Best LLM Observability Tools for Indie Developers (2026)
Most indie developers shipping an LLM feature can trace it for $0. Langfuse Cloud Hobby gives 50k units a month free and the core is MIT-licensed to self-host, Comet Opik gives 25k spans free with a $19/mo Pro plan, and Helicone's free tier logs 10,000 requests through a one-line proxy change. Here is what each tool costs, what it counts, and which one fits your stage.
Key takeaways
- Langfuse is our default pick: Cloud Hobby is free for 50k units a month with 30 days of data access, Core is $29/mo for 100k units, and the repo is MIT-licensed except its ee folders.
- Comet Opik has the cheapest paid step we found: 25k spans free, then Pro at $19/mo for 100k spans and $5 per extra 100k, with an Apache-2.0 open source version.
- Helicone is the fastest to add because it works as an OpenAI-compatible gateway: change the base URL and you log 10,000 requests a month free, with 7-day retention.
- Every vendor counts usage differently (units, traces, spans, GB, requests), so estimate your volume in their unit before comparing prices.
- Log prompt, response, model, tokens, cost, latency, user id and prompt version from day one, and redact PII before it leaves your server.
On this page 25
- TL;DR
- What LLM observability actually gives you
- The comparison table
- The tools, one by one
- Langfuse: the default for most indie developers
- Comet Opik: the cheapest paid tier
- Helicone: the five-minute option
- LangSmith: if you live in LangChain
- Arize Phoenix and Arize AX: local-first
- Laminar: agent traces, billed by GB
- Braintrust and W&B Weave: eval-first platforms
- Traceloop and OpenLLMetry: instrument once
- Sentry, Datadog and PostHog: what you may already have
- What to log from day one
- Privacy: redact before you send
- Worked example: the same app on three tools
- Decision guide by stage
- Frequently asked questions
- What is the best free LLM observability tool for indie developers?
- What are the best Langfuse alternatives?
- Do I need an LLM observability tool if I already use Sentry or Datadog?
- Can I self-host LLM observability for free?
- How do I estimate what LLM tracing will cost?
- Ship it and show it
- Sources
Verdict: Start with Langfuse. Its free Cloud Hobby plan covers 50k units a month, Core is $29/mo when you outgrow it, and you can self-host the MIT-licensed core if your data cannot leave your servers. Pick Comet Opik if you want the cheapest paid tier, or Helicone if you want logging in five minutes by swapping a base URL.
Every price and limit below comes from each vendor's own pricing page, docs or GitHub repo, checked on 13 Nov 2026. The advice in between is ours and is marked as such.
TL;DR#
- Langfuse: free for 50k units a month, Core $29/mo for 100k units plus $8 per extra 100k, MIT core you can self-host (Langfuse pricing).
- Comet Opik: 25k spans free with 60-day retention, Pro $19/mo for 100k spans plus $5 per 100k, Apache-2.0 (Comet pricing).
- Helicone: a gateway you point your OpenAI SDK at; 10,000 requests a month free, Pro $79/mo (Helicone pricing).
- LangSmith: 5,000 base traces a month free for one seat, Plus $39 per seat per month (LangChain pricing).
- Already on Sentry or Datadog? Check their AI monitoring before adding another tool.
Disclosure: ZeroRaised publishes this guide and is a launch site for bootstrapped products. ZeroRaised has no commercial relationship with any vendor listed here and earns nothing if you sign up. Facts come from the linked sources as checked on 13 Nov 2026.
What LLM observability actually gives you#
A user says "it gave me a weird answer" and you have no idea which prompt, model or retrieved chunk caused it. LLM observability tools record each request as a trace (prompt, response, model, tokens, cost, latency and nested tool or retrieval steps), and most add prompt versioning, datasets and evals. We cover the why in what is AI agent monitoring and scoring in how to evaluate LLM outputs; this post answers which tool, at what price.
The comparison table#
Units differ by vendor, so read the "billing unit" column before comparing prices.
| Tool | Free tier | Entry paid plan | Billing unit | Self-host licence | OpenTelemetry |
|---|---|---|---|---|---|
| Langfuse | Hobby: 50k units/mo, 30 days data access, 2 users | Core $29/mo, 100k units, then $8 per 100k | Units (traces, observations, scores) | MIT, except ee folders | Built on OTel |
| LangSmith | Developer: 5,000 base traces/mo, 1 seat | Plus $39/seat/mo, 10,000 base traces | Traces (base 14-day, extended 180-day retention) | Self-hosted option offered; plan not stated in docs | Not confirmed in docs we read |
| Helicone | Hobby: 10,000 requests, 1 GB, 7-day retention, 1 seat | Pro $79/mo, 1-month retention, unlimited seats | Requests and storage, usage-based above free | Apache-2.0 | Not checked |
| Arize Phoenix / AX | Phoenix: open source. AX Free: 25k spans/mo, 1 GB, 15 days | AX Pro $50/mo, 50k spans, 10 GB, 30 days | Spans and GB | Phoenix: Elastic License 2.0 | Built on OTel |
| Comet Opik | Free Cloud: 25k spans/mo, 60 days, up to 10 members | Pro $19/mo, 100k spans, then $5 per 100k | Spans | Apache-2.0 | First-party OTel support |
| Laminar | Free: 1 GB, 7 days, 1 seat, 1 project | Starter $30/mo, 3 GB, then $2/GB, 30 days | GB of data | Apache-2.0 | OTel-native SDK |
| Braintrust | Starter: 1 GB processed data, 10k scores, 14 days | Pro $249/mo, 5 GB, then $3/GB | GB processed plus scores | Not checked | Not checked |
| W&B Weave | 1 GB/mo ingestion | Pro starts at $60/mo, extra ingestion $0.10/MB | MB/GB ingested | Not checked | Not checked |
| Traceloop (OpenLLMetry) | Free Forever: up to 50K spans/mo, 24-hour retention | Enterprise only, by consultation | Spans | OpenLLMetry SDK: Apache-2.0 | Based on OTel |
| Datadog Agent (LLM) Observability | Up to 40K LLM spans/mo | Pro $160/mo (annual), 100K LLM spans | LLM spans | Not checked | Not checked |
Each figure is cited in the tool sections below.
The tools, one by one#
Langfuse: the default for most indie developers#
Langfuse's Hobby plan is free with 50k units a month, 30 days of data access and 2 users. Core is $29/mo with 100k units and unlimited users, then $8 per 100k units up to 1M. Pro is $199/mo with 3 years of data access (Langfuse pricing).
The catch is the unit. Langfuse says "A billable unit in Langfuse is any tracing data point sent to the platform", which covers traces, observations (spans, events, generations) and scores. One user request with a retrieval step and two LLM calls is several units, not one.
The repo is "MIT licensed, except for the ee folders", and the README says Langfuse is "built on OpenTelemetry" (Langfuse GitHub). Its pricing page states "Langfuse is open source and you can self-host it for free."
Our read: the best mix of a real free tier, a cheap first paid step and an exit route to self-hosting.
Comet Opik: the cheapest paid tier#
Opik's Free Cloud plan gives 25k spans a month, 60-day retention and up to 10 team members. Pro is $19/mo for 100k spans, with extra spans at $5 per 100k and extended 400-day retention at $29 per 100k spans (Comet pricing). The open source version is Apache-2.0 and has "first-party OpenTelemetry support" (Opik GitHub).
Our read: at a few hundred thousand spans a month, Opik's overage undercuts Langfuse, and its free retention is twice as long.
Helicone: the five-minute option#
Helicone is a proxy: its gateway is "an OpenAI-compatible, unified API", so you change your SDK's base URL and add a Helicone key (Helicone docs). Hobby is free with 10,000 requests, 1 GB, 7-day retention and one seat. Pro is $79/mo with 1-month retention, then usage-based pricing with no per-request rate listed (Helicone pricing). The repo is Apache-2.0 and self-hostable with Docker Compose (Helicone GitHub).
Our read: ideal for single-call apps, but a proxy sits in your request path, unlike SDK tools that log asynchronously.
LangSmith: if you live in LangChain#
LangSmith's Developer plan is $0 for one seat with 5,000 base traces a month; Plus is $39 per seat per month with 10,000 base traces and unlimited seats. Base traces are kept 14 days, extended traces 180 days, and overage is pay as you go with no rate shown (LangChain pricing). The docs say it "works with many frameworks and providers" and offers cloud, hybrid or self-hosted setups (LangSmith docs).
Our read: the smallest free tier here, and per-seat pricing bites when a second person joins. Pick it if you already use LangGraph.
Arize Phoenix and Arize AX: local-first#
Phoenix is an open source, local-first tool "built on top of OpenTelemetry", licensed under the Elastic License 2.0 (Phoenix GitHub). ELv2 is source-available, not OSI open source. The hosted Arize AX Free plan gives 25k spans, 1 GB and 15 days; AX Pro is $50/mo for 50k spans, 10 GB and 30 days (Arize pricing).
Laminar: agent traces, billed by GB#
Laminar bills by data, not events: Free is 1 GB with 7-day retention and one seat; Starter is $30/mo for 3 GB and $2 per extra GB with 30-day retention; Pro is $150/mo for 10 GB and 6 months (Laminar pricing). The code is Apache-2.0 with an "OpenTelemetry-native" SDK and Docker Compose self-hosting (Laminar GitHub).
Our read: GB billing suits agents that make many small calls; it penalises apps that log long documents in every prompt.
Braintrust and W&B Weave: eval-first platforms#
Braintrust Starter is $0 with 1 GB processed data, 10k scores and 14-day retention; Pro is $249/mo (Braintrust pricing). W&B Weave includes 1 GB of ingestion a month, with Pro from $60/mo and extra ingestion at $0.10/MB (CoreWeave/W&B pricing). Both suit eval-heavy teams with budget.
Traceloop and OpenLLMetry: instrument once#
OpenLLMetry describes itself as "Open-source observability for your GenAI or LLM application, based on OpenTelemetry" and is Apache-2.0 (OpenLLMetry GitHub). Traceloop's own hosted Free Forever plan allows up to 50K spans a month but keeps data for only 24 hours (Traceloop pricing).
Our read: the SDK is the useful part. OpenTelemetry instrumentation can point at any OTel-compatible backend later, the cleanest defence against lock-in.
Sentry, Datadog and PostHog: what you may already have#
- Sentry: its docs say agent monitoring "automatically collects information about agent runs, tool calls, model interactions, and errors", including token usage and cost (Sentry docs).
- Datadog: Agent Observability lists a free tier of up to 40K LLM spans a month and Pro at $160/mo billed annually with 100K spans, then $3.50 per 10K spans annually or $5 per 10K on demand (Datadog pricing).
- PostHog: offers LLM analytics with an OpenTelemetry install option (PostHog docs). We could not verify its LLM free allowance today.
What to log from day one#
Whichever tool you pick, capture these fields on every LLM call:
- Prompt, or a prompt id plus variables.
- Response, including tool calls.
- Model, with the exact model id.
- Tokens in and out, plus cached tokens.
- Cost, the number that tells you where to start in how to reduce LLM costs.
- Latency, total and time to first token.
- User or account id, hashed if you prefer.
- Prompt or app version, so regressions line up with deploys.
- Session id, to group chats and agent runs.
- User feedback, even a thumbs up or down.
Privacy: redact before you send#
Traces contain whatever your users typed. Our read: strip emails, phone numbers and secrets before prompts leave your server, store ids rather than names, and list the vendor as a subprocessor. If customers ask where their data goes, self-hosting Langfuse, Opik, Helicone or Laminar keeps traces on infrastructure you control.
Worked example: the same app on three tools#
Illustrative arithmetic, not a quote: a small RAG app with 3,000 requests per day for 30 days (90,000 requests), each one trace with three steps (retrieval plus two LLM calls).
| Tool | Usage in its unit | Monthly cost on list prices |
|---|---|---|
| Langfuse Core | 90,000 traces + 270,000 observations = 360,000 units. 100k included, 260k extra at $8 per 100k = $20.80 | $29 + $20.80 = $49.80 |
| Comet Opik Pro | 270,000 spans. 100k included, 170k extra at $5 per 100k = $8.50 | $19 + $8.50 = $27.50 |
| Datadog Pro (annual) | 270,000 LLM spans if all three count. 100K included, 170K extra at $3.50 per 10K = $59.50 | $160 + $59.50 = $219.50 |
Our read: the unit matters more than the headline price, and sampling matters most. Log every trace during launch, then sample routine successes and keep every error and thumbs down.
Decision guide by stage#
| Your stage | Our pick |
|---|---|
| Prototype, solo, $0 budget | Langfuse Hobby, or Phoenix locally |
| One LLM call per request, want it running tonight | Helicone Hobby |
| Launched, a few hundred thousand spans a month | Opik Pro ($19/mo) or Langfuse Core ($29/mo) |
| Customer data must stay on your servers | Self-host Langfuse, Opik or Laminar |
| Built on LangChain or LangGraph | LangSmith |
| Agent with many small steps | Laminar or Langfuse |
| Already paying for Sentry or Datadog | Try their AI monitoring before adding a tool |
| Evals are the main job, budget available | Braintrust or W&B Weave |
Our read: use an OpenTelemetry-based SDK, start free, and pay only when you hit a real limit. Shipping soon? Our bootstrapped app launch guide covers the rest.
Frequently asked questions#
What is the best free LLM observability tool for indie developers?#
For most solo developers, Langfuse. Its Cloud Hobby plan is free for 50k units a month with 30 days of data access and 2 users, and the repository is MIT-licensed (except the ee folders) if you want to self-host. Comet Opik (25k spans free, 60-day retention) and Arize AX Free (25k spans, 15 days) are close alternatives.
What are the best Langfuse alternatives?#
Comet Opik if you want a cheaper paid tier ($19/mo for 100k spans) and an Apache-2.0 licence. Helicone if you want a proxy you can add by changing a base URL. Arize Phoenix if you want an OpenTelemetry-based tool you run locally. LangSmith if you already build with LangChain or LangGraph. Laminar if you are tracing agents and prefer GB-based billing.
Do I need an LLM observability tool if I already use Sentry or Datadog?#
Maybe not at first. Sentry's docs say its agent monitoring collects agent runs, tool calls, model interactions and errors, and shows token usage and cost. Datadog's pricing page lists a free tier of up to 40K LLM spans a month. A dedicated tool adds prompt management, datasets and evals that general monitoring tools are not built around.
Can I self-host LLM observability for free?#
Yes. Langfuse (MIT except ee folders), Comet Opik (Apache-2.0), Helicone (Apache-2.0) and Laminar (Apache-2.0) all publish their code on GitHub with self-hosting instructions. Arize Phoenix is under the Elastic License 2.0, which is source-available rather than OSI open source. You pay in server costs and your own maintenance time.
How do I estimate what LLM tracing will cost?#
Count how many LLM calls, retrieval steps and tool calls one user request creates, multiply by requests per month, then convert into the vendor's unit. Langfuse bills every trace, observation and score as a unit; Opik and Datadog bill spans; Laminar, Braintrust and W&B Weave bill by data volume. Our worked example shows the same workload costing very different amounts on different tools.
Ship it and show it#
ZeroRaised is a launch directory for products built without institutional VC. Once your AI feature is traced and behaving, submit your product for free. For more tool comparisons for bootstrapped builders, join the ZeroRaised newsletter.
Sources#
- Langfuse pricing, last checked 13 Nov 2026
- Langfuse GitHub repository, last checked 13 Nov 2026
- LangChain (LangSmith) pricing, last checked 13 Nov 2026
- LangSmith documentation, last checked 13 Nov 2026
- Helicone pricing, last checked 13 Nov 2026
- Helicone quick start, last checked 13 Nov 2026
- Helicone GitHub repository, last checked 13 Nov 2026
- Arize pricing, last checked 13 Nov 2026
- Arize Phoenix GitHub repository, last checked 13 Nov 2026
- Comet (Opik) pricing, last checked 13 Nov 2026
- Opik GitHub repository, last checked 13 Nov 2026
- Laminar pricing, last checked 13 Nov 2026
- Laminar GitHub repository, last checked 13 Nov 2026
- Braintrust pricing, last checked 13 Nov 2026
- CoreWeave / W&B pricing, last checked 13 Nov 2026
- Traceloop pricing, last checked 13 Nov 2026
- OpenLLMetry GitHub repository, last checked 13 Nov 2026
- Datadog Agent Observability pricing, last checked 13 Nov 2026
- Sentry AI agent monitoring docs, last checked 13 Nov 2026
- PostHog LLM analytics docs, last checked 13 Nov 2026
Frequently asked questions
What is the best free LLM observability tool for indie developers?
For most solo developers, Langfuse. Its Cloud Hobby plan is free for 50k units a month with 30 days of data access and 2 users, and the repository is MIT-licensed (except the ee folders) if you want to self-host. Comet Opik (25k spans free, 60-day retention) and Arize AX Free (25k spans, 15 days) are close alternatives.
What are the best Langfuse alternatives?
Comet Opik if you want a cheaper paid tier ($19/mo for 100k spans) and an Apache-2.0 licence. Helicone if you want a proxy you can add by changing a base URL. Arize Phoenix if you want an OpenTelemetry-based tool you run locally. LangSmith if you already build with LangChain or LangGraph. Laminar if you are tracing agents and prefer GB-based billing.
Do I need an LLM observability tool if I already use Sentry or Datadog?
Maybe not at first. Sentry's docs say its agent monitoring collects agent runs, tool calls, model interactions and errors, and shows token usage and cost. Datadog's pricing page lists a free tier of up to 40K LLM spans a month. A dedicated tool adds prompt management, datasets and evals that general monitoring tools are not built around.
Can I self-host LLM observability for free?
Yes. Langfuse (MIT except ee folders), Comet Opik (Apache-2.0), Helicone (Apache-2.0) and Laminar (Apache-2.0) all publish their code on GitHub with self-hosting instructions. Arize Phoenix is under the Elastic License 2.0, which is source-available rather than OSI open source. You pay in server costs and your own maintenance time.
How do I estimate what LLM tracing will cost?
Count how many LLM calls, retrieval steps and tool calls one user request creates, multiply by requests per month, then convert into the vendor's unit. Langfuse bills every trace, observation and score as a unit; Opik and Datadog bill spans; Laminar, Braintrust and W&B Weave bill by data volume. Our worked example shows the same workload costing very different amounts on different tools.