Eval platforms vs eval optimization
An eval platform runs, stores and reviews evals. Slancha pilots aim to make the expensive part of that run cheaper and faster. Pilots by arrangement. Arize, Galileo, LangSmith, Braintrust and Promptfoo are platforms. Slancha pilots aim to work alongside the eval stack you already run. It is not a replacement.
Published prices, retrieved 16 October 2026, not Slancha results. The judge-model bill is separate from the platform fee on every row, including ours. Nothing here is a ranking.
What each one is for
Arize Phoenix
Self-hosted tracing, evaluation and experimentation. Arize also sells Arize AX as the hosted product.
Published price. The self-hosting docs say Phoenix is free to self-host, with no license fee and no feature gate. The license page identifies Elastic License 2.0.
You still pay. The judge model, and the machine you run Phoenix on.
Limit. Read the license before you offer Phoenix itself as a service to someone else. Hosted production monitoring is AX, priced separately.
Arize AX
Hosted tracing and evaluation. The pricing page lists online and offline evals as unlimited on the published tiers.
Published price. Free is $0 with 25,000 spans a month, 1 GB and 15-day retention. Pro is $50 a month with 50,000 spans, 10 GB and 30-day retention. Enterprise is custom.
You still pay. Unlimited platform evals are not free model calls. You still pay the judge.
Limit. Span and storage allowances are the published cap. Confirm the tier before you buy.
Galileo
Evaluation and observability for agents, including runtime guardrails. On 16 October 2026, galileo.ai/pricing returned Splunk’s pricing page.
Published price. That page lists Agent Observability starting at $100 per month per 1.2 million spans a year, billed annually. Treat the line as the published list price and confirm it.
You still pay. Span volume, and any judge model the product does not include.
Limit. The product line has moved onto the page that URL returns. Do not budget from an older write-up.
LangSmith
Tracing and evaluation, a natural fit when the app already uses LangChain. Developer is one free seat. Plus adds seats and a larger included trace allowance.
Published price. Developer is $0 per seat a month, one seat, up to 5,000 base traces a month, then pay as you go. Plus is $39 per seat a month, up to 10,000 base traces a month, then pay as you go.
You still pay. Traces past the allowance, and the judge model if you bring one.
Limit. Their docs describe plan transitions. The pricing page is the list price to confirm.
Braintrust
Experiments, evals and production traces, with review in the product. Starter includes the whole team.
Published price. Starter is $0 a month: $10 of credits, 1 GB, 10,000 judge results then $2.50 per 1,000, and 14-day retention. Pro is $249 a month: $100 of credits, 5 GB, 50,000 judge results then $1.50 per 1,000, and 30-day retention. Enterprise is custom.
You still pay. Judge-result overage, processed data, and model calls made with your own key.
Limit. The pricing page uses its own label for a judge result. The allowances above are the figures on that page.
Promptfoo
Evaluation in the repo and in CI. Community includes the evaluation features and runs locally or on your own infrastructure.
Published price. Community is free. Enterprise and on-premise are custom. Community red-team probes are capped at 10,000 a month.
You still pay. The judge model. A free runner does not make the model call free.
Limit. Team sharing, hosted monitoring and SSO are on the paid plans.
Slancha
Slancha pilots aim to work alongside the eval stack you already run. It is not a replacement. Slancha pilots aim for cheaper, faster evals, not a second place to store them.
Published price. Slancha: pilots by arrangement. There is no public install command for it.
You still pay. Your current platform, and the model calls the run still makes.
Limit. This site does not show a token or dollar saving as a result.
Use the platform, then make the run affordable
Keep the system that already holds your traces, datasets and experiments. If a model upgrade or a consistency check is too expensive to run, the missing piece is the cost of the judge calls, not another dashboard. The guide cites the public numbers. The calculator turns your set size, runs, judge price and tokens into a cost, with the assumptions labeled on the page. If the run is too slow rather than too expensive, the eval speed guide covers that.
Slancha: pilots by arrangement. There is no public install command for it.
Request a pilot