GILD verdict
Choose Langfuse when the tracer ships inside a product you sell, and Arize when the smallest self-hosted footprint or classic ML monitoring matters.
- The workflow
- an AI team picking the tracer it will self-host or resell, with agents in production and OpenTelemetry already in place
- The decisive evidence
- ReportedIn February 2026 bbv's AI Hub replaced Phoenix with Langfuse because Phoenix's Elastic License blocked inclusion in SaaS offerings, and Apify replaced its Phoenix eval runner with Langfuse gates in September 2026.456
- MeasuredOne operator measured tuned Langfuse v4 at 4 containers and 1,848 MiB against 1 Phoenix container at about 400 MiB, and made Phoenix the default on Oct 13, 2026. Bayes Impact moved its traces to Phoenix on Sep 29, 2026.12
- DocumentedLangfuse Cloud lists a rate at every volume, which puts 200,000 traces a month at US$369 on Pro in GILD's calculation from published rates.79
- GILD analysisThe AX bill at that volume is unresolved. Arize's pricing page listed no AX Pro span overage rate on Oct 28, 2026, so the 1,950,000 spans past the allowance need a quote before the two bills can be compared.
- The tradeoff
- Langfuse also keeps prompts and evals on one MIT-licensed platform. Langfuse brings ClickHouse, Redis and S3 to run when you self-host. Phoenix is lighter, but Bayes Impact found its bundled Helm setup documented as not production ready, so its database and sign-in become your work.1248112022
- What would change the call
- The call moves if Arize relicenses Phoenix under a license that allows SaaS inclusion, or publishes an AX Pro overage rate that undercuts Langfuse Pro at your span count.
Teams that switched name two reasons. bbv's AI Hub replaced Arize Phoenix with Langfuse in February 2026 because Phoenix's Elastic License kept it out of SaaS offerings, and Apify moved its eval gates to Langfuse in September 2026. Bayes Impact merged a move of its platform's traces to Phoenix on September 29, 2026, and one self-hoster measured Phoenix at one container and about 400 MiB against 1,848 MiB for tuned Langfuse v4. Arize AX adds classic ML monitoring that Langfuse does not cover.
What teams report
Switched 3
bbvOn Langfuse
bbv's AI Hub proposed leaving Phoenix in an issue opened on Nov 18, 2025, because its Elastic License blocked inclusion in SaaS offerings. A 26-commit pull request merged on Feb 12, 2026 moved tracing to OpenTelemetry and Langfuse spans and the dataset API to Langfuse.
“Replace Phoenix dataset API with Langfuse…”
mfundn, Contributor, bbvch-ai/aihub-core, bbvGitHub (bbvch-ai/aihub-core), , post Bayes ImpactOn Arize
Bayes Impact merged a pull request on Sep 29, 2026 that deletes its 673-line Langfuse exporter. Its platform now emits plain OTLP, and an OpenTelemetry Collector sends LLM traces to Phoenix.
“The Langfuse exporter (673 lines), the langfuse* packages and the two @google-cloud/opentelemetry-* exporters are gone.”
Jérémie Doucy, Contributor, bayes-platform, Bayes ImpactGitHub (bayesimpact/bayes-platform), , post Jordi Martinez, Owner, jordimo/infraOn Arize
After removing Langfuse on May 2, 2026, Jordi Martinez made Phoenix the default on every host on Oct 13, 2026 and kept Langfuse v4 as a fallback. Tuned Langfuse v4 needed 4 containers and 1,848 MiB; Phoenix needed 1 container at about 400 MiB.
“Started as "bring Langfuse back" (removed 2026-05-02, 4cfb295) and ended somewhere else. Both stacks are in the compose; Phoenix is the default everywhere, Langfuse is a tested fallback behind its profile.”
Jordi Martinez, Owner, jordimo/infra, jordimo/infraGitHub (jordimo/infra), , post
Evaluated 1
NonBioS
NonBioS looked at Phoenix only briefly, put off by the Phoenix and Arize naming and the website's usability, then chose Langfuse over Helicone (Mar 5, 2025).
“We looked at Phoenix Arize very briefly, but the naming was a bit confusing. Is it Phoenix, or is it Arize? The website also seemed challenging from a usability perspective, and we were left wondering if that would translate to the tool as well.”
Amit, Author, NonBioS blog, NonBioSNonBioS blog, , write-up
Choosing between Langfuse and Arize
Choose Langfuse if
- You resell or embed the stack in a SaaS product: Langfuse's core is MIT, and bbv's AI Hub left Phoenix in February 2026 because its Elastic License blocked that.
- You want to self-host the full platform, prompt management included, free under the MIT license without an Enterprise contract.
- You need data access past 30 days on a self-serve plan: 90 days on Core and 3 years on Pro.
Choose Arize if
- You want the lightest self-hosted tracer: Phoenix ran as one container at about 400 MiB in one operator's October 2026 comparison, against four containers and 1,848 MiB for tuned Langfuse v4.
- The same team also runs classification, regression, ranking, or object detection models and wants drift and performance monitoring beside agent tracing.
- You want a low entry price for a small team: AX Pro is US$50 a month, though Arize's pricing page listed no span overage rate on Oct 28, 2026.
Side by side
| Dimension | Langfuse | Arize |
|---|---|---|
| What teams report | bbv's AI Hub replaced Phoenix with Langfuse for tracing and datasets, merged Feb 12, 2026, to use an MIT license in SaaS offerings. Apify replaced its Phoenix eval runner with Langfuse-based gates, merged Sep 17, 2026.456 | Bayes Impact deleted its Langfuse exporter and sends plain OpenTelemetry to Phoenix, merged Sep 29, 2026. One operator made Phoenix the default and kept Langfuse v4 as a fallback, merged Oct 13, 2026.12 |
| PlansList prices in US dollars. Both include unlimited users on paid plans; Langfuse Hobby allows 2 users. | Hobby free with 50,000 units a month and 2 users; Core US$29 a month; Pro US$199 a month; Enterprise US$2,499 a month; self-hosting free under the MIT license78 | AX Free US$0; AX Pro US$50 a month; AX Enterprise custom17 |
| Billing unitAn agent run with 10 observations and 1 score is 12 Langfuse units and about 10 Arize spans, so compare on your own trace shape. | A unit: each trace, each observation inside it, and each score, including scores from Langfuse's own evaluators9 | Spans and gigabytes ingested18 |
| Included volume | 50,000 units a month on Hobby; 100,000 on Core and above7 | 25,000 spans and 1 GB a month on Free; 50,000 spans and 10 GB a month on Pro17 |
| Usage price | 100,000 units included on Core and above, then US$8 per 100,000 to 1 million, US$7 to 10 million, US$6.50 to 50 million, and US$6 above7 | Pro includes 50,000 spans and 10 GB a month, and Arize's pricing page lists no overage rate (Oct 28, 2026); Enterprise volume is custom1723 |
| 200,000 traces a month, 10 usersGILD's calculation from published rates, Oct 28, 2026. Inputs: 200,000 traces a month, 10 observations and 1 code score per trace, 20 KB per trace, 10 users, 30 days retained, no judge calls. | US$369 a month on Cloud Pro: US$199, plus US$72 for the first 900,000 units past the 100,000 included at US$8 per 100,000, plus US$98 for the next 1.4 million at US$7. Each trace is 12 units, 2.4 million in all.79 | US$50 AX Pro base fee, plus 1,950,000 spans past the 50,000 included at a rate Arize no longer publishes, so the total needs a quote. Self-hosted Phoenix on one 4-vCPU, 8 GiB VM: about US$48 a month in infrastructure, before operations time.1718 |
| Self-hosted footprint | Web and worker containers with PostgreSQL, ClickHouse, Redis and S3. One operator measured tuned v4 at 4 containers and 1,848 MiB resident (Oct 13, 2026).211 | Phoenix runs as one container on SQLite by default or PostgreSQL; the same operator measured about 400 MiB on PostgreSQL. Bayes Impact notes the Phoenix in its Helm chart is documented as not production ready, with no authentication (Sep 29, 2026).1221 |
| Retention | Data access for 30 days on Hobby, 90 days on Core, and 3 years on Pro and Enterprise. Project retention policies from Pro up, minimum 3 days; self-hosted data is kept until you set a policy10 | 15 days on Free, 30 days on Pro, custom on Enterprise17 |
| Self-hosting and license | Free under the MIT license with all core features and unlimited usage; an Enterprise key adds project RBAC, retention policies, audit logs, and server-side masking8 | AX on Kubernetes under Enterprise; Phoenix, a separate open-source product, free to self-host under the Elastic License 2.0, which bbv found blocked inclusion in SaaS offerings41922 |
| Classic ML models | Not covered; Langfuse traces LLM applications and agents | Predictions and actuals for classification, regression, ranking, and object detection, with drift, performance, and data quality monitoring20 |
| Cloud regions | US (Oregon), EU (Ireland), Japan (Tokyo), and a HIPAA region, all on AWS; regions are fully separate12 | US, EU, or Canada |
| Compliance | SOC 2 Type II and ISO 27001 reports from Pro, a BAA in the HIPAA region from Pro, and a GDPR DPA on every plan13 | SOC 2 Type II and GDPR from Pro; HIPAA, enterprise SSO, and audit logs on Enterprise17 |
| Ownership | ClickHouse, which acquired Langfuse in January 2026 and states no planned licensing changes16 | Arize AI, part of Dynatrace since October 1, 2026, which also maintains Phoenix24 |
What practitioners report, side by side
Each product's figures come from its own practitioners, for the group you pick.
Overall signal
Teams and leaders rate Langfuse +46, 20 points above Arize at +26. Most of the gap is in complaints, with 4 of Arize's 27 sources negative against 0 of Langfuse's 26.
The score is how far positive sources lead negative ones, per 100 sources from teams and leaders that take a side. Each author behind these scores wrote one of them, so each source is one vote.
A lead of a fifth of the sources reads positive and a lead of half strong positive, with the same lines below zero. The dots run positive, mixed, negative. Run a pilot on one production app to measure which failures it surfaces and how long tracing them takes.
Footprint on the AI Application Platform stack
They differ on one of 9 layers. On Guardrails, Langfuse touches it and Arize supports it. Coverage is each product's placement on this stack, as its product page shows it. Core is a job the product owns; Supports and Touches are jobs it helps with. Coverage says nothing about how well a product does them.
| Layer | Langfuse | Arize |
|---|---|---|
| Operate | ||
| Guardrails | Touches | Supports |
The 8 layers where they match
| Layer | Langfuse | Arize |
|---|---|---|
| Model | ||
| Model APIs | Touches | Touches |
| Model gateway | Touches | Touches |
| Build | ||
| Orchestration | Touches | Touches |
| Tool use | Touches | Touches |
| Data | ||
| Retrieval | Touches | Touches |
| Evaluate | ||
| Evaluation | Core | Core |
| Operate | ||
| Observability | Core | Core |
| Cloud deployment | Touches | Touches |
What practitioners praise
Among teams and leaders, Free, Simple Self-Hosting and OpenTelemetry & OpenInference reach 5 sources from 3 authors only for Arize, at 6 of 6 and 5 of 6. Langfuse needs a pilot on the same production traffic before you choose on either. Both carry Trace-Level Debugging. It is praised in all 5 Langfuse sources that discuss it, against 10 of 11 for Arize.
| Praised for | Langfuse | Arize |
|---|---|---|
| Trace-Level Debugging | 5 of 5 | 10 of 11 |
| Free, Simple Self-Hosting | Not a theme in its sources | 6 of 6 |
| OpenTelemetry & OpenInference | Not a theme in its sources | 5 of 6 |
Each count is the sources that praise a theme, or raise it as a concern, out of those that discuss it. A row pairs the two products when both carry a theme under the same or an equivalent name.
A count ending in "so far" has under 5 sources from 3 authors among teams and leaders. "Not a theme in its sources" means GILD found no theme of that name in that product's sources.
What practitioners raise as concerns
Among teams and leaders, Upgrades & Migrations reaches 5 sources from 3 authors only for Langfuse, at 5 of 6. Arize needs a pilot on the same production traffic before you choose on it. Scale & Self-Host Operations reaches 5 sources from 3 authors only for Arize, at 6 of 6. Langfuse needs a pilot on the same production traffic before you choose on it.
| Raised as a concern | Langfuse | Arize |
|---|---|---|
| Upgrades & Migrations | 5 of 6 | Not a theme in its sources |
| Scale & Self-Host Operations | Not a theme in its sources | 6 of 6 |
24 sources
- bayesimpact/bayes-platform pull request #877: send traces to Phoenix instead of Langfuse (Sep 29, 2026)
- jordimo/infra pull request #15: standardise on Phoenix, keep Langfuse v4 as fallback (Aug 6, 2026)
- NonBioS: The NonBioS LLM Observability Pick (Mar 5, 2025)
- bbvch-ai/aihub-core pull request #851: replace Arize Phoenix with an open-source alternative (Dec 22, 2025)
- bbvch-ai/aihub-core pull request #907: replace Arize Phoenix with Langfuse (Feb 12, 2026)
- apify/apify-mcp-server pull request #1376: delete the Phoenix eval runner (Sep 17, 2026)
- Langfuse pricing
- Langfuse self-hosted pricing
- Langfuse docs: Billable units
- Langfuse docs: Data retention
- Langfuse docs: Self-host Langfuse
- Langfuse docs: Data regions and availability
- Langfuse security and compliance
- Langfuse docs: Prompt management
- Langfuse docs: OpenTelemetry
- Langfuse blog: Langfuse joins ClickHouse (Jan 16, 2026)
- Arize AX pricing
- Arize AX docs: Pricing and usage
- Arize AX docs: Self-hosted
- Arize docs: ML models (predictions, drift, and performance)
- Arize Phoenix docs: Overview
- Arize Phoenix docs: License
- Arize: Compare Arize
- Dynatrace: Dynatrace completes acquisition of Arize (Oct 1, 2026)