The toolsTen AI observability tools, one card each
Order: by how many of five jobs a tool does on a plan you can start without sales (traces of each step, evaluations, prompt management, cost per call, a free self-hosted build); ties go to the cheaper paid plan.
1Opik
Best for a low flat pricePaid from$19/mo
Comet's open-source tracing and evaluation platform. Free: 25,000 spans a month, kept 60 days. Pro: $19 a month flat for 100,000 spans, then $5 per 100,000.
Good
- Lowest paid plan and cheapest extra spans of the ten
- Apache-2.0; OTLP, agent graphs, evaluations and a prompt library
Watch out
- Self-hosted Opik has no user management
- 400-day retention costs $29 per 100,000 spans more
2Langfuse
Best for open-source self-hostingPaid from$29/mo
Open-source tracing, evaluations and prompts, owned by ClickHouse since January 2026. Hobby: free, 50,000 units a month, 30 days. Core: $29 a month for 100,000 units, then $8 per 100,000.
Good
- MIT core: self-host all core features free, SSO included
- 37,651 GitHub stars; OTLP, agent graphs, MCP tracing
Watch out
- Each trace, observation and score is a unit: a 10-step run uses 11+
- Audit logs and retention controls need the paid licence
3Arize (Phoenix and AX)
Best for free self-hosting, no feature gatesPaid from$50/mo
Phoenix is free to self-host; AX is the hosted platform. AX Free: 25,000 spans and 1 GB, kept 15 days. AX Pro: $50 a month for 50,000 spans, kept 30 days. Owned by Dynatrace since 1 October 2026.
Good
- Phoenix self-hosting has “no feature gates”
- OTLP in; cost rolled up to the trace and the project
Watch out
- Phoenix is Elastic License 2.0: source-available, not open source
- No published price for spans past the AX plan
4PostHog AI Observability
Best next to product analyticsPay as you go$0 + usage
LLM tracing inside PostHog's product analytics. The first 100,000 events a month are free; the next 50,000 cost $0.00035 each, then the price steps down to $0.000175, $0.0001 and $0.00008 an event by 1M. No seat charges.
Good
- Cost per call, evaluations, prompts and sessions beside product analytics
- Metadata kept 1 year on the free plan; OTLP over HTTP
Watch out
- Prompt and response text kept 30 days on every plan
- Self-hosting is “officially unsupported”
5Laminar
Best for long retention on a small planPaid from$30/mo
Open-source agent tracing billed by data size. Free: 1 GB a month, 7 days, 1 seat. Pro: $30 a month for 3 GB, 3 months of data and unlimited seats, then $2 a GB.
Good
- Apache-2.0; traces LLM calls, tools and control flow
- Dollar cost on every LLM span; OTLP over gRPC and HTTP
Watch out
- No prompt management
- Billed by the GB, so long prompts cost more
6LangSmith
Best for LangChain and LangGraph teamsPaid from$39/seat
LangChain's tracing and evaluation platform. Developer: free, 1 seat, 5,000 base traces a month. Plus: $39 a seat a month with 10,000 base traces in total, then $0.005 a trace. Base traces are kept 14 days.
Good
- Every agent step is a run; turns link up as threads
- OTLP, LLM-as-judge evals, prompt environments
Watch out
- $500 per 100,000 extra traces, the dearest volume here
- Self-hosting is an Enterprise add-on only
7W&B Weave
Best if you already use Weights & BiasesPaid from$60/mo
Weights & Biases' tracing and evaluation toolkit, now sold as part of CoreWeave Forge. Free: 1 GB of data a month. Pro: from $60 a month with 1.5 GB, then $0.10 a MB.
Good
- Agent timelines with nested tool calls; MCP tracing
- OTLP endpoint; SOC 2 Type II on every plan
Watch out
- $0.10 a MB is $100 a GB, the highest data price here
- No retention stated; self-hosting only on a paid licence
8Datadog LLM Observability
Best if you already run DatadogBilled yearly$160/mo
Datadog's LLM product, now called Agent Observability. Free: 40,000 LLM spans a month, kept 15 days. Pro: from $160 a month billed yearly for 100,000 LLM spans, then $3.50 per 10,000.
Good
- Only LLM spans are billed, not tool or retrieval spans
- Beside your APM, logs and alerts; MCP client tracing
Watch out
- $200 a month month-to-month, $240 on demand
- No self-hosting; 15-day retention unless you pay more
9Braintrust
Best for evaluation-led teamsPaid from$249/mo
An evaluation-first platform with tracing. Starter: free, 1 GB of data, 10,000 scores, 14 days. Pro: $249 a month for 5 GB and 50,000 scores, then $3 a GB.
Good
- Cost rolls up from child spans to the parent
- Unlimited users on every plan; MCP server tracing
Watch out
- The highest entry price of the ten; scores billed on top
- No self-hosting: BYOC on Enterprise keeps a SaaS control plane
10Griffin
Best for cost per use caseWatch plan$0
A TypeScript fetch wrapper from Winding AI that records what each AI call cost, grouped by use case, after the call finishes. It isn't a tracing tool: one record per call, no spans. It also replays real calls on cheaper models with a judge (replay grading).
Price: Watch $0. Test: $10 of replays free, then $20 packs. Route: 5% of routed spend after the first $50.
Good
- Replays real calls on cheaper models, grades each answer, routes only to a model that passed, and falls back to the original model if a routed call fails
- Cost, tokens, latency and errors per call, by use case, recorded outside the request path
Watch out
- One of the five jobs: no traces, eval suite or prompt management
- Hosted only (MIT SDK); the wrapper is TypeScript