Mid-tier Mainstream models Proxy required (the site makes no China-related statements)

Baseten Review: Pricing & Comparison

An inference cloud for open-weight models with a published $-per-million-token sheet and per-minute dedicated GPUs

Last verified: 2026-10-15 · Visit official site →

An inference cloud that lays its price sheet on the table

Baseten (baseten.co) is best known as a platform for deploying and hosting open models, and it also offers token-priced Model APIs — the path for people who would rather not run their own deployment. The category matters: it hosts open-weight models only, with no OpenAI, Claude or Gemini, so it is not a relay replacement.

Pricing is the most interesting part, because nearly all of it is published:

Model APIs (per 1M tokens, input/cache/output, USD):
- GPT OSS 120B: $0.10 / — / $0.50
- GLM-5.3-Flash: $0.15 / $0.03 / $0.50
- DeepSeek V4.1 Flash: $0.30 / $0.007 / $1.20
- NVIDIA Nemotron 3 Ultra: $0.60 / $0.12 / $2.40
- DeepSeek V4 Pro 0813: $1.32 / $0.132 / $3.96
- GLM-5.3: $1.40 / $0.14 / $4.40
- GLM-5.3 Fast: $2.10 / $0.21 / $6.60
- Kimi K3: $3.00 / $0.30 / $15.00
Dedicated GPUs (per minute): T4 $0.01052, L4 $0.01414, A10G $0.02012, A100 80GB $0.06667,
H100 MIG 40GB $0.0625, H100 80GB $0.10833, B200 180GB $0.16633

Basic is $0/month pay-as-you-go; Pro and Enterprise have no published price and are quote-only. Billing runs on the principle that you only pay for the compute you actually use, down to the minute, with no charge for idle time — considerably fairer than a monthly GPU reservation for bursty traffic.

Coverage

The library is organised by category: roughly 40+ LLMs plus transcription, TTS, image generation, embeddings, biology and computer-use. Named models include GLM-5.3/-Flash/-Fast, DeepSeek V3.1/V4.1/V4 Pro, Kimi K2.5-K3, Qwen3.5/3.6, Gemma 4, MiniMax M2.5/M3, NVIDIA Nemotron 3, GPT OSS 20B/120B and Llama 3.3 70B. No total count is published.

Note that only about a dozen models sit in the token-priced Model APIs — the rest of the library effectively requires a dedicated deployment, with entirely different cost and operational implications. Check which bucket your target model falls into before committing.

Credibility and risk

The customer roster is the most verifiable item here: Canva, Cursor, Notion, HubSpot, Ramp and Patreon are all named on the site, which carries more weight than self-reported scale figures.

On the other side, the free-credit amount is unpublished, Pro/Enterprise are quote-only, and claims such as 2-3x lower latency come from customer case studies, i.e. self-reported. Our recommendation: use the signup credits to measure real prices and latency on the Model APIs first, then decide whether a dedicated deployment is worth it.

Information verified 2026-10-15. Prices, dedicated-GPU rates and the model list reflect what baseten.co and its pricing page displayed on that date.

DeepInfra: low-latency inference for open-source models, transparent and extremely low pricing, supports crypto payment, includes image generation models
Together AI: large-scale open-source model compute aggregation, 100+ models via one API, suited to both research and enterprise use
RunPod: per-second billed GPU cloud, community compute pool, 70% cheaper than AWS
Novita AI: 200+ open-source models, dedicated image generation, low-cost global inference

Quick facts

Pricing model	Model APIs are billed per token (per 1M tokens, input/cache/output): GLM-5.3-Flash $0.15/$0.03/$0.50, DeepSeek V4.1 Flash $0.30/$0.007/$1.20, DeepSeek V4 Pro 0813 $1.32/$0.132/$3.96, GLM-5.3 $1.40/$0.14/$4.40, GLM-5.3 Fast $2.10/$0.21/$6.60, Kimi K3 $3.00/$0.30/$15.00; the Basic plan is $0/month pay-as-you-go; dedicated GPUs bill per minute (T4 $0.01052, A10G $0.02012, A100 80GB $0.06667, H100 80GB $0.10833, B200 180GB $0.16633), and the site states that you only pay for compute you use, down to the minute, with no charge for idle time
Model coverage	The library lists roughly 40+ LLMs plus transcription, TTS, image generation, embeddings, biology and computer-use categories; concrete coverage includes GLM-5.3/-Flash/-Fast, DeepSeek V3.1/V4.1/V4 Pro, Kimi K2.5-K3, Qwen3.5/3.6, Gemma 4, MiniMax M2.5/M3, NVIDIA Nemotron 3, GPT OSS 20B/120B and Llama 3.3 70B — all open-weight; no total model count is published
Latency / SLA	The site leads on production inference that will not break your product or your bank, and a customer case study claims Notion achieved 2-3x lower latency; custom SLAs and self-hosting/VPC are enterprise-only. Vendor and customer claims, with no published latency figures
Mainland direct connect	Proxy required (the site makes no China-related statements)
Best for	Developers / Enterprise
Referral program	No public affiliate programme found (the /referral path returns 404).

Pros

A complete per-token price sheet that also shows cache and output rates — the most predictable cost model of the vendors added in this round
The customer roster is independently checkable: Canva, Cursor, Notion, HubSpot, Ramp and Patreon are all named on the site, which is stronger evidence than self-reported scale numbers
Dedicated GPUs bill by the minute with an explicit no-idle-charge policy, which is fairer than monthly GPU reservations for bursty workloads

Cons

It hosts open-weight models only, with no OpenAI, Claude or Gemini — it cannot replace a relay for access to closed frontier models
The token-priced Model APIs catalogue covers only about a dozen models; most of the library requires a dedicated deployment with a completely different cost structure
The free-credit amount is not published and Pro/Enterprise are quote-only, so costs are hard to forecast before you scale

Compare more AI API relays

See the full comparison board — filter by price tier, model coverage, and mainland direct-connect status.

Back to the comparison board →