LangSmith
The broadest agent platform we have scored, and the one most likely to surprise you on the invoice: Engine rescans every 6 hours at 10-15 LCU a run, which is 60 to 90 dollars a day per project if you leave the spend limit blank.
PRICING
| Developer | $0 | 1 seat, 5,000 base traces per month then pay as you go, 14-day retention, community support, Fleet 5 LCU per month, Sandboxes 5 LCU and 1 LSU per month capped at 10 sandboxes. No deployment, no Engine, no bulk export. Cloud only. Hard monthly cap of 5,000 traces with no payment method on file |
| Plus | $39/seat/mo | 10,000 base traces per month then pay as you go, unlimited seats, one free small serverless deployment, Engine, Fleet 25 LCU per month, bulk data export, support portal and email. Cloud only. No SSO |
| Enterprise | Custom | Cloud, hybrid (SaaS control plane, your VPC data plane) or fully self-hosted on Kubernetes, custom SSO, ABAC and RBAC, support SLA, deployed engineers, custom terms and infosec review, annual invoice |
| Startup | Discounted | VC-backed early-stage companies, up to $10,000 in credits and higher free trace allotments |
Three meters stack, and the second and third are the ones that bite. First, seats at $39 before a single trace is logged. Second, traces: base traces are $0.0005 each and keep 14 days, extended traces keep 400 days and cost ten times as much, and retention extension is enabled by default on new online evaluators and automation rules, so switching on a judge silently multiplies the cost of every trace it touches. LangSmith also gives you usage limits, not spend limits, which is not the same thing. Third, Engine: initialisation runs 30-40 LCU and every scheduled scan runs 10-15 LCU, four scans a day, which is roughly $60 to $90 per day per project unless an admin sets a monthly LCU cap. The cap field is blank by default. Note that several third-party pricing guides published in 2026 quote $2.50 per 1,000 base traces; the official docs say $0.0005 per trace, and the docs win.
checked 2026-08-14 · vendor pricing page
Element scores
Strengths
LangSmith is no longer an observability tool, it is the whole agent lifecycle in one account, and that is the honest reason to buy it. Tracing is the deepest here and it is not locked shut: LangSmith ingests plain OpenTelemetry from any language, documents its attribute mapping against the GenAI, TraceLoop, OpenInference and Logfire conventions, supports collector fanout to a second backend, and ships LANGSMITH_OTEL_ONLY so you can point the same instrumentation somewhere else. That is a real answer to the lock-in complaint, and most reviews of this product have not caught up with it. Evaluation is complete on every tier, offline and online, with LLM-as-judge, pairwise, code and human evaluators, assertions, annotation queues and versioned datasets with splits. Engine is the standout: it clusters recurring failures out of production traces, diagnoses them against your connected repository, opens a pull request, generates an evaluator and ground-truth examples so the same failure cannot come back unnoticed, and reopens the issue itself if it recurs. Around it sits infrastructure nobody else in this category ships, namely managed deployment with scale-to-zero and cron, ephemeral sandboxes for agent-generated code, and an LLM Gateway that does PII and secrets redaction, spend caps and model fallbacks in the request path. Protocol support is strong: a hosted remote MCP server on OAuth 2.1 across four cloud regions plus BYOC and self-hosted, and any deployed agent can be published as an MCP server itself. Compliance is enterprise-grade, with SOC 2 Type II, HIPAA and GDPR, AES-256 at rest, TLS 1.2 or higher in transit, annual third-party audits and penetration tests.
Honest dings
Cost control is the thing to fix on day one. Three meters stack, seats at $39 each before you log anything, traces at $0.0005 for 14 days or ten times that for 400, and LCU-metered products on top. Retention extension is on by default for new online evaluators and automation rules, so the moment you switch on a judge you are paying ten times per trace for every trace it scores. LangSmith gives you usage limits, not spend limits, which is a real distinction when a chatty agent emits thousands of runs an hour. Engine is the sharpest edge: it rescans every 6 hours whether or not anything is wrong, at 10-15 LCU a scan and $1.50 an LCU, so a project left with a blank spend limit runs at roughly $60 to $90 a day. The cap is blank by default and an Organization Admin has to set it. Engine also runs exclusively on LangChain-managed inference, BYOK is explicitly unsupported, so enabling it means your trace data is processed with the vendor's LLM keys. Self-hosting is Enterprise-only and closed-source, which is the clean separation from Langfuse and Opik: with those two you can run the real product for free, here you are booking a sales call. For EU buyers there is one more thing worth knowing before procurement gets there: GCP EU and AWS US regions exist and GDPR compliance is claimed, but LangChain states in its own regions FAQ that it has no EU legal entity to contract with today. Finally, Engine's scan scope filter only accepts run name and metadata, not feedback key or score threshold, so you cannot point it precisely at the traces your evaluators already flagged.
Sources (15) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- langchain.com/pricing — Official pricing verified live 14/08/2026: Developer $0 with 1 seat and 5k base traces, Plus $39 per seat with 10k base traces, Enterprise custom. LCU $1.50 and LSU $1.00 definitions, deployment metering rates by resource, Fleet allowances of 5 and 25 LCU, Sandbox allowances and rates, LLM Gateway feature matrix, hosting options (Cloud, Hybrid, Self-Hosted), SSO and RBAC availability by tier, and the statement that LangSmith does not train on customer data (accessed 2026-08-14)
- docs.langchain.com/langsmith/usage-and-billing — Official trace pricing and the primary correction to third-party guides: base charge is $0.0005 per trace with 14-day retention, extended retention is 400 days at ten times the price ($0.005 per trace, $0.0045 per upgrade). Also the finding that retention extension is enabled by default on new online evaluators and automation rules, that LangSmith offers usage limits rather than spend limits, and the full rate-limit table by plan (accessed 2026-08-14)
- docs.langchain.com/langsmith/engine — Official Engine documentation and the source of the cost finding: initialisation 30-40 LCU, recurring scans every 6 hours at 10-15 LCU, $1.50 per LCU, spend limit blank by default. Also the closed loop (detect, diagnose, pull request, generate evaluator and ground-truth examples, auto-reopen), the explicit statement that Engine uses LangChain-managed inference and does not support BYOK, and the scope-filter limitation to run name and metadata only (accessed 2026-08-14)
- docs.langchain.com/langsmith/trace-with-opentelemet… — Official OpenTelemetry documentation: OTLP ingestion from any OTel-compatible application, full attribute mapping tables for GenAI, TraceLoop, OpenInference and Logfire conventions, collector fanout to multiple backends, LANGSMITH_OTEL_ONLY to send traces elsewhere, trace attachments for multimodal agents, and the warning that a span whose parent never reaches LangSmith is silently dropped (accessed 2026-08-14)
- docs.langchain.com/langsmith/langsmith-mcp-server — Official MCP documentation: hosted remote MCP server with OAuth 2.1 and dynamic client registration on GCP US, GCP EU, GCP APAC and AWS US, plus BYOC and self-hosted from v0.15, and the full tool list covering threads, prompts, runs, datasets, experiments and billing. Several tools are documentation-only rather than executable (accessed 2026-08-14)
- docs.langchain.com/langsmith/evaluation — Official evaluation documentation: offline and online evaluation flows, human, code, LLM-as-judge and pairwise evaluators, experiment configuration with repetitions, concurrency and caching, benchmarking, unit tests, regression tests and backtesting, and the note that cloud, hybrid and self-hosted all include observability, evaluation, prompt engineering and deployment (accessed 2026-08-14)
- docs.langchain.com/langsmith/manage-datasets — Official dataset documentation: automatic versioning on every change, human-readable version tags, splits and filtered views, export to CSV, JSONL and OpenAI fine-tuning format, exporting filtered experiment traces back into datasets, and confirmation that Engine can generate ground-truth examples from production traces (accessed 2026-08-14)
- docs.langchain.com/langsmith/regions-faq — Official regions FAQ: SOC 2 Type 2 certified, HIPAA compliant, GDPR compliance claimed, four functionally equivalent cloud regions available on all plans including free, and the explicit statement that LangChain has no legal entity in the EU for customer contracting today (accessed 2026-08-14)
- docs.langchain.com/langsmith/shared-responsibility-… — Official security posture: SOC 2 Type II, GDPR and HIPAA with annual third-party audits and penetration testing, AES-256 at rest and TLS 1.2 or higher in transit, tenant isolation on GCP, and the customer's own responsibility to mask PII at source before it leaves their environment (accessed 2026-08-14)
- langchain.com/about — Official company facts: LangChain started as Harrison Chase's side project in late 2022, company founded with Ankush Gola in early 2023, headquartered in San Francisco with offices in New York, Boston and Amsterdam, works with 35% of the Fortune 500, over 1 billion open-source downloads, over 1 billion events per day ingested on LangSmith, backed by IVP, Benchmark and Sequoia (accessed 2026-08-14)
- techcrunch.com/2025/10/21/open-source-agentic-start… — Independent confirmation of the $125M raise at a $1.25B valuation (October 2025), the funding position behind the current platform build-out (accessed 2026-08-14)
- llmtools.cc/blog/langfuse-vs-langsmith — Independent head-to-head (26/07/2026): documents that LangChain cut its old free tier and a visible wave of users moved to self-hosted Langfuse, prices LangSmith at roughly 25x Langfuse at production scale, and confirms self-hosting is Enterprise-only with a closed-source backend (accessed 2026-08-14)
- apiscout.dev/guides/langsmith-vs-langfuse-vs-braint… — Independent three-way comparison (15/03/2026): confirms 5k traces per month free and $39 per seat, rates the zero-config LangChain and LangGraph tracing as best in class, and flags per-seat pricing plus limited self-hosting as the trade-off (accessed 2026-08-14)
- bigdataboutique.com/blog/llm-observability-tools-co… — Independent comparison (30/03/2026): rates LangSmith's evaluation framework as the most complete of the three, and notes that the closed-source backend means you cannot inspect or modify the platform and that non-LangChain stacks require more manual integration work (accessed 2026-08-14)
- chatforest.com/reviews/langsmith-llm-observability-… — Independent review (06/05/2026): frames the central tension between being the default for the largest LLM framework community and being aspirationally framework-agnostic, and calls the Enterprise-only self-hosting restriction the deciding question against fully open alternatives (accessed 2026-08-14)