Weights & Biases pricing

About

Weights & Biases (W&B) is the experiment-tracking layer that became an MLOps standard. Founded in 2018 by Lukas Biewald and Chris Van Pelt, its open SDK (wandb) — a few lines you drop into a training script to log metrics, hyperparameters, and artifacts — spread to over 1 million users including OpenAI, Meta, NVIDIA, and Toyota. The platform has since grown four product surfaces: W&B Models (experiment tracking, sweeps, registry, artifacts), W&B Weave (LLM observability, tracing, evaluations, and LLM-as-a-judge scorers for GenAI apps), W&B Serverless Inference (a per-token endpoint for 31 open-weight models), and the new CoreWeave Sandboxes.

The headline corporate fact: on May 5, 2025, CoreWeave completed its acquisition of W&B for a reported ~$1.7 billion (W&B was last valued at $1.25B in 2023). The deal slots a developer-facing platform on top of CoreWeave’s GPU cloud, creating a path from raw compute to experiment tracking to inference. CoreWeave publicly committed to keeping the platform interoperable — customers can still use any infrastructure provider, deployment option, and foundation model — but the pricing now bundles CoreWeave Sandboxes and a first-party inference endpoint, both signs of the parent’s footprint.

For the most current information, visit Weights & Biases.

Pricing summary : How Weights & Biases’s pricing model works

W&B runs a freemium plan fee plus three usage meters. You pick a tier — Free ($0), Pro (starts at $60/month), or custom Enterprise — and each bundles a quota of seats, storage, and Weave data ingestion. The plan fee is only the floor; the variable bill comes from the meters.

The three things that actually scale your bill:

• Storage (GB) — model artifacts, datasets, and run history. Free includes 5 GB/mo, Pro 100 GB/mo, and overage runs $0.03/GB.
• W&B Weave data ingestion (GB) — the volume of logged LLM traces, evals, and monitoring data for GenAI apps. Free includes 1 GB/mo, Pro 1.5 GB/mo, and overage is $0.10/MB (i.e. ~$100/GB — far pricier than model storage, because it’s the live observability stream).
• W&B Inference (per token) — Serverless calls to 31 open-weight models, billed per token in arrears, from $0.03 per 1M input (OpenAI GPT OSS 20B / 120B) up to $1.74 input (DeepSeek V4-Pro) / $4.40 output (Z.AI GLM 5.1). Flagship models add a discounted cached-input rate; Pro ships a $5/mo credit.
• ARIA (per token) — a newer token-priced product, listed as its own metered line on the compare table; free “for a limited time” on Free and Pro, included on Enterprise. Its rate card is already published: ARIA runs on OpenAI models and adds an Agent Token Rate of $0.50 per 1M tokens on top of model rates, for combined rates of $3.00 in / $0.75 cached / $15.50 out (gpt-5.4) up to $5.50 / $1.00 / $30.50 (gpt-5.5) per 1M tokens.

Seats sit alongside these: Free allows up to 5 model seats, Pro up to 10. The legacy W&B Models meter — “tracked hours” (cumulative compute hours of logged training) — still surfaces in the academic license (“unlimited tracked hours”) and the older per-user tiering, but it is no longer a billed meter: W&B’s billing docs state tracked hours are unlimited on every plan. The current published cloud tiers lead with storage + ingestion + inference.

What makes this different: most observability tools meter one thing. W&B meters four (seats, storage, traces, tokens) because it fused an MLOps tool, an LLM-observability tool, and an inference endpoint into one bill — so modeling your cost means modeling all four levers at once.

Pricing by product

Pro is gated to “early-stage teams fewer than 50 employees” — exceed that and W&B requires a move to Enterprise. Academic research is free forever (all Pro features, unlimited tracked hours, 200 GB storage, 25 GB/mo Weave ingestion, up to 100 seats).

Sales motions across products: self-serve credit-card checkout (Free/Pro), product-led growth via the free wandb SDK, and sales-led (Enterprise custom quotes, annual invoicing, single-tenant/private hosting).

W&B Serverless Inference per-token examples (per 1M tokens, as of the 2026-07-29 rate card): Z.AI GLM 5.2 $0.76 in / $0.14 cached / $2.42 out; Z.AI GLM 5.1 $1.40 in / $0.26 cached / $4.40 out; Z.AI GLM 5 $1.00 in / $3.20 out; MiniMax M3 $0.23 in / $0.05 cached / $0.96 out (added 2026-07-29); Moonshot AI Kimi K2.7 Code $0.71 in / $0.15 cached / $3.50 out; Moonshot AI Kimi K2.6 $0.65 in / $0.15 cached / $3.41 out; NVIDIA Nemotron 3 Ultra $0.75 in / $0.15 cached / $2.75 out; DeepSeek V4-Pro $1.74 in / $0.14 cached / $3.48 out; DeepSeek V4-Flash $0.14 in / $0.07 cached / $0.28 out; Google Gemma 4 31B $0.10 in / $0.34 out (no cached tier); OpenAI GPT OSS 20B $0.03 in / $0.13 out; OpenAI GPT OSS 120B $0.03 in / $0.17 out; JetBrains Mellum2 12B A2.5B and IBM Granite 4.1 8B both $0.05 in / $0.10 out; Qwen3 235B A22B-2507 $0.10 in / $0.10 out; Meta Llama 3.1 8B $0.22 in / $0.22 out. Cheaper models publish a two-part rate (input / output); newer flagship models add a discounted cached input tier in the middle.

Two vendor surfaces, two rate cards — mostly reconciled. The marketing “Available models” list on wandb.ai/site/inference/ and the canonical Token-Based Pricing card at wandb.ai/site/pricing/inference used to disagree on OpenAI’s GPT-OSS rates; as of the 2026-07-29 capture both surfaces now show identical GPT OSS 20B ($0.03 in / $0.13 out) and GPT OSS 120B ($0.03 in / $0.17 out) prices, and the same 15-45% cuts landed on both pages for GLM 5.2, Kimi K2.7 Code, Kimi K2.6, and Gemma 4 31B. One gap remains: the rate card carries 31 models (adding MiniMax M3 alongside the pre-existing Z.AI GLM 5), while the marketing list has 30 and still omits Z.AI GLM 5 entirely. W&B’s own docs point Pro overage billing at the rate card: “W&B bills Pro plan users for overages monthly, based on model-specific pricing” links to wandb.ai/site/pricing/inference. Price off the rate card, not the model gallery, when the two ever diverge again.

Hidden costs : What Weights & Biases users actually pay

The $60 Pro fee is rarely the real bill — the three meters are. The single biggest surprise is Weave data ingestion overage at $0.10/MB. That works out to roughly $100 per GB, and an LLM app that logs verbose traces, large prompts, and eval results can produce several GB a month without anyone noticing — so a team that budgeted “$60 for W&B” can see a four-figure ingestion line.

For a Pro team running real GenAI workloads, the math is roughly:

Other line items to watch: the 50-employee gate means growing teams are pushed off self-serve Pro into custom Enterprise quotes; HIPAA, SSO, audit logs, and BYO-bucket are Enterprise-only, so any regulated use jumps straight to a sales contract; and Inference is a separate per-token meter layered on top of everything — the $5/mo Pro credit barely dents real usage. CoreWeave Sandboxes also carry only a limited-time credit ($10/mo on Free / $25/mo on Pro, customizable on Enterprise) before becoming a paid add-on.

Want to estimate your own W&B bill? Use the Weights & Biases pricing calculator to model costs based on usage patterns. For broader context on metered observability spend, see the LLMOps cost-tracking gaps teams hit.

Pricing evolution : Weights & Biases pricing history and changes

Cadence

Tracked range: 2018–present. Snapshot images attached for the 2026 captures; earlier tiers reconstructed from archived pricing pages and third-party teardowns (ZenML’s WandB pricing guide, G2’s Weights & Biases pricing listing).

Notable changes

• 2018 — Launched the wandb SDK; free logging plus paid team/enterprise seats and a “compute hours” meter.
• 2022 — Meter terminology shifted to “tracked hours”; Starter/Pro sold at ~$50/user/mo, tiered by cumulative tracked hours (to $100/$150), $1/hr overage, $0.03/GB storage.
• 2024 — W&B Weave shipped, introducing Weave data ingestion (GB) as a new value metric for GenAI observability.
• 2025-05-05 — CoreWeave completed its ~$1.7B acquisition, folding W&B into its GPU cloud with an interoperability pledge.
• 2026 — Tiers restructured to Free / Pro ($60/mo) / Enterprise, metered on storage ($0.03/GB), Weave ingestion ($0.10/MB), and per-token Serverless Inference, plus CoreWeave Sandboxes credits.
• 2026-07-14 — Added ARIA, a new token-priced product, as its own line on the pricing compare table (free “for a limited time” on Free and Pro, included on Enterprise); refreshed the Serverless Inference catalog, with the top output rate now $4.40/1M (Z.AI GLM 5.2, up from the ~$2.75 ceiling). Core tiers and all three meters were unchanged — an additive product launch, not a repricing.
• 2026-07-21 — Repriced DeepSeek V4-Flash from $0.01 in / $0.01 out to $0.14 input / $0.07 cached / $0.28 output per 1M tokens — 14x on input, 28x on output — and gave it a cached-token tier the old two-part rate did not have. Because V4-Flash had been the catalog’s undercut, the cheapest published input rate across the whole model list rose 3x, from $0.01 to $0.03 (OpenAI GPT OSS 20B). Read against the rest of the card, the $0.01 rate looks like an introductory price on an “Experimental”-badged 1M-context model rather than a durable floor: the reset lands V4-Flash in line with the cheap tier while still leaving it an order of magnitude under DeepSeek V4-Pro ($1.74 in / $3.46 out). The practical bite is on long-context and agent loops — a full 1M-token prompt on V4-Flash went from about a cent to about 14 cents (7 cents on cached input), so the models that made 1M context feel free are no longer free. Nothing on the cloud page moved: Free $0, Pro from $60/mo, custom Enterprise, storage $0.03/GB, Weave ingestion $0.10/MB, the $5/mo Pro Inference credit, Sandboxes credits, and free-for-now ARIA were byte-identical to the prior capture. The structural thesis is unchanged — what the week clarified is which meter carries price risk.
• 2026-07-29 — Eight days after the V4-Flash reset, the Inference rate card moved again — this time down for most of the models it touched. Z.AI GLM 5.2 fell roughly 45% on both input ($1.39 → $0.76) and output ($4.40 → $2.42); Moonshot AI’s Kimi K2.7 Code (-24% in / -12.5% out) and Kimi K2.6 (-32% in / -15% out) both dropped; Google Gemma 4 31B fell modestly on input ($0.12 → $0.10) and lost its $0.09 cached-input tier entirely. The one increase: OpenAI GPT OSS 120B’s output rate rose 21% ($0.14 → $0.17) even as its input rate fell to match the 20B model’s $0.03 floor. A new model, MiniMax M3, joined the catalog at $0.23 in / $0.05 cached / $0.96 out, taking the roster from 30 to 31. The move also quietly closed a discrepancy flagged the week before: the marketing “Available models” page and the canonical Token-Based Pricing rate card had disagreed on GPT-OSS pricing since at least August 2025 (marketing still showing $0.05/$0.20 and $0.15/$0.60 launch rates); both surfaces now show the same $0.03/$0.13 and $0.03/$0.17 figures — though the marketing list still omits Z.AI GLM 5 outright, so the two pages are reconciled on price but not on catalog completeness. Cloud tiers, storage, and Weave ingestion were untouched.

What’s unique : Weights & Biases’s distinctive pricing mechanics

1. Four meters under one bill. Seats, storage (GB), Weave ingestion (GB), and Inference (tokens) all bill independently. That reflects W&B’s history as three products (Models, Weave, Inference) stitched together — and it means buyers must forecast four distinct levers, not one usage number. The July 2026 arrival of ARIA — a second token-priced product, free for a limited time — hints at a fifth line to come, so the surface buyers must model is still widening rather than consolidating; watch for the day ARIA’s introductory pricing ends.

2. The asymmetric storage-vs-ingestion price. Model storage overage is a cheap $0.03/GB; Weave trace ingestion is $0.10/MB — roughly 3,000x more per byte. W&B prices the live observability stream (hot, queryable LLM traces) far above cold model artifacts, a deliberate signal of where the marginal cost and the value sit.

3. The 50-employee Pro tripwire. Pro is explicitly restricted to teams under 50 employees; outgrow it and you must move to Enterprise. It’s a rare case of a vendor coding its self-serve-to-sales handoff directly into the pricing page rather than relying on usage thresholds.

4. The stable half and the volatile half. The published plan fees ($0 / $60 / custom) and the two data meters ($0.03/GB storage, $0.10/MB Weave ingestion) have held byte-identical across every 2026 capture. The per-model Inference rate card is the part that actually moves — and by 2026-07-29 it had moved twice in eight days: 2026-07-21 reset DeepSeek V4-Flash from $0.01 to $0.14 in / $0.28 out, and 2026-07-29 cut five other models 15-45% (Z.AI GLM 5.2, Moonshot Kimi K2.7 Code and K2.6, Google Gemma 4 31B) while raising OpenAI GPT OSS 120B’s output rate 21% and adding a new model, MiniMax M3. Two events in eight days, one an increase and the other mostly cuts, confirms this isn’t a single surprising spike but a standing cadence — rates move in both directions on this card, without notice, on a roughly weekly rhythm. The “Experimental” badge W&B leaves on those models functions as a standing repricing licence — it flags which rates are provisional without ever committing to a notice period. So the half of the bill buyers can diff is the stable half, and the half that reprices (in either direction) is the one they have to re-check per model, per week.

Strengths & weaknesses

Billing UX : Weights & Biases billing controls and transparency

• Billing controls — Pro is billed monthly or annually upfront; model seats are prorated when added mid-term, with no credit for removed seats. Weave ingestion, Inference, and storage are billed monthly in arrears on actual usage. Enterprise is invoiced annually upfront.
• Usage visibility — The platform surfaces storage and ingestion consumption against quota; the published pricing table and FAQ explain how each meter is calculated. Per-model Inference rates are listed on a dedicated Token-Based Pricing rate card. The billing settings page charts usage per category (Storage, Weave, Inference, Training, Sandboxes) with gold/red threshold lines at 85% and 100% of allowance, and emails usage and spending alerts to billing admins at those thresholds.
• Inference spending caps — Every account tier carries a default Inference spending cap: $100/month on Free, $6,000/month on Pro, and $700,000/year on Enterprise, raised only by manual review with an account executive. W&B requires prepayment for paid Inference access, and you can set your own monthly Inference budget (requests stop being served once it is hit) plus an email notification threshold.
• Tracked hours are not billed — Despite surviving in the academic-license copy, tracked hours are unlimited and unbilled on every current plan, so they carry no usage alert.
• Payment options — Self-serve credit-card checkout and a 30-day Pro free trial; Enterprise moves to annual invoicing with contract terms, single-tenant hosting, and security addenda (HIPAA BAA, CMEK).

Strategic wins : Why Weights & Biases’s pricing decisions worked

1. Free SDK as the category land-grab

The free wandb logging SDK embedded W&B into ML training workflows at OpenAI, Meta, and thousands of labs before any sales motion. Owning the developer’s first import wandb made W&B the default — distribution CoreWeave ultimately paid a reported ~$1.7B to acquire. See usage-based pricing fundamentals.

2. Free-forever academic seeding

Handing researchers the full Pro feature set for free — plus 200 GB storage, 25 GB/mo Weave ingestion, and 100 seats, roughly 2x the storage and 16x the ingestion a paying Pro team gets — trains the next generation of ML practitioners on W&B, who then bring it into industry. It’s a long-cycle funnel disguised as generosity. Related: how AI companies structure pricing.

3. Pricing the hot stream above the cold store

Charging $0.10/MB for live Weave traces but $0.03/GB for model artifacts aligns price with the genuinely expensive thing (queryable, retained observability data). It’s a clean example of choosing the right usage metric per product surface rather than one blunt unit.

Areas to improve : Gaps in Weights & Biases’s pricing approach

1. Four-meter forecasting

Seats + storage + ingestion + tokens is a lot to model. Teams routinely undershoot the Weave ingestion line because $0.10/MB doesn’t read as expensive until it’s a multi-GB monthly bill. Better in-product forecasting would reduce bill shock.

2. Governance gated to Enterprise

SSO, HIPAA, audit logs, and BYO-bucket all require a custom Enterprise contract. A security-conscious 20-person team that needs SSO has no self-serve path — it jumps from $60 Pro straight to sales.

3. The 50-employee cliff

Coding a headcount cap into Pro is clean for W&B but abrupt for customers: a team that crosses 50 employees is forced into renegotiation regardless of actual usage, which can feel punitive mid-project.

4. Rate-card moves land without a changelog

The 2026-07-21 DeepSeek V4-Flash reset (14x input, 28x output) arrived as a silent edit to the model catalog — no dated note on the inference page, no deprecation window on the old rate, and the model still carries the same “Experimental” badge it had at $0.01. Teams that built an agent loop around the cheapest 1M-context model found out by reading their next invoice. Eight days later, on 2026-07-29, the pattern repeated: five more models were repriced 15-45% and a new model (MiniMax M3) joined the catalog, again with no changelog entry. That second move also quietly fixed a real problem — the marketing model list and the canonical rate card had disagreed on OpenAI GPT-OSS pricing since at least August 2025 — but even a customer-favorable correction like that shipped undocumented, so buyers who’d been pricing off the (wrong) marketing page had no signal the numbers had changed underneath them. Publishing per-model rate history, or even a short notice period before a move of this magnitude in either direction, would cost W&B nothing and would let buyers re-plan instead of reconcile.

Key takeaways

1. A free SDK can be the whole moat. W&B’s import wandb default drove the adoption that made it the MLOps standard — and the reported ~$1.7B CoreWeave exit.
2. Meter each surface on its own value unit — then version every rate card. Storage GB, Weave trace ingestion, and per-token Inference each price what that surface actually costs, but only the plan page gets versioned: the Inference card reset DeepSeek V4-Flash from $0.01 to $0.14 in / $0.28 out on 2026-07-21 with no notice, lifting the whole catalog’s cheapest input rate 3x to $0.03 — then moved again eight days later on 2026-07-29, cutting five other models 15-45% while raising GPT OSS 120B’s output 21%. Two undocumented moves in eight days confirm this is a standing cadence, not a one-off: a loss-leading catalog entry is a price increase waiting to happen, and a discounted one is a cut waiting to happen just as quietly.
3. Price the hot stream above the cold store. $0.10/MB for live traces vs $0.03/GB for artifacts is a deliberate, defensible asymmetry.
4. Free-for-academics is a multi-year funnel. Switching off the paid meter for research seeds the practitioners who later buy.
5. Acquisition reshaped the bundle, not the philosophy. CoreWeave added Sandboxes and first-party Inference, but the freemium-plus-meters structure held.

UBP implications

1. When you sell multiple products, meter each on its own unit — and give every rate card the change discipline of a plan page. Invest in a unified forecasting view, or buyers will model the plan fee and miss the meters that dominate the bill. W&B’s tier page is stable and diff-able; its per-model Inference card is neither, and it moved twice in eight days — the 2026-07-21 DeepSeek V4-Flash reset (14x input, 28x output) and the 2026-07-29 five-model reprice (15-45% cuts alongside a 21% output hike on GPT OSS 120B) — showing this isn’t a one-time gap but a recurring cost of an undiffable rate card. Dated entries, a notice window on moves in either direction, and an explicit “introductory” label on rates you intend to normalise keep your cheapest line from becoming a promise you break.
2. Differentiate hot vs cold data in your price. Live, queryable observability streams justify a far higher per-byte rate than cold artifact storage; charging one flat rate leaves money or trust on the table.
3. A free, uncrippled developer entry point compounds. W&B’s free SDK and free-forever academic tier built the distribution that made the company acquirable — proof that the top of the funnel is a pricing decision, not just marketing.

Sources

• Weights & Biases pricing page (accessed 2026-07-21)
• W&B Token-Based Pricing rate card (per-model Inference + ARIA rates) (accessed 2026-07-21)
• W&B Serverless Inference product page (accessed 2026-07-21)
• W&B docs: Serverless Inference usage limits and spending caps (accessed 2026-07-21)
• W&B docs: billing settings, usage categories and alerts (accessed 2026-07-21)
• CoreWeave: completes acquisition of Weights & Biases (accessed 2026-06-16)
• Weights & Biases release notes (accessed 2026-07-21)

Bottom line

Weights & Biases prices like the multi-product platform it became: a free, sticky experiment-tracking SDK at the top, a $60/mo Pro tier for small teams, and custom Enterprise for everyone regulated or over 50 employees — all sitting on three usage meters (storage at $0.03/GB, Weave trace ingestion at $0.10/MB, and per-token Serverless Inference). The free personal and free-forever academic tiers are a deliberate land-grab that built the 1M-user distribution CoreWeave paid a reported ~$1.7 billion for in May 2025. The real friction is forecasting: the plan fee is the floor, and Weave ingestion overage is where unprepared GenAI teams get a four-figure surprise. The published tiers have not moved all year; the per-model Inference card has, twice in eight days — DeepSeek V4-Flash reset from $0.01 to $0.14 in / $0.28 out on 2026-07-21, then five more models moved 15-45% and a new model (MiniMax M3) joined the catalog on 2026-07-29 — so bookmark the rate card, not just the pricing page; it now reprices roughly weekly.

Want to compare Weights & Biases against other AI infrastructure companies? Browse the pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Inference rate card repriced again; MiniMax M3 added

Jul 2026

Five more Serverless Inference models were cut 15-45% per 1M tokens (Z.AI GLM 5.2 ~45% to $0.76 in/$2.42 out; Moonshot Kimi K2.7 Code and K2.6 down 25-32%; Google Gemma 4 31B down, losing its cached-input tier), OpenAI GPT OSS 120B's output rose 21% to $0.17, and a new model, MiniMax M3, joined the catalog (31 models total). Cloud tiers, storage, and Weave ingestion were unchanged.

DeepSeek V4-Flash repriced 14x on Serverless Inference

Jul 2026

The entry-level DeepSeek V4-Flash model went from $0.01 input / $0.01 output per 1M tokens to $0.14 input / $0.07 cached / $0.28 output, adding a cached-token tier and lifting the Inference catalog's cheapest input rate to $0.03 (OpenAI GPT OSS 20B). Cloud tiers (Free / Pro $60 / Enterprise), storage ($0.03/GB) and Weave ingestion ($0.10/MB) were unchanged.

ARIA token-priced product added to the pricing page

Jul 2026

W&B added ARIA — a new token-based-pricing product — as a distinct metered line on the pricing compare table, free 'for a limited time' on Free and Pro and included on Enterprise, alongside a refreshed Serverless Inference model catalog. Core tiers (Free / Pro $60 / Enterprise) and the storage ($0.03/GB) + Weave ingestion ($0.10/MB) meters were unchanged.

Restructured tiers: storage + ingestion + per-token Inference

Jan 2026

Post-acquisition pricing settled on Free / Pro ($60/mo) / Enterprise, metered on storage ($0.03/GB overage), Weave ingestion ($0.10/MB overage), and per-token Serverless Inference, plus CoreWeave Sandboxes credits ($10/$25 per mo).

CoreWeave completes ~$1.7B acquisition

May 2025

CoreWeave closed its acquisition of W&B for a reported ~$1.7 billion (W&B last valued at $1.25B in 2023), folding the developer platform into CoreWeave's GPU cloud while pledging continued interoperability across infra providers and models.

W&B Weave launches for LLM observability

Jan 2024

W&B added Weave — tracing, evaluations, LLM-as-a-judge scorers, and production monitoring for GenAI apps — introducing a new value metric: Weave data ingestion (GB of logged traces), distinct from Models tracked hours.

'Compute hours' meter renamed to 'tracked hours'

Jan 2022

W&B Models pricing shifted terminology from cumulative 'compute hours' to 'tracked hours'; the Starter/Pro tier was sold at roughly $50/user/mo, tiered by cumulative tracked hours (to $100/$150) with $1/hr overage and $0.03/GB storage.

Launched as wandb experiment tracking

Jan 2018

Lukas Biewald and Chris Van Pelt founded W&B; the free experiment-tracking SDK (wandb) became the de-facto standard for logging ML training runs, monetized via team/enterprise seats and a 'compute hours' meter.