Qwen3.5 27B $0.30/1M, GLM 5.3 Flash $0.075/1M, DeepSeek V4 Flash $0.14/1M, DeepSeek V4 Pro $1.60/1M, Kimi K3 $3.00/1M. It sits in the same price band as DeepInfra with a wider multimodal catalog.

Output speed, gpt-oss-120b150 tok/s (Artificial Analysis)
Models hosted200+
Dedicated endpointsYes — serverless and dedicated endpoints, plus GPU instances
Llama/DeepSeek/Qwen-class priceLlama 3.3 70B $0.135/1M in; DeepSeek V4 Flash $0.14/1M in; Qwen3.8 Flash $0.15/1M in
OpenAI-compatibleYes (also a Hugging Face Inference Provider)
Free tierTrial credits

Watch out: Throughput is low — 150 tok/s on gpt-oss-120b, roughly a third of Groq and a twelfth of Cerebras — so this is a batch and background-job platform, not an interactive one. Published pricing is largely input-side on the marketing page, so you must check output rates per model before modelling costs. Enterprise assurances, dedicated-endpoint rate cards and regional residency options are less developed than at Nebius or Baseten.

GPT OSS 120B $0.05/1M in; Llama 3.3 70B $0.135/1M in; DeepSeek V4 Flash $0.14/1M in; Qwen3 Coder 30B $0.07/1M in; Kimi K3 $3.00/1M in

Nebius Token Factory

Nebius rebranded AI Studio to Token Factory during 2026 (docs moved to docs.tokenfactory.nebius.com) and repositioned it as the enterprise inference tier of its GPU cloud. It serves 60+ open models — DeepSeek, Qwen, Llama, Mistral, GPT OSS, GLM and NVIDIA Nemotron — on an OpenAI-compatible API with shared and dedicated tiers, optional EU or US regional deployment, custom autoscaling and a 99.9% uptime SLA, and you can serve your own fine-tunes through the same dashboard. Blended prices range from $0.08/1M for Nemotron 3.5 Lightning (BF16) to $4.20/1M for Kimi K3, a 54x spread; Qwen3 32B Base is $0.12/1M blended, gpt-oss-120b Base runs 313 tok/s at $0.12/1M. Volume discounts start at 10% at 1M tokens/month and reach 50% at 1B/month, with a further 20-30% off via the batch API.

Output speed, gpt-oss-120b307-313 tok/s (Artificial Analysis)
Models hosted60+
Dedicated endpointsYes — 99.9% SLA, custom autoscaling, EU/US regions
Llama/DeepSeek/Qwen-class priceQwen3 32B $0.12/1M blended; Qwen3 30B A3B $0.12/1M; Kimi K3 $4.20/1M
OpenAI-compatibleYes
Free tierSignup credits; limited free models

Watch out: The AI Studio to Token Factory rename means older docs, SDK references and blog links are stale — budget time for that. The catalog skews heavily toward NVIDIA Nemotron variants (the cheapest and fastest options are all Nemotron), so the models that make Nebius look good on price/speed charts may not be the ones you want. Throughput is mid-pack at 307 tok/s on gpt-oss-120b, and the flagship-model prices at the top of the range ($4.20/1M for Kimi K3) are above Together and DeepInfra.

$0.08/1M blended (Nemotron 3.5 Lightning BF16) to $4.20/1M (Kimi K3); Qwen3 32B $0.12/1M; gpt-oss-120b Base $0.12/1M; volume discounts 10% at 1M tok/mo rising to 50% at 1B/mo; batch API a further 20-30% off

Parasail

Parasail orchestrates rented GPU capacity across 40+ data centres in 15+ countries and exposes it as serverless, dedicated or batch inference. Its distinguishing move is breadth by construction rather than curation: it advertises per-token access to essentially any open-weight model on Hugging Face (2M+), with about 40 curated endpoints (Qwen, Llama, Mistral, DeepSeek, gpt-oss) already warm. Serverless spans $0.01/1M for BGE-M3 input to $15.00/1M at the top end, with gpt-oss-20b at $0.04/1M input; batch is priced in bands by parameter count and precision, starting at $0.02-$0.04 per 1M for 0-4.1B models at FP4. Artificial Analysis measures 126 tok/s and $0.07/1M blended on gpt-oss-120b. Dedicated GPU reservations run B300 down to RTX 5090, billed per minute, rates on request. It raised a $32M Series A in April 2026.

Output speed, gpt-oss-120b126 tok/s (Artificial Analysis)
Models hosted~40 curated endpoints; 2M+ HF checkpoints on demand
Dedicated endpointsYes — reserved GPUs billed per minute, rates on request
Llama/DeepSeek/Qwen-class pricegpt-oss-20b $0.04/1M in; range $0.01-$15.00/1M across models
OpenAI-compatibleYes
Free tierTrial credits

Watch out: Dedicated GPU pricing is quote-only, which makes it impossible to compare against DeepInfra or Runpod without a sales call. Throughput is slow at 126 tok/s on gpt-oss-120b, and because capacity is rented across many third-party data centres, latency, region and reliability characteristics vary more than at a vendor operating its own fleet. It is a much smaller company than Together or Fireworks, so weigh vendor risk before making it a single point of failure.

Serverless $0.01-$15.00 per 1M depending on model; gpt-oss-20b $0.04/1M in; gpt-oss-120b ~$0.07/1M blended; batch from $0.02-$0.04/1M for sub-4B models at FP4; dedicated GPU rates on request

Hyperbolic

Hyperbolic runs an open-access GPU marketplace aggregating supply from many operators, with a serverless inference API bolted on for about 25 open models — OpenAI-compatible chat, vision, image generation (FLUX, Stable Diffusion, LoRA) and audio (Melo TTS). Reported on-demand GPU rates in 2026 have moved substantially: H100 SXM quoted at $1.50/hr in June and $3.19/hr by August, H200 $2.40 to $3.99, B200 $3.50 to $5.99, with consumer cards from $0.16/hr (RTX 3070) and $0.30/hr (RTX 4090). Reserved clusters carry 3-12 month commitments with up to 40% discount. Serverless inference has been quoted at Llama 3.3 70B $0.40/1M, DeepSeek R1 $3.00/1M and Llama 3.1 405B $4.00/1M, with images from $0.01.

Output speed, gpt-oss-120bNot measured by Artificial Analysis
Models hosted~25 open models
Dedicated endpointsYes — reserved GPU clusters, 3-12 month commits, up to 40% off
Llama/DeepSeek/Qwen-class priceLlama 3.3 70B ~$0.40/1M; DeepSeek R1 ~$3.00/1M (catalog appears dated)
OpenAI-compatibleYes
Free tierSignup credits

Watch out: Confidence in these figures is low: Hyperbolic's own pricing page did not resolve during research and the numbers come from secondary aggregators, and the quoted GPU rates roughly doubled between June and August 2026 — assume anything here is stale. The serverless model list looks a generation behind (Llama 3.1 405B, DeepSeek R1 rather than V4), and Artificial Analysis lists a $0.21/1M blended price for gpt-oss-120b with no throughput measurement at all, which suggests thin capacity. Wrong choice for a production API where you need a stable catalog and predictable latency.

H100 SXM $1.50-$3.19/hr, H200 $2.40-$3.99/hr, B200 $3.50-$5.99/hr, RTX 4090 $0.30/hr (rates moved sharply during 2026); serverless Llama 3.3 70B ~$0.40/1M, DeepSeek R1 ~$3.00/1M; reserved clusters up to 40% off

Replicate

Replicate's model is that every model is a versioned, containerised artifact (a Cog image) with a typed schema, so you can run community models, push your own, and fork someone else's without touching infrastructure. Billing is predominantly per GPU-second rather than per token: CPU $0.000100/s, T4 $0.000225/s ($0.81/hr), L40S $0.000975/s ($3.51/hr), A100 80GB $0.001400/s ($5.04/hr), H100 $0.001525/s ($5.49/hr), scaling to 8x H100 at $0.012200/s ($43.92/hr, committed-spend contracts only). Some language models are billed per token instead — DeepSeek-R1 is listed at $3.75/1M input and $0.01 per thousand output tokens. Private models normally bill setup, idle and active time; fast-booting fine-tunes bill active time only.

Output speed, gpt-oss-120bNot measured by Artificial Analysis
Models hostedThousands of community models (image, video, audio, LLM)
Dedicated endpointsYes — private/dedicated deployments, multi-GPU on committed spend
Llama/DeepSeek/Qwen-class priceMostly per GPU-second; DeepSeek-R1 $3.75/1M input, $0.01/1k output
OpenAI-compatiblePartial — native Replicate API is primary; OpenAI-compatible endpoint for some LLMs
Free tierSmall trial allowance

Watch out: Per-GPU-second billing is a poor fit for LLM chat traffic — you pay for cold starts and idle, and a sustained H100 at $5.49/hr is more than double DeepInfra's $2.20/hr for the same card. The LLM catalog and per-token coverage lag the dedicated text providers badly (DeepSeek-R1 rather than V4), throughput is not benchmarked by Artificial Analysis, and multi-GPU tiers require committed-spend contracts. Wrong choice as a primary LLM backend; right choice for diffusion and audio.

H100 $0.001525/s ($5.49/hr), A100 80GB $0.001400/s ($5.04/hr), L40S $0.000975/s ($3.51/hr), T4 $0.000225/s ($0.81/hr), 8x H100 $43.92/hr; DeepSeek-R1 $3.75/1M input

Runpod

Runpod sells two things — Serverless (autoscaling GPU workers with a queue, billed per second of active work) and Pods (rented VMs, split into cheaper Community Cloud and more reliable Secure Cloud). September 2026 serverless rates: B300 $9.98/hr, B200 $8.64/hr, H200 $5.93/hr, H100 $4.79/hr, A100 80GB $2.72/hr, L40S/6000 Ada $1.75/hr, RTX 4090 $1.10/hr, A5000/L4/3090 $0.69/hr. Pods are much cheaper: H100 SXM $2.69 community / $3.49 secure, H200 $3.59/$4.59, A100 PCIe $1.19/$1.59, L40S $0.79/$1.09, RTX 4090 $0.34/$0.74, A5000 $0.16/$0.27. There are also public per-token endpoints and per-image/per-request endpoints, but the catalog is small and secondary to the compute business.

Output speed, gpt-oss-120bN/A — depends on the stack you deploy
Models hostedA small set of public endpoints; primarily bring-your-own-container
Dedicated endpointsYes — Pods (community and secure) plus reserved capacity
Llama/DeepSeek/Qwen-class priceNo competitive per-token card; H100 pod $2.69-$3.49/hr
OpenAI-compatibleOnly via the vLLM template or an OpenAI-compatible worker you deploy
Free tierNone ongoing

Watch out: Community Cloud is aggregated third-party hardware with no meaningful availability guarantee — nodes can disappear and network/disk performance varies, so it is not a production substrate. Serverless cold starts on large models are painful because weights must load into VRAM, and you are still responsible for the whole serving stack. The public per-token endpoint catalog is thin and not a credible alternative to Together or DeepInfra.

Serverless H100 $4.79/hr, A100 $2.72/hr, RTX 4090 $1.10/hr; Pods H100 SXM $2.69 (community) / $3.49 (secure), A100 PCIe $1.19/$1.59, RTX 4090 $0.34/$0.74, A5000 $0.16/$0.27

fal

fal is the generative-media counterpart to the text-focused platforms: it hosts open and licensed image and video models on an inference runtime tuned for diffusion and transformer-video workloads, and bills by output unit rather than tokens or GPU time. Representative September 2026 rates: Wan 2.5 at $0.05/second of video, Kling 2.5 Turbo Pro at $0.07/second, Veo 3 at $0.4/second, Ovi at $0.2/video; Seedream V4 at $0.03/image, Flux Kontext Pro at $0.04/image, Qwen image at $0.02/megapixel. Serverless GPU compute is also available from about $1.89/hr for H100 up to $4.49/hr for B300 if you need to run your own workflow. Output-unit billing means you can price a feature per user action without modelling GPU seconds.

Output speed, gpt-oss-120bN/A — no LLM serving
Models hostedHundreds of image/video/audio models
Dedicated endpointsYes — private serverless deployments and dedicated GPU compute
Llama/DeepSeek/Qwen-class priceN/A — media only; $0.02-$0.04/image, $0.05-$0.4/video-second
OpenAI-compatibleNo — native fal queue/REST API
Free tierTrial credits

Watch out: It does not serve LLMs, so it cannot be your text-inference platform — it is strictly complementary to Together, Fireworks or DeepInfra. Per-output pricing is easy to reason about but expensive at volume compared with renting a GPU and running the same open weights yourself, and some of the headline models (Veo 3, Kling) are proprietary rather than open weights, so a portion of the catalog is not portable elsewhere. No tokens/sec metric applies.

Video $0.05/s (Wan 2.5), $0.07/s (Kling 2.5 Turbo Pro), $0.4/s (Veo 3), $0.2/video (Ovi); images $0.02-$0.04 each; GPU compute from $1.89/hr H100 to $4.49/hr B300

CoreWeave Serverless Inference

CoreWeave added a pay-per-token serverless inference product on top of its large GPU fleet, with tracing and evals built in, aimed at customers already buying reserved capacity who want an elastic overflow tier. Artificial Analysis measures it at 74 tok/s and $0.03/1M blended on gpt-oss-120b — the cheapest blended rate in that comparison. The underlying GPU business publishes instance-level hourly rates (GB200 NVL72 $42.00/hr, HGX B200 $68.80/hr, HGX H100 $49.24/hr for the full node, A100 $21.60/hr, L40S $18.00/hr in North America), but the per-token inference columns are largely 'contact sales'. Practically, this is an add-on for existing CoreWeave accounts rather than a self-serve developer API.

Output speed, gpt-oss-120b74 tok/s (Artificial Analysis)
Models hostedSmall curated set; count not published
Dedicated endpointsYes — this is the core business (reserved GPU clusters)
Llama/DeepSeek/Qwen-class priceNot published — contact sales; gpt-oss-120b ~$0.03/1M blended
OpenAI-compatibleYes
Free tierNone

Watch out: Almost nothing is self-serve: the published pricing page routes per-token inference to 'contact sales', so you cannot benchmark cost without a commercial conversation, and confidence in the $0.03/1M figure is low because it comes from a third-party benchmark rather than a rate card. Throughput is the weakest in the comparison at 74 tok/s on gpt-oss-120b. The catalog is small relative to Together or Novita. Wrong choice for a startup that wants to swipe a card and start calling an endpoint today.

gpt-oss-120b ~$0.03/1M blended (Artificial Analysis); GPU nodes HGX H100 $49.24/hr, HGX B200 $68.80/hr, GB200 NVL72 $42.00/hr, A100 $21.60/hr, L40S $18.00/hr; most per-token rates are contact-sales

Lambda

Lambda is a GPU cloud with among the cleanest published on-demand rates: H100 SXM $3.99-$4.29/hr, B200 SXM6 $6.69-$6.99/hr, GH200 $2.29/hr, A100 SXM/PCIe $1.99-$2.79/hr, plus 1-Click Clusters at $6.16/GPU/hr for 16x H100 down to $5.54 at 256 GPUs, and $9.86 down to $8.87/GPU/hr for HGX B200 configurations on two-week-to-one-year terms. The important 2026 change is that its OpenAI-compatible per-token Inference API is being wound down — Lambda's own inference page now directs customers to deploy on NVIDIA GPU instances instead. Treat Lambda as a compute vendor, not an inference API vendor.

Output speed, gpt-oss-120bN/A — inference API winding down
Models hostedEffectively none going forward
Dedicated endpointsGPU instances and 1-Click Clusters only
Llama/DeepSeek/Qwen-class priceNo ongoing per-token card; H100 $3.99/hr
OpenAI-compatibleWas yes; API deprecated
Free tierNone

Watch out: Do not build on the Lambda Inference API: it is explicitly winding down and Lambda has not published a shutdown date, model list or ongoing rate card, so any per-token integration you write here is throwaway work. As a pure GPU cloud there is no managed serving stack, no model catalog and no OpenAI-compatible endpoint unless you deploy one yourself. Reserved terms beyond one year require a sales conversation, and 1-Click Clusters carry a two-week minimum.

H100 SXM $3.99-$4.29/hr, B200 SXM6 $6.69-$6.99/hr, GH200 $2.29/hr, A100 $1.99-$2.79/hr; 1-Click Clusters 16x H100 $6.16/GPU/hr, 256x H100 $5.54/GPU/hr, 16x B200 $9.86/GPU/hr; Inference API winding down · deprecated

OpenRouter

OpenRouter is a routing and billing layer rather than an inference platform: one API key and one OpenAI-compatible endpoint reach 500+ models across 80+ providers, including most of the vendors in this chapter, with automatic failover when a provider degrades and per-provider routing preferences (cheapest, fastest, or a named list). Its practical value is discovery and de-risking — you can A/B the same open model across Together, DeepInfra, Novita and Parasail without four integrations, and you get normalised usage and cost analytics across all of them. Pricing is provider pass-through plus a platform fee; the marketing site does not publish the markup percentage, so confirm the current fee schedule and any BYOK terms in the docs before committing.

Output speed, gpt-oss-120bVaries by routed provider (46-1,849 tok/s across the pool)
Models hosted500+ models across 80+ providers (routed, not hosted)
Dedicated endpointsNo — routes to providers' endpoints; supports provider pinning and BYOK
Llama/DeepSeek/Qwen-class priceProvider pass-through plus undisclosed platform fee
OpenAI-compatibleYes — this is the entire point
Free tierRotating free models with strict rate limits

Watch out: You add a network hop and a second billing relationship to every request, which costs latency and makes p99 debugging harder — you are now diagnosing someone else's routing decision. The markup is not publicly documented, and routing to whichever provider is cheapest means quantisation, context window and throughput can silently change under you between requests unless you pin providers. It also gives you no leverage to negotiate committed-spend discounts, so it is the wrong permanent home for a workload you already understand.

Pass-through of each provider's per-token rate plus an undisclosed platform fee (not published on the marketing site); credits purchased in advance, e.g. $10 or $99