Observability Platform Pricing: Examples & Companies

What is it

Observability Platform Pricing is pricing for LLM and ML observability platforms — tracing, evaluation, and monitoring of model behavior in production. The product ingests the telemetry an AI application throws off — traces, spans, events, evaluation scores, logged payloads — and the pricing usually follows the pipe: a free ingestion quota, metered volume above it, and a retention window that decides how long the data stays queryable.

The LLM-native core of the category converged on a remarkably uniform shape — meter ingestion, give away seats, sell retention — and the nine telemetry vendors (Langfuse, Arize AI, Galileo, Comet’s Opik, Braintrust, HoneyHive, Athina AI, Patronus AI, PromptLayer) each vary the recipe without leaving it. What differs from vendor to vendor is which unit gets counted — a span, a whole trace, a composite unit, an ingested byte — and that choice quietly decides the bill.

Two adjacent flavors stretch the category at its edges. Weights & Biases folds LLM observability (its Weave product) into a broader MLOps platform that also bills seats, model storage, and inference tokens. And the FinOps pair — Finout and Vantage — are cost observability: they watch cloud and AI spend and price by the size of the bill being monitored rather than by telemetry ingested. Together the twelve show that “observability pricing” spans a spectrum from pure per-span metering to flat subscriptions keyed to the problem being observed.

The whole category sits under two structural pressures. Open source is everywhere — Arize’s Phoenix, Comet’s Opik, Langfuse’s self-host path, and W&B’s free wandb SDK — so managed pricing is disciplined by the self-hosting alternative. And the category is consolidating fast: ClickHouse acquired Langfuse in January 2026, Cisco closed on Galileo in May 2026, and CoreWeave acquired Weights & Biases for a reported ~$1.7B — which makes pricing (and data-export) stability itself a buying criterion.

How it works

For the LLM/ML-telemetry vendors the standard bill is platform fee + (ingestion − quota) × rate, with retention and seats as packaging levers. The FinOps outliers replace the ingestion meter with a flat tier keyed to monitored spend:

Worked example — dual-axis fan-out. A RAG app sending 1M spans and 25 GB a month to Arize AI’s Pro plan pays the base fee plus overage on two axes — trace-spans above the included allowance and ingested GB above 10 — with the data axis, not the raw span count, typically doing most of the damage on payload-heavy agent traces. The same behavior on Braintrust hits its processed-data meter, which counts every byte moving through the platform under a $249 Pro fee, while on Weights & Biases the same trace stream runs through Weave at $0.10/MB — roughly $100/GB, about 3,000x the $0.03/GB it charges for cold model storage. Instrument first, then price: the usage-event tracking guide covers how to measure your own fan-out before committing to a tier.

Companies using this

Twelve in-corpus companies carry the observability product segment: the LLM/ML-telemetry group — Arize AI, Athina AI, Braintrust, Comet (Opik), Galileo, HoneyHive, Langfuse, Patronus AI, and PromptLayer — plus Weights & Biases on the MLOps side and the FinOps cost-observability pair, Finout and Vantage. Ten of the twelve ship a free tier (Finout is the exception, quote-only across every tier), and all twelve gate their top tier behind sales.

Patterns observed

The telemetry core prices ingestion, gives away seats, and charges for memory. A per-user tax would fight the meter, so seats are free or unlimited almost everywhere. The money moves to two places instead: metered overage above the free ingestion quota, and retention as a quiet upsell — Patronus AI gates its entire free tier on a rolling 2-week data window rather than on features, while Braintrust and Arize both roughly double their retention window exactly where the paid tier begins. Keeping data queryable, not receiving it, is where the second dollar comes from.

Defensive unit design is the second pattern. Every vendor that survived repricing simplified toward a unit the buyer can audit. Comet prices Opik spans flat per account; Athina AI deliberately meters executions while ingested logs cost zero; PromptLayer folds requests, agent runs, and eval cells into a single transaction, trading precision for forecastability — the legibility-versus-cost-tracking trade-off the choosing the right usage metric guide frames directly. Even where meters multiply, they stay named and published: Braintrust’s three lines and Weights & Biases’s four are each individually visible on the pricing page rather than rolled into an opaque total.

Free tier as an acquisition funnel, backed by open source. Because telemetry is only valuable at production volume, the free quota is the top of the funnel, and behind most tiers sits an open-source escape hatch (Phoenix, Opik, Langfuse self-host, W&B’s wandb SDK) that both seeds adoption and disciplines the managed price. The FinOps outliers reach the same “adopt free, grow into paid” outcome from the other direction: Vantage pairs a genuinely free tier with high-traffic free tools like EC2Instances.info to seed developers before any paid plan.

Counterexamples & variants

Multi-meter cloud bills, not SaaS tiers. Braintrust breaks the “one ingestion meter” mold with three independent meters closer to a cloud invoice than a SaaS plan, with the processed-data line as the recurring surprise. Weights & Biases goes further: four simultaneous meters and a deliberately asymmetric price that runs its hot, queryable Weave trace stream at roughly 3,000x its cold model-storage rate, pricing observability far above artifacts. W&B also codes its self-serve-to-sales handoff into the pricing page with a hard 50-employee cap on Pro — a rare structural tripwire instead of a usage threshold.

Metering work done, not data received. Athina AI inverts the category’s core assumption: telemetry in is free, and only compute the platform itself initiates — prompt runs, eval cells, flow steps — burns credits. Patronus AI breaks it the other way, with no ingestion meter at all on its self-serve tier, just a hard 2-week data window, the real money living in a quoted enterprise plan plus an optional pay-as-you-go evaluation API. Both prove “observability pricing” can meter evaluation and computation rather than raw data volume.

FinOps observability, priced by the bill it watches. The sharpest variant is the cost-observability pair, where neither vendor meters telemetry at all — the “observed” object is the customer’s own spend. Vantage gates plans by monitored cloud spend at fixed monthly rates it advertises as pricing that “doesn’t contribute to your cost problems” — usage-shaped without being a percentage of spend. Finout prices for the certainty its finance buyers want: a flat annual fee tiered by committed cloud spend (reported at roughly 1% of the bill, ~$1,000/mo up to ~$500k spend), unlimited users, and “no surprise overage charges.” The price scales with the size of the problem, not with events ingested.

What this means for buyers vs vendors

For buyers

Run a week of production-shaped traffic through two or three free tiers and read the actual meters before choosing — the same workload registers as wildly different volumes depending on whether the unit is a trace (Galileo), a span (Arize, Comet), a composite unit (Langfuse), or a byte, where a per-MB line can produce four-figure surprises on payload-heavy traces. Price the retention you actually need for audits rather than the tier default, and separate your two questions: if you are buying application observability, the LLM-telemetry cohort is the field; if you are buying cost observability, Finout and Vantage price by your cloud bill, not your trace volume, and the comparison is a different one. In a consolidating category, ask what happens to your rates and your data-export path if the vendor is acquired — three of these twelve were, inside eighteen months.

For vendors

The telemetry cohort’s settled playbook is clear: meter ingestion, free the seats, ladder retention, and keep an open-source pressure valve. The live design question is bill-shock control — graduated curves (Langfuse) and flat team accounts (Comet’s Opik) both beat raw linear overage on buyer trust, and if your true cost driver is bytes rather than events, surface it as its own published axis the way Arize does rather than letting a hidden data meter aggregate into an invoice the buyer cannot reconstruct — the lesson Braintrust’s processed-data line and W&B’s Weave meter both teach. The FinOps variant points to a second strategy entirely: when your buyer is finance, price for certainty the way Finout does — a flat committed-spend fee with no overage — even if it means leaving usage upside on the table, because a predictable invoice is itself the product for that buyer.

Explore this theme in the knowledge graph →