Nesyona Research // Data Study

The LLMOps Stack: OpenTelemetry, Self-Host and Pricing-Model Comparison 2026

Cite this dataset: DOI 10.5281/zenodo.20738671 (CC-BY 4.0)

Which production LLM-ops tools are OpenTelemetry-native, self-hostable, and open-source, and how do they price? A 19-tool, four-layer comparison.

Last updated:
Bottom line: Across 19 production LLM-ops tools placed in a four-layer stack (gateway, observability, evaluation, guardrails), OpenTelemetry support is the dividing line that decides portability and pricing model is the column that decides the bill at scale. The strongest tools leak across layers: Portkey spans gateway, observability, and guardrails; Langfuse spans observability and evaluation; Maxim spans evaluation, observability, and ships its own gateway. The honest open-source default for observability is Langfuse (MIT, self-hostable, OTel-native since v3); for a self-governed gateway, Portkey or LiteLLM.

The comparison matrix

Nineteen tools across the four layers, on the five dimensions that decide interoperability and cost. yes and no are read from each vendor's own documentation; partial denotes tier-limited, proxy-based, ingest-only, or unconfirmed-native. Pricing model is the column to read at scale, not the headline price. The full machine-readable matrix is in data.json.

ToolLayerOTel-nativeSelf-hostOpen sourcePricing modelEval / Guardrails built in
PortkeyGateway +obs +guardrailsyesyesApache-2.0 coreusage (per log)guardrails + observability
LiteLLMGatewayyesyesOSS + Enterprisefree OSS / Enterprise quoterouting only
Cloudflare AI GatewayGatewaypartialno (cloud only)nofree (with Workers)no
Kong AI GatewayGatewayyesyescore OSS, AI plugins paidEnterprise licensepolicy-level
OpenRouterGateway (aggregator)partialno (cloud only)noflat 5.5% on creditsno
AIMLAPIGateway (aggregator)partialno (cloud only)nousage (pay-as-you-go)no
LangfuseObservability +evalyes (v3)yesMITusage (per unit)eval + prompt management
LangSmithObservability +evalpartial (late)partial (enterprise)noseat + per-traceeval
HeliconeObservability +gatewaypartialyesOSSusage (per request)light eval
Arize PhoenixObservability +evalyes (OpenInference)yesOSSfree OSS / usageeval
Traceloop / OpenLLMetryObservabilityyes (reference)yes (library)Apache-2.0free library / platformplatform only
Datadog LLM ObservabilityObservabilityyes (GenAI convention)no (cloud only)nousage (per span)eval
MaximEvaluation +obs +gatewayyes (Bifrost)partial (enterprise)Bifrost Apache-2.0usage / Enterpriseobservability + guardrails
BraintrustEvaluation +obspartial (ingests OTLP)partial (enterprise)nousage / Enterpriseobservability
PromptfooEvaluationno (not the focus)yes (local-first)OSSfree OSS / Enterprisered-team
DeepEval / Confident AIEvaluation +obsyespartial (OSS library only)DeepEval OSSusage (cloud)observability (cloud)
Prediction GuardGuardrails +inferenceyes (events)no (hosted)nousage / Enterpriseeval-style checks
Guardrails AIGuardrailsyes (telemetry)yesApache-2.0free OSS / Provalidation
NeMo GuardrailsGuardrailsnoyesApache-2.0free OSSrails only

Pricing models and OTel status reflect each vendor's public documentation as of June 2026 and change often; verify on the vendor's own page before a purchase decision. The narrative companion to this dataset is The LLMOps Stack 2026.

Methodology

Each cell is read from the vendor's own public documentation, GitHub repository, or pricing page as of June 2026. OpenTelemetry support is marked yes (native) only where confirmed from the tool's own docs; partial denotes tier-limited, proxy-based, ingest-only, or unconfirmed-native. Vendor performance claims (for example latency benchmarks) are not encoded here.

The dataset tracks five dimensions: otel_native, self_host, open_source, pricing_model, and eval_guardrails_built_in, plus each tool's primary layer and the layers it spans. Rankings and the four-layer reference model were fixed before any monetization check; there is no paid placement. Defunct (HumanLoop) and acquired (Protect AI) products are excluded from the live comparison.

Open dataset. The full matrix is published at data.json under a CC-BY 4.0 license, free to share and adapt with attribution to Nesyona / Vincent Couey (ORCID).

Save
Dashboard