Baseten Pricing: Per-Minute GPU Rates, Model API Tokens & the Conversion Math (2026)

Baseten Prices Two Ways, and They Are Not Comparable

Baseten sells inference under two billing models, and the most common mistake is comparing the wrong one against your current provider. Numbers below are from the official Baseten pricing page as of July 2026.

A dollar-per-hour GPU rate and a dollar-per-million-tokens rate live on different axes. You cannot say which is cheaper until you know how many tokens the GPU produces per hour, which depends on the model and how many requests you run in parallel. The rest of this page lists both rate cards, then shows the conversion.

Dedicated Deployment GPU Rates (July 2026)

Every dedicated instance is billed per minute of active compute. Baseten states you pay only for the time the model is using compute, which its FAQ defines to include deploying, scaling, and serving predictions. Hourly equivalents below are the per-minute rate times 60.

The B200 is Baseten's first Blackwell-generation instance. CPU-only instances also exist, from $0.00058/min (1x2) to $0.01382/min (16x64), for pre- and post-processing steps that do not need a GPU. The per-minute rate is the platform-inclusive price: it covers Baseten's autoscaling, deployment tooling, and managed serving layer on top of the raw GPU.

Model API Token Rates (July 2026)

Separate from dedicated GPUs, Baseten runs a serverless Model API on pre-optimized open models, billed per token. This is the rate card you compare directly against Morph, OpenRouter, Fireworks, and other per-token providers. Prices per million tokens.

These are competitive per-token rates. The point of the dedicated option is not to beat them at low volume, it is to beat them at high, steady volume where a saturated GPU produces tokens far below the serverless price. The next section is the math that tells you where that crossover sits.

Converting a Per-Minute GPU Rate to $/M Tokens

A dedicated GPU has a fixed hourly cost regardless of how many tokens it generates. So the effective cost per token is entirely a function of throughput. The equation:

Tokens per second is the number that decides everything. It has two components that pull in opposite directions:

• Single-stream speed is the tokens/s one request sees. For GLM-5.2 on Baseten, Artificial Analysis measures about 332 tok/s. At that speed one B200 stream produces ~1.195M tokens/hr.
• Concurrency multiplies total GPU output. A GPU serving 20 requests at once produces roughly 20x the tokens per hour (until it saturates), which divides the effective per-token cost by the concurrency. This is the entire economic case for dedicated.

So the honest comparison is not "$9.98/hr vs $4.40/M output." It is "$9.98/hr spread across however many tokens I actually push through this GPU per hour." Push a little, the per-token cost is terrible. Saturate it, the per-token cost drops below any serverless rate. The worked example makes the range concrete.

Worked Example: GLM-5.2 on a Dedicated B200

Take a B200 at $9.98/hr running GLM-5.2, using the ~332 tok/s single-stream output speed from Artificial Analysis (July 2026). Here is the effective per-token cost at three concurrency levels, holding total throughput proportional to concurrency until the GPU saturates.

One idle-ish stream on a dedicated B200 costs about $8.35 per million tokens, well above Baseten's own GLM-5.2 Model API at $1.40 input / $4.40 output, and above Morph's morph-glm52-744b at $1.10 / $4.10. Saturate the same GPU at high concurrency and the effective rate falls to roughly $0.42/M, cheaper than any per-token option. The break-even is around 5 concurrent streams for this model on this GPU.

This is why the "is Baseten cheaper?" question has no single answer. If your traffic is steady enough to keep a GPU near saturation, dedicated wins on cost. If it is bursty or low-volume, you pay for idle GPU hours and per-token serverless is cheaper. The shape of your traffic decides, not the rate card.

Autoscaling, Scale-to-Zero, and Cold-Start Billing

Baseten's per-minute granularity is what makes scale-to-zero a real saving rather than a rounding trick: an idle replica at min_replica=0 costs nothing, and you are billed from the minute a replica starts. The catch is the cold start.

The design trade-off is direct. Scale-to-zero removes the idle bill but adds a cold-start latency the docs say "can take minutes for large models," and that wake-up window is billable. To eliminate cold starts, Baseten recommends keeping warm replicas, which puts you back on a 24/7 GPU bill. There is no setting that gives you both instant response and zero idle cost on a dedicated deployment; the Model API is the option that does (someone else keeps the GPU warm and you pay per token).

Effective Cost: Baseten Dedicated vs Morph vs OpenRouter

Put the three ways to run open coding models side by side. Baseten dedicated is per-minute GPU (cost depends on utilization), Morph is a fixed per-token endpoint that stays warm, and OpenRouter routes to third-party per-token providers with a small platform fee. Morph rates below are the published per-token prices per million tokens.

Morph's fast coding models are per-token and always warm, so there is no idle GPU bill and no cold start to design around. Verified Morph rates per million tokens: morph-glm52-744b $1.10 / $4.10, morph-qwen35-397b $0.50 / $3.50 ($0.30 cached), morph-qwen36-27b $0.289 / $2.40, morph-minimax3-428b $0.60 / $2.40, and morph-dsv4flash $0.139 / $0.278. For a coding agent, the deciding factor is usually throughput on the code path and the absence of a cold-start tax, not the raw GPU rate.

The general rule: dedicated GPU pricing wins when you can keep the hardware saturated on steady traffic, and per-token pricing wins on everything variable, bursty, or latency-sensitive. Most teams that aren't running one model at constant high load are cheaper on a per-token endpoint. For a deeper cost breakdown across providers, see LLM cost optimization.

Enterprise, Self-Hosting, and Flex Compute

Baseten's plan tiers are Basic (pay-as-you-go, $0 base), Pro (priority GPU access with volume discounts), and Enterprise (custom SLAs). What Enterprise adds beyond higher limits:

Enterprise per-token and per-GPU rates are negotiated, not published. The published Basic and Pro rates are the ones on the pricing page and in the tables above. New accounts get complimentary signup credits; there is no standing free monthly allowance beyond that.

Frequently Asked Questions

How does Baseten pricing work?

Two models. Dedicated Deployments rent a GPU by the minute (T4 $0.01052/min up to B200 $0.16633/min), billed only while the replica runs. Model APIs charge per token like a standard LLM API (GLM-5.2 $1.40 / $4.40 per million tokens). Dedicated suits steady or latency-sensitive traffic on your own model; Model APIs suit instant, variable serverless usage. Rates as of July 2026 from baseten.co/pricing.

How much does a Baseten H100 or B200 cost?

H100 80GB is $0.10833/min ($6.50/hr) and B200 180GB is $0.16633/min ($9.98/hr). Other instances: T4 $0.63/hr, L4 $0.85/hr, A10G $1.21/hr, A100 80GB $4.00/hr, H100 MIG 40GB $3.75/hr. Billed per minute of active compute, not per full hour.

How do I convert Baseten's per-minute GPU rate to cost per token?

Divide the hourly GPU cost by tokens produced per hour: effective $/M tokens = (GPU $/hr) / (tokens/sec × 3,600 / 1,000,000). A B200 at $9.98/hr serving GLM-5.2 at ~332 tok/s single-stream makes about 1.2M tokens/hr, so ~$8.35/M on one stream. Run 20 streams and it drops to ~$0.42/M. Concurrency is the whole game.

Is Baseten cheaper than per-token providers?

Only when the GPU stays busy. A single low-concurrency dedicated stream can run ~$8/M tokens, more expensive than Baseten's own GLM-5.2 Model API ($1.40 / $4.40) or Morph's morph-glm52-744b ($1.10 / $4.10). Saturate the GPU and dedicated wins. For variable or bursty traffic, per-token is usually cheaper because you never pay for idle capacity.

Does Baseten charge for cold starts and scale-to-zero?

Baseten bills per minute for active compute, which its FAQ says includes deploying, scaling, and serving. Scale-to-zero (min_replica=0) makes idle replicas free, but the next request triggers a cold start the docs warn can take minutes for large models, and that wake-up is billable. Keeping warm replicas avoids cold starts but means paying for idle GPUs 24/7.

What models does Baseten's Model API offer?

As of July 2026: GLM-5.2 ($1.40 / $0.26 cached / $4.40), DeepSeek V4 ($1.74 / $3.48), GLM 5 ($0.95 / $3.15), Kimi K2.6 ($1.00 / $3.90), NVIDIA Nemotron 3 Super ($0.60 / $2.40), and gpt-oss-120b ($0.10 / $0.50), per million tokens. Most offer cached-input discounts. Confirm the current list on the pricing page before building.

What is Baseten's free tier?

New accounts get complimentary signup credits. Basic is pay-as-you-go with no monthly base; Pro adds priority GPU access and volume discounts; Enterprise adds custom SLAs, self-hosting, and flex compute against existing cloud commitments. There is no standing free monthly allowance beyond the initial credits.