Baseten vs DeepInfra (2026): DeepInfra's H100 Is $1.79/hr vs Baseten's $6.50/hr, and Why You'd Still Pay More

Who Wins Per Workload

The decision is per workload, not per platform. DeepInfra wins on price and convenience; Baseten wins wherever running your own model with control over serving is the point.

Quick Comparison

Baseten is the control-and-deployment platform; DeepInfra is the price-and-convenience platform.

Billing Models: Replica-Minute vs Per-Token

This is the core difference, and it decides which platform is cheaper for you.

Baseten dedicated deployments bill per replica-minute. Every running replica costs the underlying GPU rate continuously, whether it serves one request or a thousand. That is efficient at high, steady utilization where you pack a GPU with batched traffic, and expensive when traffic is low or spiky, because warm replicas still bill. Baseten also offers Model APIs for pre-optimized open models priced per token, so you can mix both inside one account.

DeepInfra serverless is the inverse: you pay per input and output token on shared infrastructure and pay nothing when you send nothing. There is no replica to keep warm and no idle cost, but there is no free tier and billing is postpaid. For dedicated needs, DeepInfra rents whole GPUs by the hour at roughly a third of Baseten's rate.

Per-Token Pricing: DeepInfra Undercuts on Every Shared Model

On serverless per-token rates, DeepInfra is cheaper than Baseten's Model APIs on the models both list. DeepInfra also resells Anthropic models, which Baseten does not.

DeepInfra prices cached input separately on most models (DeepSeek V4-Pro cached input is $0.10/M, GLM-5.1 cached is $0.205/M), which matters for agentic workloads that resend a large system prompt each turn. Its catalog also covers DeepSeek V3 ($0.32/$0.89), DeepSeek R1-0528 ($0.50/$2.15), MiniMax-M2.5 ($0.15/$1.15), and Qwen3-Max ($1.20/$6.00). Baseten's Model API catalog includes NVIDIA Nemotron 3 Ultra ($0.60/$2.40) and Nemotron 3 Super ($0.30/$0.75), which DeepInfra does not headline.

Dedicated GPU Pricing: A 3.6x Gap on the H100

If you rent whole GPUs, DeepInfra is far cheaper on every comparable card.

Baseten's posted GPU rates run roughly 3x to 4x DeepInfra's on comparable hardware (H100 $6.50 vs $1.79 is 3.6x; B200 $9.98 vs $2.79 is 3.6x). The gap reflects what you get: Baseten's rate bundles its inference stack, engine selection, autoscaling, and managed deployment. DeepInfra rents closer to raw capacity. Baseten also serves a deeper lineup of small GPUs (T4, L4, A10G, H100 MIG) for cost-tuning smaller models, which DeepInfra does not break out on its dedicated menu.

Cost on a Real Workload

Cold Starts and Scale-to-Zero

DeepInfra's popular models are always warm; Baseten's scale-to-zero trades idle cost for a first-request delay.

DeepInfra serves high-traffic open models on shared endpoints that never scale down, so there is no cold start for the popular models. When a model is saturated you can hit a limit, after which DeepInfra adds capacity. There is no replica for you to keep warm and no idle bill.

Baseten's autoscaling default is min_replica=0. Baseten's docs state that a deployment scaled to zero replicas incurs no charges, but the next request triggers a cold start that can take minutes for large models. During wake-up, billing is per minute even though the replica is not yet serving responses. To eliminate cold starts, the docs recommend min_replica of at least 2 for production, which means paying for warm idle capacity. That is the control-versus-cost tradeoff in one knob.

Neither platform is built for the coding-agent apply loop. If applying model-generated code edits is the bottleneck, that is a different tool (Morph Fast Apply, ~10,500 tok/s, with published benchmarks).

Rate Limits and Account Model

DeepInfra caps concurrency at the account level; Baseten's ceiling is whatever you provision in replicas.

DeepInfra's 200 concurrent-request cap is the practical ceiling on serverless throughput per account; sustained high concurrency past that means either contacting them or moving to dedicated GPUs. Baseten has no shared cap because each deployment is your own replicas, and you scale the count to match load.

Compliance and Data Retention

Both are SOC 2 Type II; the differences are HIPAA, ISO 27001, and how each handles your prompts.

DeepInfra states zero retention on inference: prompts and completions are deleted from disk and memory after a short retention period, with only metadata logged (request ID, cost, sampling parameters). The exception is Google models, where Google logs prompts and responses for abuse detection. DeepInfra is SOC 2 and ISO 27001 certified, with technical and organizational measures for GDPR and HIPAA. Baseten is SOC 2 Type II certified and HIPAA compliant, and because you run dedicated deployments, retention is governed by how you configure them.

When to Use Baseten

• Custom or fine-tuned models in production. Package the model with Truss, pick the GPU, and control autoscaling and serving logic. This is the platform's core.
• You need engine-level control. Per-model selection between TensorRT-LLM and SGLang for a specific model that has to be fast, instead of a fixed shared menu.
• Scale-to-zero with a tunable cost knob. min_replica=0 means no idle bill when traffic stops; min_replica of 2 or more removes cold starts when it matters. You decide per deployment.
• HIPAA and a free start. SOC 2 Type II, HIPAA, and free starter credits for experimentation before you commit.
• High, steady utilization. When a dedicated replica stays near full, per-minute billing you control can beat per-token rates because you keep the batching margin.
• GPT-OSS 120B specifically. Baseten's $0.10/$0.50 Model API rate is the cheaper of the two on this model.

When to Use DeepInfra

• Lowest cost per token. DeepSeek V4-Pro at $1.30/$2.60, GLM-5.1 at $1.05/$3.50, and Kimi K2.6 at $0.75/$3.50 undercut Baseten's Model API on every shared model both list.
• Cheapest dedicated GPUs. H100 at $1.79/hr and B200 at $2.79/hr, roughly a third of Baseten's posted rates.
• Zero setup, no idle cost. Call a popular open model behind an OpenAI-compatible URL and pay only for tokens, with no replica to keep warm.
• Resold Anthropic models. Claude Opus 4.8 ($5/$25), Sonnet 4.6 ($3/$15), and Haiku 4.5 ($1/$5) under the same account, which Baseten does not offer.
• Zero-retention serverless. Prompts and completions deleted after a short window, only metadata logged, with SOC 2 and ISO 27001.

Frequently Asked Questions

Is Baseten or DeepInfra cheaper?

DeepInfra is cheaper on both dedicated GPUs and per-token calls. A dedicated H100 80GB is $1.79/hr on DeepInfra versus $6.50/hr on Baseten, a 3.6x gap. On tokens, DeepSeek V4-Pro is $1.30/$2.60 per 1M on DeepInfra versus $1.74/$3.48 on Baseten, and GLM-5.1 is $1.05/$3.50 versus $1.30/$4.30. The one model where Baseten is cheaper is GPT-OSS 120B at $0.10/$0.50.

What is the difference between Baseten and DeepInfra?

DeepInfra is serverless-first: call a shared per-token endpoint for popular open models with no infrastructure to manage and no free tier. Baseten is a deployment platform: ship a model, pick the GPU, pay per replica-minute, with TensorRT-LLM or SGLang engine selection and scale-to-zero. Baseten also sells per-token Model APIs for pre-optimized open models. DeepInfra gives the lowest price per token; Baseten gives control over how your own model is served.

Does Baseten have cold starts?

Yes, for scaled-to-zero deployments. Baseten's default is min_replica=0, so a deployment with no traffic incurs no charges, but the next request triggers a cold start that the docs say can take minutes for large models, and billing is per minute during wake-up even before the replica serves. The docs recommend min_replica of at least 2 for production to remove cold starts, at the cost of warm idle capacity. DeepInfra's shared endpoints stay warm.

What are DeepInfra's rate limits?

DeepInfra allows 200 concurrent requests per account. Billing is postpaid with monthly invoices plus mid-month invoicing at usage-tier thresholds of $20, $100, $500, $2,000, and $10,000. There is no free tier. Baseten has no shared concurrency cap because you run dedicated replicas and scale them yourself.

Does DeepInfra retain my prompts?

No, except for Google models. DeepInfra states zero retention on inference: prompts and completions are deleted from disk and memory after a short retention period, with only metadata (request ID, cost, sampling parameters) logged. The exception is Google models, where Google logs prompts and responses for abuse detection. DeepInfra is SOC 2 and ISO 27001 certified.

When is Baseten worth 3.6x DeepInfra's GPU price?

When you run your own or fine-tuned model and use what the premium buys: per-model engine selection between TensorRT-LLM and SGLang, custom-model deploys via Truss, scale-to-zero with configurable autoscaling, and HIPAA-eligible deployment. If you only need to call a popular open model behind one API for the lowest price, DeepInfra's $1.30/$2.60 DeepSeek V4-Pro and $1.79/hr H100 win and the Baseten controls go unused.