Fireworks vs Baseten Pricing (2026): Same Per-Token Rates, $6.50/hr vs $7.00/hr H100

Serverless Per-Token Pricing: Model by Model

Both platforms run the same popular open models on a serverless API. On the models that overlap, the rates are within a few cents of each other.

Cached input is discounted on both: Fireworks gives cached input a 50% discount (DeepSeek V4 Pro drops to $0.145/M cached), and Baseten lists cached rates per model (DeepSeek V4 $0.145/M, Kimi K2.6 $0.16/M, GLM 5.1 $0.26/M). Fireworks also bills batch inference at 50% of serverless, which Baseten does not surface as a serverless discount because batch on Baseten runs on the same per-minute dedicated GPU.

Dedicated GPU Pricing: Where They Diverge

This is where the two platforms actually differ. Baseten bills dedicated GPUs per minute and publishes a full ladder from T4 to B200. Fireworks bills on-demand dedicated GPUs per second and lists only the large accelerators.

Baseten is per-minute billed: H100 80GB is $0.10833/min ($6.50/hr), B200 180GB is $0.16633/min ($9.98/hr), A100 80GB is $0.06667/min ($4.00/hr), and there is an H100 MIG 40GB slice at $0.0625/min ($3.75/hr) for models that fit in 40GB. CPU-only instances run from $0.00058/min. Fireworks dedicated GPUs are billed per second and scale to zero, but the catalog starts at the H100 tier.

Neither is cheap by market standards. DeepInfra lists H100 at $1.79/hr and B200 at $2.79/hr; Modal works out to about $3.95/hr for H100. Both Fireworks and Baseten charge for the serving runtime and tooling on top of the silicon, so the per-hour number is not the whole cost story.

Who Wins Per Workload

The deciding question is whether you want to own the GPU or never see it. Match your workload to the column.

Cost on a Real Workload

Take Kimi K2.6 serving 50M output tokens per day (and assume input is small enough to ignore for the estimate). At the shared serverless rate of $4.00/M output, that is the same bill on both: 50M x $4.00/M = $200/day, about $6,000/mo. Per-token, there is no gap to exploit.

Rate Limits

Fireworks publishes hard numbers; Baseten throughput is whatever your replica configuration allows.

On Fireworks you get 10 requests per minute with no payment method and a fixed 6,000 RPM ceiling once you add one. Spending tiers gate the monthly budget: Tier 1 is $50/mo with a valid card, Tier 2 is $500/mo after $50 spent or added, Tier 3 is $5,000/mo, and Tier 4 is $50,000/mo. So you can size your expected throughput against published limits before writing a line of code.

Baseten dedicated deployments do not have a shared RPM cap. Each replica targets a concurrency of 1 request by default (model-specific guidance ranges from 1 for image generation up to 256 for Whisper async batch), and the deployment autoscales from a min replica count to a max replica count you set. The default autoscaling window is 60 seconds, configurable from 10 to 3,600 seconds. Throughput is a function of how many replicas you allow, not a quota you share with other tenants.

Cold Starts and Scale to Zero

Scale-to-zero is the Baseten selling point, and the cold start is the catch.

Fireworks serverless endpoints stay warm on shared infrastructure, so there is effectively no cold start on a catalog model. The trade is that you do not control the GPU.

Baseten scales to zero by default (min_replica=0). The docs are explicit: a deployment scaled to zero incurs no charges, but the next request triggers a cold start that "can take minutes for large models," and during wake-up "billing is per minute even though the replica isn't yet serving responses." To eliminate cold starts in production, Baseten's docs recommend min_replica of at least 2, which keeps GPUs warm and removes the idle savings. So scale-to-zero is real, but for latency-sensitive production you are usually back to paying for warm capacity.

Fine-Tuning

Fireworks runs fine-tuning as a managed per-token service; Baseten runs training on the same per-minute GPU you deploy to.

On Fireworks, DPO is 2x the SFT rate at each size, and the fine-tuned model then serves at the base model's per-token rate with no inference premium. Baseten does not bill fine-tuning per token: you train on a dedicated GPU at the same per-minute rate you serve on (H100 $6.50/hr, A100 80GB $4.00/hr), so a fine-tune costs whatever the GPU costs to keep running. Fireworks is the cleaner choice for a LoRA on a catalog base model you call intermittently; Baseten wins when you want to train and serve any architecture on hardware you control.

Compliance and Data Retention

Both clear the standard enterprise bar. Fireworks adds an explicit zero-retention default.

Fireworks states it "does not log or store prompt or generation data for any open models without explicit user opt-in," with TLS 1.2+ in transit and AES-256 at rest. Both providers are SOC 2 Type II certified and HIPAA compliant. If a no-prompt-logging default is a hard requirement, Fireworks documents it on open models; Baseten documents SOC 2 Type II and HIPAA on its pricing page but does not advertise a zero-retention default in the same terms.

When to Use Fireworks AI

• Zero-ops token APIs on catalog models. Call DeepSeek V4, Kimi K2.6, GLM 5.1, or GPT-OSS without touching a GPU, on warm endpoints with no cold start.
• Published, predictable rate limits. 6,000 RPM ceiling and spending tiers from $50/mo to $50,000/mo let you size capacity before launch.
• Managed fine-tuning. LoRA SFT from $0.50/M tokens, served at the base model rate with no inference premium.
• Cost-sensitive batch jobs. Batch inference at 50% of serverless and cached input at 50% cut the bill on large offline runs.
• Documented zero-retention. No prompt or generation logging on open models without opt-in, with TLS 1.2+ and AES-256.

When to Use Baseten

• Custom or non-catalog models. Deploy any architecture to a dedicated GPU instead of being limited to a serverless catalog.
• Cheap small GPUs. T4 at $0.63/hr, L4 at $0.85/hr, A10G at $1.21/hr, A100 80GB at $4.00/hr, all per-minute billed, none of which Fireworks surfaces.
• Slightly cheaper big GPUs. H100 80GB at $6.50/hr beats Fireworks' $7.00/hr; B200 at $9.98/hr beats $10.00/hr.
• Scale-to-zero on bursty workloads. No idle charge at min_replica=0, as long as minutes-long cold starts are acceptable.
• Cheaper open-weight serverless. GPT-OSS 120B at $0.10/$0.50 and GLM 5.1 at $1.30/$4.30 undercut Fireworks on those models.

Neither 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).

Frequently Asked Questions

Is Fireworks AI or Baseten cheaper?

On serverless per-token APIs they are nearly identical: DeepSeek V4 Pro is $1.74/M input and $3.48/M output on both, and Kimi K2.6 is $0.95/$4.00 on both. Baseten is slightly cheaper on GPT-OSS 120B ($0.10/$0.50 vs $0.15/$0.60) and GLM 5.1 ($1.30/$4.30 vs $1.40/$4.40). On dedicated GPUs Baseten is cheaper per hour: H100 80GB at $6.50/hr vs Fireworks $7.00/hr, and B200 at $9.98/hr vs $10.00/hr. Both sit well above DeepInfra, which lists H100 at $1.79/hr.

What is the main difference between Fireworks AI and Baseten?

Fireworks is serverless-first: a per-token API across a curated catalog, plus on-demand dedicated GPUs from $7.00/hr that scale to zero. Baseten is dedicated-first: deploy any model to per-minute GPUs ($0.63/hr T4 to $9.98/hr B200) with scale-to-zero by default, plus a serverless Model API. Fireworks optimizes for zero-ops token serving; Baseten optimizes for control over the GPU and custom deployments.

What are the rate limits on Fireworks AI vs Baseten?

Fireworks publishes hard numbers: 10 RPM without a payment method, a 6,000 RPM ceiling with one, and spending tiers from $50/mo (Tier 1) to $50,000/mo (Tier 4). Baseten does not publish a flat RPM cap because dedicated deployments scale on a concurrency target (default 1 request per replica) up to your configured max replicas, so throughput is set by replica count rather than a shared quota.

Does Fireworks AI or Baseten handle cold starts better?

Fireworks serverless endpoints stay warm on shared infrastructure, so there is effectively no cold start. Baseten scales to zero by default (min_replica=0), which means the first request after idle triggers a cold start that the docs say can take minutes for large models, and billing is per minute during wake-up. Baseten's docs recommend min_replica of at least 2 for production to eliminate cold starts, which removes the idle savings.

What does Baseten cost to fine-tune and serve a custom model?

Baseten bills the GPU per minute for both training and serving on dedicated deployments, so a fine-tune costs whatever the GPU costs to keep running: H100 at $6.50/hr, A100 80GB at $4.00/hr. Fireworks runs fine-tuning as a managed per-token service: up to 16B is $0.50/M tokens LoRA SFT or $1.00/M full SFT, 16.1B to 80B is $3.00/$6.00, and over 300B is $10.00/$20.00, with DPO at 2x the SFT rate. On Fireworks the fine-tune then serves at the base model's per-token rate.