Predibase pricing

About

Predibase is a fine-tuning and serving platform for open-source large language models, founded in 2021 by Devvret Rishi (CEO), Piero Molino (chief science officer), and Travis Addair (CTO). Its DNA is open source: Molino created Ludwig, the declarative deep-learning framework, and Addair maintained Horovod, the distributed-training library — and Predibase commercialized that lineage into a managed platform for customizing and serving smaller, cheaper open models instead of paying frontier-API rates. Its best-known pieces are LoRA Land (a collection of fine-tuned adapters that beat GPT-4 on narrow tasks), LoRAX (a serving stack that packs many LoRA adapters onto one GPU), and Reinforcement Fine-Tuning (RFT) built on GRPO.

The company raised over $28 million from Felicis, Greylock, and Sancus Ventures. In June 2025, Rubrik (NYSE: RBRK) — the data-security and backup vendor — announced it would acquire Predibase, reportedly for between $100 million and $500 million, “to accelerate agentic AI adoption from pilot to production at scale.” Rubrik’s pitch is that bolting Predibase’s fine-tuning and serving onto its data-governance footprint lets enterprises build accurate AI agents on top of their own governed data.

The acquisition is the single most important pricing fact here: the Predibase product and documentation still operate, but the standalone rate card is no longer cleanly public. predibase.com/pricing now redirects toward Rubrik’s site (which returns an access-denied wall to automated capture), and the predibase.com TLS certificate has expired. The prices below are therefore reconstructed from Predibase’s own documentation and reputable third-party summaries that quoted the published rate card; treat them as directional and confirm with sales. For current details, see Predibase’s docs.

Pricing summary : How Predibase’s pricing model works

Predibase is pure usage-based, metered on two distinct units rather than seats or a flat subscription:

1. Fine-tuning — per 1M tokens processed. You pay for the tokens your training job consumes, with the rate scaling by model size and technique. Standard supervised fine-tuning (SFT) with LoRA starts at $0.50 per 1M tokens (models up to 16B); Reinforcement Fine-Tuning (RFT GRPO) is the premium meter at up to $20.00 per 1M tokens.
2. Serving — per second of GPU time. Dedicated/private serverless deployments bill by the second for whatever GPU replicas are scaled up, from $2.14/GPU/hr (L4) to $4.80/GPU/hr (A100). Autoscaling and scale-to-zero mean idle deployments stop charging, and LoRAX serves many adapters on one GPU.
3. Free shared inference + trial. The Developer tier includes a 30-day, $25 free-credit trial and free rate-limited shared serverless endpoints (1M tokens/day, 10M tokens/month) for testing. Enterprise (H100s, no rate limits, VPC) is sales-quoted.

What makes this different: most managed-LLM platforms price inference per output token; Predibase prices the work that produces a custom model (fine-tuning tokens) and the infrastructure that serves it (GPU-seconds) separately. The LoRAX angle is the kicker — because dozens of fine-tuned adapters share one GPU, the per-second serving meter spreads a single GPU’s cost across many models, which is the economic argument for owning small fine-tunes instead of renting one big frontier API.

Pricing by product

Fine-tuning, per 1M tokens processed (reconstructed from Predibase docs, June 2026):

Dedicated serving (private serverless deployments), per GPU-hour, billed per second:

Sales motions across products: the Developer tier (free trial + free shared inference) and pay-as-you-go fine-tuning/serving are self-serve (PLG); H100-class hardware, removal of shared-endpoint rate limits, VPC deployment, and custom credits are sales-led Enterprise (sales@predibase.com).

Hidden costs : What Predibase users actually pay

Predibase’s two-meter model is clean on paper, but the real bill mixes training and serving:

The two cost traps are familiar to anyone running GPU infra. First, dedicated serving is the recurring spend — a single A100 left running around the clock dwarfs the cost of the fine-tune that produced the model, so the savings story depends on packing many adapters onto one GPU (LoRAX) and aggressively scaling to zero. Second, RFT is dramatically pricier than SFT: at up to $20.00 per 1M tokens, a reinforcement fine-tune can cost 40x a comparable LoRA SFT run, so reaching for RL when SFT would do is an easy way to blow the budget.

Want to estimate your own Predibase bill? Use the Predibase pricing calculator to model your fine-tuning and serving costs based on tokens and GPU hours.

Pricing evolution : Predibase pricing history and changes

Cadence

Tracked range: 2024–present. Note the live rate card became hard to capture after the June 2025 Rubrik acquisition (cert expired, pricing redirects); figures are from Predibase docs and third-party summaries.

Notable changes

• Early 2024 — Predibase introduced what it billed as the first purely serverless solution for fine-tuned LLMs, moving to per-1M-token fine-tuning pricing (early card roughly $0.36 to $3.21 per 1M tokens) plus free rate-limited shared inference for testing.
• 2025 — Added Reinforcement Fine-Tuning (GRPO) as a premium meter at $10.00–$20.00 per 1M tokens, well above SFT, pricing the heavier compute of RL-based tuning directly into the token rate.
• June 25, 2025 — Acquired by Rubrik (reported $100M–$500M). Product continues, but the standalone public rate card became hard to reach as predibase.com/pricing began redirecting toward Rubrik and the predibase.com certificate expired.

The trajectory is a small, transparent open-model-tuning platform that built a clean two-meter usage model, then folded into a larger security/data-governance vendor — at which point its formerly public pricing receded behind the acquirer’s chrome.

What’s unique : Predibase’s distinctive pricing mechanics

1. Two meters, not one. Predibase prices fine-tuning (per 1M tokens processed) and serving (per second of GPU) as separate units. Most managed-LLM APIs collapse everything into per-output-token; Predibase makes the cost of building a custom model explicit and distinct from running it.

2. Technique-tiered token pricing. The per-token fine-tuning rate isn’t flat — it steps up by model size and method, from $0.50 (small SFT LoRA) to $20.00 (RFT GRPO). The price card itself nudges you toward the cheapest technique that solves the task.

3. Adapter-packing economics via LoRAX. Because LoRAX serves many fine-tuned adapters on a single GPU, the per-second serving meter can amortize one GPU across dozens of customized models — the structural argument for owning small fine-tunes over renting one large frontier API.

Strengths & weaknesses

Billing UX : Predibase billing controls and transparency

• Billing controls — Usage is pay-as-you-go: per-1M-token fine-tuning jobs and per-second dedicated serving, with autoscaling and scale-to-zero so idle deployments stop charging. The Developer tier runs on a $25 free-credit pool; Enterprise adds credits via sales.
• Usage visibility — Settings → Billing surfaces invoices and remaining credits; shared-endpoint usage is capped by explicit daily/monthly token rate limits (1M/day, 10M/month on Developer) so testing can’t silently run up a bill.
• Payment options — Self-serve credit-based billing for Developer/usage; sales-led contracts and invoicing for Enterprise (H100s, VPC, removed rate limits) via sales@predibase.com.

Strategic wins : Why Predibase’s pricing decisions worked

1. Pricing the “own a small fine-tune” argument directly

By metering fine-tuning per token and serving per GPU-second — and packing many adapters onto one GPU with LoRAX — Predibase made the cost case for customizing small open models instead of paying frontier per-token rates. The pricing structure is the pitch. See how AI companies structure pricing.

2. A genuinely low-friction on-ramp

A 30-day, $25-credit trial plus free rate-limited shared inference lets developers fine-tune and test before paying a cent — a PLG wedge that suits a tool aimed at ML engineers who want to prove value on a real task first. Related: outcome-based pricing trends.

3. Letting the meter pick the technique

Tiering fine-tuning by size and method (SFT LoRA at $0.50 up to RFT at $20.00 per 1M tokens) prices compute intensity transparently, steering teams to the cheapest technique that works rather than defaulting to the most expensive. See choosing the right usage metric.

Areas to improve : Gaps in Predibase’s pricing approach

1. Post-acquisition pricing opacity

The biggest gap is no longer a fee — it’s that the public rate card receded after the Rubrik deal (redirects, expired cert). For a tool that won trust on transparent per-token pricing, a clearly maintained standalone pricing page would reduce uncertainty. See bill shock and cost unpredictability.

2. Serving is the silent recurring cost

Dedicated GPU serving at $2.14–$4.80/GPU/hr dwarfs the cost of the fine-tune that produced the model. Clearer in-product guardrails (default scale-to-zero, idle-deployment warnings) would help teams avoid leaving an A100 running 24/7 by accident.

3. Enterprise gating of H100s

Capping H100-class serving to sales-led Enterprise means the highest-throughput option has no self-serve price, which forces a sales conversation just to benchmark the fastest hardware — friction for teams that want to size a deployment before committing.

Key takeaways

1. Predibase is pure usage-based on two meters — per-1M-token fine-tuning ($0.50 up to $20.00) and per-second GPU serving ($2.14–$4.80/GPU/hr) — with a free $25 trial. For the underlying model, see the introduction to usage-based pricing.
2. Fine-tuning is priced by technique and size, so the cheapest method that solves the task is the cheapest bill — RFT can cost 40x a base SFT LoRA run.
3. Serving, not training, is the recurring spend — a dedicated A100 left running around the clock costs more than the fine-tune, so scale-to-zero and LoRAX adapter-packing matter.
4. The free $25 trial plus rate-limited shared inference is a real PLG on-ramp for ML engineers to prove value before paying.
5. Rubrik’s June 2025 acquisition is the dominant pricing fact — the product continues but the standalone rate card is no longer cleanly public.

UBP implications

1. Separate the meter for building from the meter for running. Predibase prices fine-tuning (tokens) distinctly from serving (GPU-seconds), which makes the true cost of a custom model legible — a useful pattern for any platform where production and consumption are different cost drivers.
2. Let the price card steer behavior. Tiering by technique and model size nudges customers to the cheapest method that works, aligning the meter with both cost and customer outcome.
3. Acquisition can quietly erase pricing transparency. When a usage-priced tool is absorbed by a larger vendor, its formerly public rate card can recede behind the acquirer — a reminder that “public pricing” is only as durable as the company publishing it.

Sources

• Predibase pricing page — redirects toward Rubrik / bot-walled post-acquisition (accessed 2026-06-15)
• Predibase docs — Usage & Billing (accessed 2026-06-15)
• Predibase docs — Inference overview (accessed 2026-06-15)
• Rubrik — Rubrik to Acquire Predibase (accessed 2026-06-15)
• TechCrunch — Rubrik acquires Predibase (accessed 2026-06-15)
• eesel AI — What is Predibase? (2025 overview, quotes the rate card) (accessed 2026-06-15)

Bottom line

Predibase is a clean example of usage-based pricing built on two distinct meters: per-1M-token fine-tuning (from $0.50 for small LoRA jobs up to $20.00 for reinforcement fine-tuning) and per-second GPU serving ($2.14–$4.80/GPU/hr), wrapped in a free $25-credit Developer trial and a sales-led Enterprise tier for H100s and VPC. The structure is the pitch — meter the work of building a small custom model separately from the GPU that serves it, and use LoRAX to spread one GPU across many adapters. The dominant caveat is the June 2025 Rubrik acquisition: the product still runs, but the standalone rate card has receded behind the acquirer, so confirm current numbers with Predibase. Browse the pricing blueprint for more fully-researched company profiles, or compare Predibase against other AI platform companies.

Want to compare Predibase against other AI infrastructure companies? Browse the pricing blueprint.