Software advice
Best LLM Observability Tools in 2026
An LLM observability tool is software that records every model call and agent step in an application as a trace, with inputs, outputs, tokens, latency and cost, so engineers can debug failures and measure quality. Many also run evaluations and manage prompts. It is not a general application performance monitor and it does not host or train models.
Quick answer
Langfuse ranks first for open-source code you can self-host or run in an EU region. Arize Phoenix suits a free local start, and Opik has the lowest paid entry at $19 monthly. Braintrust fits teams whose main need is evaluation.
A team shipping an LLM feature soon needs to answer three questions: what did the model receive, what did it return, and did the change we just made make things better or worse. LLM observability tools answer them by storing traces and scoring outputs. This ranking compares twelve vendors on trace quality, price, data handling, how hard it is to leave, and which team each suits. Prices come from vendor pricing pages read in October 2026 and change often. Each vendor's own pricing, security and terms pages were read directly. We did not run benchmark tests, so quality is described through documented features and design trade-offs, not scored.
The 12 best LLM observability tools in 2026, ranked
There is no score. Vendors can pay for more visibility on this website, for example a higher place in a ranking or more room on a page. Commercial relationships can also affect which products we feature and how prominently. How we rank.
| # | Tool | Best for | Free plan | Entry price |
|---|---|---|---|---|
| 1 | Langfuse | Teams that want open-source tracing they can self-host or run in an EU cloud region. | Yes | $29 per month (Core) |
| 2 | Arize Phoenix | Engineers who want a local, open-source trace viewer first and a managed upgrade path later. | Yes | $50 per month (Arize AX Pro) |
| 3 | Opik (Comet) | Small teams that want the lowest paid entry price and an open-source option to fall back on. | Yes | $19 per month (Pro Cloud) |
| 4 | Braintrust | Teams that treat evaluation as a release gate and want datasets, scorers and traces together. | Yes | $249 per month (Pro) |
| 5 | Helicone | Teams that want request logging plus a gateway with caching, rate limits and fallbacks. | Yes | $79 per month (Pro) |
| 6 | LangSmith | Teams already building agents with LangChain or LangGraph who want tracing from the same vendor. | Yes | $39 per seat per month (Plus) |
| 7 | Portkey | Teams that want one gateway for many model vendors with logs, guardrails and budgets attached. | Yes | $49 per month (Production) |
| 8 | HoneyHive | Teams running agents in production who want online evaluations and annotation queues. | Yes | Quoted |
| 9 | Lunary | Small teams that want chat replays, topic classification and human review at a flat per-user price. | No | $20 per user per month (Team) |
| 10 | Traceloop | Teams that already use OpenTelemetry and want LLM spans sent to a backend of their choice. | Yes | Quoted |
| 11 | Datadog LLM Observability | Companies that already run Datadog for infrastructure and want AI traces on the same screens. | Trial | Quoted |
| 12 | Confident AI | Teams that already use DeepEval and want shared evaluation runs, datasets and production traces. | Yes | $200 per month (Starter) |
- 1
Langfuse takes first place because the same open-source code runs as a hosted service and on your own servers, and the cloud offers US, EU or Japan data regions. Tracing, prompt management and evaluation sit in one product, and pricing is counted in units with all users unlimited above Hobby. The acquisition by ClickHouse is the open question, since roadmap priorities may shift.
Strong on
- Self-hosting guides exist for Docker Compose, Kubernetes and Terraform on AWS, GCP and Azure
- Core and Pro plans have unlimited users, so seat count does not drive the bill
- SOC 2 Type II and ISO 27001 reports are listed for the Pro plan
Weak on
- Hobby keeps only 30 days of data and allows two users
- Extra usage is billed at $8 per 100k units, which grows with chatty agents
Skip it if: Teams that want a single vendor to also supply the gateway, caching and key management in front of the model. Hobby is free with 50k units a month. Core adds 100k units and 90 days of data access.
- 2
Arize Phoenix
Arize Phoenix ranks second because it starts with one command on a laptop and needs no account, then connects to the managed Arize AX platform when a team wants hosted storage. The managed side has a free tier with 25k spans a month. The split between two products with different limits is confusing, and AX retention on the free tier is only 15 days.
Strong on
- Phoenix runs locally with a single uvx command and can be self-hosted
- AX Free and AX Pro both list unlimited users and unlimited evals
- Self-hosted deployment is offered on the AX Enterprise tier
Weak on
- AX Free keeps traces for 15 days and AX Pro for 30 days
- Two products, Phoenix and AX, mean two sets of docs and limits to learn
Skip it if: Buyers who want one flat product without separate open-source and managed editions to evaluate. Phoenix is open source and free to run. The managed Arize AX Pro tier starts at $50 monthly.
- 3
Opik sits third because its Pro Cloud plan is the cheapest paid tier in this ranking, and the open-source version carries the same feature set as the hosted ones according to Comet. It adds test suites and an agent playground next to tracing. Comet is better known for experiment tracking, so the LLM side has a younger track record.
Strong on
- Pro Cloud costs $19 monthly with up to 50 team members and 100k spans
- Free Cloud allows up to 10 team members and 60-day retention
- Enterprise tier lists SOC 2, ISO 27001, ISO 9001, HIPAA and GDPR compliance
Weak on
- Free and Pro Cloud both cap retention at 60 days unless customized
- Single sign-on and service accounts require the Enterprise tier
Skip it if: Companies that need single sign-on or audit controls on a self-service plan without talking to sales. Free Cloud covers 25k spans a month. Pro Cloud raises that to 100k spans and 50 members.
- 4
Braintrust takes fourth place because evaluation is its center: scorers, experiments, datasets and playgrounds are included with unlimited users even on the free Starter plan. Billing is by processed data and scored outputs, which suits eval-heavy work. The jump to $249 for Pro is steep for small teams, and the free tier keeps data for only 14 days.
Strong on
- Starter includes unlimited users, projects, datasets, playgrounds and experiments
- Deployment options include a data plane run in your own cloud or self-hosted
- SOC 2 Type II, HIPAA and GDPR are listed in the security documentation
Weak on
- Starter retention is 14 days and Pro is 30 days, with longer storage billed per gigabyte
- Pricing mixes platform fee, data volume and score counts, so forecasting takes a calculator
Skip it if: Solo developers who only need basic request logging and would pay for evaluation features they never use. Starter is free with 1 GB of processed data. Pro includes 5 GB, 50k scores and 30-day retention.
- 5
Helicone
Helicone ranks fifth because it combines logging with a gateway, so caching, rate limits and automatic fallbacks come in the same product as the dashboards. Integration is meant to be one line. The pricing page shows a notice that Helicone is no longer part of Mintlify, with details to come, so ownership and roadmap deserve a direct question before you commit.
Strong on
- Gateway features include caching, rate limits and automatic fallbacks
- Pro and Team plans list unlimited seats, and Team adds SOC 2 and HIPAA compliance
- Self-hosting is documented for AWS, GCP and Azure
Weak on
- Hobby keeps data only 7 days and ingests 10 logs a minute
- Pro shows usage-based charges on top of the $79, so the monthly bill varies
Skip it if: Buyers who need a vendor with a settled ownership story and a recent privacy policy. Hobby is free with 10,000 requests and 1 GB storage. Pro adds unlimited seats; usage-based charges apply.
- 6
LangSmith lands sixth: tracing and evaluation come from the maker of LangChain and LangGraph, so framework integration is tight, and cloud data can sit in the US or EU. Per-seat billing plus per-trace charges makes cost grow with both headcount and traffic. The free plan covers a single seat and keeps base traces only 14 days.
Strong on
- Cloud data can be hosted in LangChain's cloud in the US or EU
- Enterprise offers hybrid and fully self-hosted deployment in your own VPC
- The terms state LangChain will not use customer data to train its products
Weak on
- Plus bills $39 per seat and then each trace beyond 10k base traces
- Self-hosting and hybrid deployment are limited to the Enterprise plan
Skip it if: Small teams on other frameworks who would pay per seat for features tied to LangChain. Developer is free for one seat with 5k base traces. Plus charges per seat plus pay-as-you-go traces.
- 7
Portkey comes seventh because it is a gateway first and an observability tool second: a universal API, fallbacks, load balancing and guardrails sit in front of the model calls, and logs are a byproduct. That suits multi-vendor setups where one team owns routing for many apps. Pure tracing and evaluation depth is thinner than at the specialist tools above it.
Strong on
- Every plan includes a universal API, key management, routing and observability
- Production plan adds semantic caching, guardrails and role-based access control
- Enterprise lists private cloud, VPC hosting, SOC 2 Type 2, GDPR and HIPAA
Weak on
- Free Developer plan keeps logs for 3 days and is described as not for production
- Data residency guarantees and custom security controls require Enterprise
Skip it if: Teams that already have a gateway and only want tracing and evaluation of agent runs. Developer is free with 10k recorded logs. Production includes 100k logs, with $9 per extra 100k requests.
- 8
HoneyHive
HoneyHive sits eighth: the free Developer plan is generous on features, with tracing, online and offline evaluation, datasets and prompt versioning for up to five users, but capped at 10,000 events and 30 days of retention. There is no published paid tier between free and Enterprise, so any growth beyond the free limits means a sales conversation first.
Strong on
- Free plan includes automated and human evaluations, CI/CD integration and prompt deployments
- Annotation queues turn expert review into reusable evaluation data
- Enterprise offers self-hosted, hybrid or single-tenant deployment with SAML SSO
Weak on
- No self-service paid plan is listed, so leaving the free tier means contacting sales
- Free plan allows only 10,000 events a month and community support
Skip it if: Buyers who want to see a published price for 50k or 500k events before talking to anyone. Developer is free with 10,000 events a month and 5 users. Enterprise pricing is custom.
- 9
Lunary
Lunary ranks ninth. Its Team plan offers one year of history, chat replays, topic classification and full API access at $20 per user, and it can be self-hosted on Enterprise. Payment is required before a new workspace opens, so there is no free trial of the product for new accounts. Seats top out at ten before Enterprise.
Strong on
- Team plan keeps one year of history and unlimited projects
- Full API access and custom dashboards are included at the entry tier
- Enterprise lists self-hosting, PII masking and SOC 2 and ISO 27001 reports
Weak on
- New workspaces require payment first, since the free plan applies only to existing users
- Team is capped at 10 seats and events above 50k cost $10 per 50k
Skip it if: Developers who want to try the product for free before paying or run a large team on one plan. Team includes 50,000 events, 300 playground queries and 1 year of history. Payment is required before access.
- 10
Traceloop
Traceloop closes the list because of ownership uncertainty and thin retention. Its OpenLLMetry library is Apache-2.0 licensed and sends spans to more than 25 observability platforms, which keeps tracing portable. The pricing page now says Traceloop is joining ServiceNow, and the free tier keeps data for 24 hours, so check the product's future before building on the hosted service.
Strong on
- OpenLLMetry is Apache-2.0 licensed and connects to 25+ observability platforms
- Free plan lists up to 5 seats and 50K spans a month
- On-premise deployment, including air-gapped setups, is stated as available
Weak on
- Free plan retains data for only 24 hours
- The announcement that Traceloop is joining ServiceNow leaves the standalone roadmap unclear
Skip it if: Buyers who need a long-term standalone vendor with a settled roadmap and published paid tiers. Free forever covers 50K spans a month, 5 seats and 24-hour retention. Enterprise pricing is custom.
- 11
Datadog LLM Observability
Datadog ranks eleventh because its LLM tracing is an add-on to a much larger monitoring suite, which suits existing customers more than new buyers. Documentation describes traces made of spans with inputs, outputs, latency and errors, plus topic clustering of production traffic and scans that redact sensitive data. Billing follows ingested LLM spans, and an agent workflow creates several. Teams without Datadog adopt a whole platform to get it.
Strong on
- Docs describe traces of LLM calls and agent steps, with dashboards for cost and latency
- Automatic scanning redacts sensitive data and detects prompt injection attempts
- Python SDK integrates with OpenAI, LangChain, AWS Bedrock and Anthropic
Weak on
- Per-span billing means one agent workflow can create several billable spans
- Product is not supported on two US government Datadog sites
Skip it if: Teams with no other Datadog use who want a standalone prompt and evaluation tool with a fixed price. Billing counts LLM spans ingested, and one agent workflow can produce several spans. No dollar amount was visible.
- 12
Confident AI
Confident AI closes the list because its center of gravity is evaluation, with tracing as the second half. The docs describe it as the hosted home for DeepEval, with 50+ metrics, regression tests in CI and online evaluations on production traffic. The free plan allows 2 seats and 1 GB-month of trace spans, and the first paid step is $200 a month. Price jumps are steep, and the privacy page returned an error.
Strong on
- Free plan is available at no cost with 2 seats, 1 project and 1 GB-month of spans
- Paid plans list unlimited retention, billed by GB-month ingested or retained
- Self-hosting through Docker in AWS, Azure or GCP is described on the docs page
Weak on
- Price steps from $200 per month on Starter to $2,000 per month on Team
- Privacy and trust pages returned errors, so data terms were not checked
Skip it if: Developers who only need cheap request logging and cost tracking without evaluation datasets or metrics. Starter includes 5 projects and 5 GB-months of traces. Extra data costs $1 per GB-month ingested or retained.
How these LLM observability tools were judged
What should a good LLM trace show beyond the prompt and response?
A useful trace shows the full chain: retrieval steps, tool calls, nested agent steps, token counts, latency and cost per span. Langfuse, Arize Phoenix and Opik document agent graphs and session views for multi-turn chats, and Braintrust adds scorers that run on stored traces. Gateway tools such as Helicone and Portkey see requests only as they pass through, so work done inside your own code can be invisible. Check whether the vendor accepts OpenTelemetry, because that decides how much instrumentation you rewrite later. Also test how the interface behaves with long, multi-step traces from your own application, since that is where viewers differ most.
How do LLM observability tools charge, and what does a small team pay?
Pricing units differ, which makes direct comparison hard. Langfuse counts units, Arize AX and Opik and Traceloop count spans, Helicone counts requests, Braintrust bills processed data and scores, and LangSmith bills per seat plus traces. As of October 2026 paid entry points range from $19 monthly for Opik Pro Cloud to $249 for Braintrust Pro. HoneyHive and Traceloop publish no paid tier. Free plans exist at most vendors but keep data for between 24 hours and 60 days. Estimate your monthly trace count first, then price each vendor on that number, not on the headline.
Where do my prompts and customer data go when I use a tracing service?
Traces often contain customer text, so storage location and retention matter more than in ordinary monitoring. Langfuse offers US, EU or Japan regions on its cloud, and LangSmith hosts in the US or EU. Braintrust, Lunary, Traceloop, HoneyHive and Arize list self-hosted or private deployments on their top tiers. Compliance claims vary by tier: SOC 2 reports are tied to Pro or Enterprise plans at several vendors. LangChain's terms state it will not train on customer data. For the others, read the data processing terms before sending production traffic.
How hard is it to switch LLM observability tools later?
Switching cost sits in two places: the instrumentation in your code and the history stored at the vendor. OpenTelemetry-based collection lowers the first, and Traceloop's OpenLLMetry is built for that purpose. Open-source servers such as Langfuse, Phoenix and Opik lower the second, since you can keep running your own copy. Gateway tools add a different dependency, because your requests route through them. Ask each vendor for a bulk export of traces, datasets and prompts, and check retention limits, since a 14-day window means history disappears before a migration can happen.
Which LLM observability tool fits my team size and stack?
Match the tool to the workflow. A solo developer testing an agent fits Arize Phoenix locally or a free Langfuse Hobby account. A team on LangChain or LangGraph gets the closest integration from LangSmith. Teams that release only after passing evaluation suites fit Braintrust or HoneyHive. Companies calling several model vendors may prefer Portkey or Helicone for the gateway features. Regulated buyers should shortlist vendors that publish self-hosting guides and regional hosting. Run a two-week trial on real traffic from one service before moving everything.
How to choose a LLM observability tool in three steps
- Count spans or requests before comparing prices. List your monthly model calls, then multiply by the steps each call produces, because a single agent run can emit dozens of spans. Vendors count different things: units, spans, requests or processed gigabytes. Convert your traffic into each vendor's unit and read the plan that covers it, including the retention that comes with that plan. Then add seats if the vendor bills per user. This one spreadsheet removes most of the confusion between a $19 plan and a $249 plan, because the cheaper plan can cost more once traffic is added.
- Decide where trace data may live. Ask legal or security which fields may leave your network. If customer text must stay inside your infrastructure, shortlist tools with a published self-hosting guide, such as Langfuse, Arize Phoenix, Opik or Helicone. If an EU region is enough, check which vendors offer one on a self-service plan. Request the vendor's data processing agreement and its retention terms before trial traffic begins, and keep personal data out of traces with masking where the tool offers it. Write down the retention period you need, too, since free tiers delete older traces after as little as one day.
- Run one real service through two finalists. Pick one production-like service and instrument it with two tools for two weeks, using OpenTelemetry where supported so the code change stays small. Compare how fast you find a failing run, whether evaluation scores match your own judgment, and how the interface handles your longest traces. Export the data from each at the end. The tool that survives this test and exports cleanly is the one to pay for. If both fail on export, treat that as a warning about how hard leaving would be.
Frequently asked questions
What is the best LLM observability tool in 2026?
Langfuse ranks first here because its open-source code runs both as a hosted service and on your own servers, with US, EU or Japan regions in the cloud. Arize Phoenix ranks second for a free local start and Opik third for the lowest paid entry. The best choice still depends on trace volume and data rules, so test two finalists.
Is there a free LLM observability tool?
Yes. Langfuse Hobby offers 50k units monthly, Opik's free cloud tier covers 25k spans, Traceloop's free plan 50K spans, Helicone 10,000 requests and Portkey 10k logs. Arize Phoenix is open source and free to self-host. Free tiers keep data from 24 hours to 60 days, so they suit testing more than long-term records.
What is the difference between LLM observability and evaluation?
Observability records what happened in each call and agent step. Evaluation scores whether the output was good, using datasets, human review or an LLM acting as judge. Braintrust and HoneyHive emphasize evaluation, while Helicone and Portkey emphasize logging at the gateway. Most tools in this ranking now do both to some degree, with different depth.
Can I self-host an LLM observability tool?
Yes, with several. Langfuse publishes guides for Docker Compose, Kubernetes and Terraform. Arize Phoenix runs locally with one command, Opik offers an open-source version, and Helicone documents self-hosting on AWS, GCP and Azure. LangSmith limits self-hosting to Enterprise, and Lunary and Traceloop also list it under custom plans.
Do LLM observability vendors train on my prompts?
Policies differ and should be read per vendor. LangChain's terms say it will not use customer data to train on or improve its products, including large language models. For the other vendors in this ranking, the policy was not checked for this question, so read the data processing terms and the retention settings before sending production traffic.
Do I need a gateway like Helicone or Portkey for observability?
No. A gateway logs requests as they pass through and adds caching, fallbacks and rate limits, but it does not see work inside your own code. Tracing libraries and SDKs record those internal steps. Many teams use both: a gateway for routing and an SDK-based tool such as Langfuse for the full agent trace.
Which LLM observability tool is cheapest for a small team?
As of October 2026, Opik Pro Cloud at $19 per month is the lowest listed paid plan, followed by Langfuse Core at $29, LangSmith Plus at $39 per seat and Portkey Production at $49. Several free tiers cost nothing. Compare using your own monthly span count, because overage charges differ.
Does OpenTelemetry work with LLM observability tools?
Yes, increasingly. Langfuse lists OpenTelemetry support for Java, Go and custom languages, Opik documents native OpenTelemetry support, and Traceloop's OpenLLMetry extends OpenTelemetry to model calls. Using it keeps instrumentation portable, so changing the backend later means changing an endpoint instead of rewriting code throughout the application.