The best LLM evaluation tools in 2026 are Braintrust for a polished hosted eval workflow, Langfuse for open-source tracing and evals you can self-host free, LangSmith for LangChain teams, Inspect for rigorous statistics, DeepEval for the largest metric library, and Promptfoo for evals and red-teaming in CI. Kojen fits teams that want confidence intervals and pass^k on every run and a loop that also tunes, red-teams and deploys the agent.
One change matters for anyone choosing now: OpenAI's hosted Evals went read-only on October 31, 2026 and shut down on November 30, 2026.
Best for: a polished hosted eval and experiment workflow
Well-built experiments, scorers and side-by-side comparison, online scoring of production logs, and the Loop agent for suggesting prompt and scorer edits. Unlimited users on the free tier. No confidence intervals, red-teaming or self-hosting below Enterprise.
Pricing: Starter $0 (unlimited users, 1 GB, 10k scores, 14-day retention); Pro $249/month; Enterprise custom. Source · Open source: no
Strengths
A mature, polished eval and tracing product, with unlimited users even on the free tier.
Loop works across logs, experiments, datasets and scorers in one place.
Gateway, environments and online scoring of production traffic in the same product.
Limits
No confidence intervals or significance tests, and no pass@k beyond grouping trials.
Measures and suggests, but doesn't train weights, red-team or serve models.
Best for: open-source tracing and evals you can self-host free
MIT-licensed and free to self-host with all core features, with cheap cloud plans. Strong tracing, datasets and LLM judges; no repetitions per item, statistics or prompt optimizer.
Pricing: Cloud: Hobby free (50k units/month); Core $29/month; Pro $199/month; Enterprise $2,499/month. Self-hosted open source: free. Source · Open source: yes (MIT (except ee/ folders))
Strengths
Genuinely open source (MIT) with free, unrestricted self-hosting.
Inexpensive usage-based cloud pricing, with long retention on Pro.
An MCP server that can write prompts, datasets and scores as well as read them.
Limits
No repetitions per dataset item, so no pass@k-style reliability metrics, and no statistical tests.
No prompt optimizer, red-teaming or training: it observes and evaluates, you improve elsewhere.
Best for: teams building on LangChain or LangGraph
Tracing, evaluation, annotation queues and agent hosting in one product, with repetitions reported as mean and standard deviation. Per-seat pricing.
Pricing: Developer $0 (1 seat, 5k base traces/month); Plus $39/seat/month (10k traces); Enterprise custom. Usage beyond that is pay as you go. Source · Open source: no
Strengths
The natural choice for LangChain and LangGraph teams, with production hosting for agents.
Strong tracing, annotation queues and prompt management.
Cloud, hybrid and self-hosted deployment options.
Limits
Per-seat pricing, and the free tier has one seat.
Reliability stats stop at mean and standard deviation; no intervals or pass^k.
Best for: rigorous statistics and agent evals in code
The UK AI Security Institute's framework has the best statistics we found (clustered standard errors, bootstrap and Wilson intervals, pass@k and pass^k reducers), sandboxed agent runs and 200+ pre-built evals. No hosted team platform.
Pricing: Free, MIT. Source · Open source: yes (MIT)
Strengths
Rigorous statistics: clustered standard errors, bootstrap and Wilson intervals, pass@k and pass^k.
Sandboxed agent evaluation (Docker, Kubernetes, Modal and more), including external agents.
A credible public-sector maintainer and a large benchmark library.
Limits
Researcher-oriented: no hosted team platform or production tracing.
No red-team attack generation and no prompt optimization.
Best for: evaluations you can trust, plus improving what they measure
Every rate comes with a confidence interval, several samples per case give pass@k and pass^k, judges come from a different model family than the target in simulations, and red-team, prompt tuning and deployment share the same test cases. It doesn't trace apps running elsewhere. We make Kojen.
Pricing: Free to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source · Open source: no
Strengths
Measures and improves in one loop: evaluations, red-teaming, prompt tuning, model search and deployment share the same project and test cases.
Honest statistics by default: every rate has a confidence interval, tuned prompts must beat the current one in a paired test, and pass@k / pass^k come from several samples per case.
Agent-native: the full loop is exposed over an MCP server and a REST API with an OpenAPI spec, with the same 24 tools as the in-app agent.
Limits
Young product: weight training is early access on 8B models, LoRA adapters can't be served yet, and full fine-tuning and RLHF are previews.
No tracing SDK for apps running elsewhere, so it is not a replacement for an observability platform.
Best for: the largest library of ready-made metrics
Open-source metrics for RAG, agents, multi-turn conversations and MCP, built-in GEPA and MIPROv2 prompt optimizers, and a hosted platform with tracing. Paid tiers start at $200 a month.
Pricing: Free (2 seats, 1 project); Starter $200/month; Team $2,000/month; Team++ and Enterprise custom. Priced per organization, not per seat. Source · Open source: yes (Apache-2.0)
Strengths
One of the largest metric libraries, including agent, multi-turn and MCP metrics.
Prompt optimizers (GEPA, MIPROv2) built in.
A hosted MCP server and tracing platform on top of the open-source library.
Limits
Red-teaming is a separate library, and in the hosted product it is Enterprise-only.
Steep jump between paid tiers, from $200 to $2,000 a month.
Free, MIT-licensed YAML test suites with the broadest red-team coverage, now owned by OpenAI and recommended by OpenAI as the replacement for its hosted Evals. No statistics or production monitoring.
Pricing: Community: free forever, including self-hosting and red-teaming up to 10k probes a month. Enterprise and On-Premise: custom pricing. Source · Open source: yes (MIT)
Strengths
The widest red-team coverage available, with presets for OWASP LLM and agentic Top 10s, NIST AI RMF, MITRE ATLAS and the EU AI Act.
MIT-licensed config-as-code that lives in your repo and runs in CI.
OpenAI now names it as the migration path off OpenAI's own hosted Evals, which shut down on November 30, 2026.
Limits
No statistics built in: no confidence intervals and no pass@k metric.
Now owned by OpenAI, which some teams testing other vendors' models will weigh.
Doesn't improve the model it finds problems in; there is no tuning or serving.
Best for: OpenTelemetry-native tracing with free self-hosting
Free to self-host without feature gates, experiment repetitions, and automated Prompt Learning on Arize AX. Source-available under the Elastic License 2.0.
Free (2 seats, 1 project); Starter $200/month; Team $2,000/month; Team++ and Enterprise custom. Priced per organization, not per seat. Source
Community: free forever, including self-hosting and red-teaming up to 10k probes a month. Enterprise and On-Premise: custom pricing. Source
Free to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source
Automatically rewrites instructions or examples to raise a metric.
Partial
Loop agent suggests prompt edits you accept
No
Prompt versioning and A/B labels, no optimizer
Partial
Optimization jobs (promptim) via API; Polly edits prompts on request
No
Yes
PromptOptimizer with GEPA and MIPROv2
Partial
promptfoo optimize: one prompt against one provider
Yes
Prompt tuning plus search across 11 hosted models
Red-teaming
Generates adversarial attacks (prompt injection, jailbreaks, data leaks) and scores them.
No
Only a prompt-injection-detector cookbook
No
No
Partial
Safety and agentic benchmarks; no attack generator
Yes
DeepTeam library; hosted red-teaming is Enterprise-only
Yes
Dozens of plugins; Crescendo, GOAT, tree and Best-of-N strategies; OWASP/NIST/MITRE presets
Yes
Prompt injection, jailbreak, data exfiltration, harmful content, PII leak
Production tracing
Logs and traces live traffic from your app.
Yes
Tracing and online scoring
Yes
OpenTelemetry-based tracing
Yes
Partial
Inspect View log viewer, not production traffic
Yes
Tracing in Confident AI
Partial
OpenTelemetry traces per test case, not production monitoring
Partial
Request logs, latency, errors and feedback for Kojen endpoints only; no tracing SDK
Official MCP server
An MCP server from the vendor that lets agents drive the product.
Yes
Hosted MCP server
Yes
Authenticated MCP server, read and write
Yes
Python server and hosted remote MCP with OAuth
Partial
Agents under test can use MCP tools; no Inspect MCP server
Yes
Hosted MCP server
Yes
promptfoo mcp, plus MCP red-team testing
Yes
ai.kojen/kojen in the MCP Registry, OAuth sign-in
Self-host
You can run it on your own infrastructure.
Partial
Enterprise only; you run the data plane, Braintrust the control plane
Yes
Free, no core-feature limits
Partial
Enterprise add-on
Yes
Partial
Libraries run locally; platform on-prem is Enterprise
Yes
Free in the Community tier
No
Hosted only
What separates a good evaluation tool
Most tools now run datasets and LLM judges well. The differences are in what they tell you about uncertainty. A pass rate of 80% on 20 cases has a 95% interval of roughly 58% to 92%, so a change from 80% to 85% on that set means very little; only Inspect and Kojen report intervals out of the box. Likewise, an agent that succeeds 75% of the time per try succeeds on all three of three tries only about 42% of the time, which is why pass^k matters for anything customer-facing.
The other split is between tools that observe production traffic (Braintrust, Langfuse, LangSmith, Phoenix, Weave) and tools that test before release (Inspect, DeepEval, Promptfoo, Kojen). Many teams need one of each.
How we made this list
We scored each tool on what an evaluation needs to be trustworthy: datasets, LLM judges, repeated samples per case (pass@k / pass^k), confidence intervals, and whether it helps you act on the result (prompt optimization, red-teaming).
Order reflects how broadly useful each tool is for evaluating LLM apps and agents in production, not a single score.
Every feature claim links to the vendor's own pricing page, docs or repository, checked on November 15, 2026. Where we couldn't confirm a feature, the table says "Not verified".
We make Kojen. We include it where it does the job and say where others are stronger; it is not placed first by default.
No vendor paid to be included, and there are no affiliate links.
Frequently asked questions
What is the best open-source LLM evaluation tool?
Langfuse (MIT) for a full platform with tracing you can self-host; Inspect (MIT) for rigorous statistics and agent evals; DeepEval (Apache-2.0) for the widest metric library; Promptfoo (MIT) for evals and red-teaming in CI.
Which LLM evaluation tools report confidence intervals?
Of the tools we checked in November 2026, Inspect (stderr, bootstrap and Wilson interval metrics) and Kojen (an interval on every rate) do. Braintrust, LangSmith, Langfuse, Phoenix, Weave, DeepEval and Promptfoo report averages, deltas or standard deviations.
What is the difference between pass@k and pass^k?
pass@k is the chance at least one of k tries succeeds and rises with k; pass^k, introduced with τ-bench, is the chance all k tries succeed and falls with k. Use pass^k for agents that must be reliable every time.
What happened to OpenAI Evals?
OpenAI's hosted Evals dashboard, API and graders went read-only on October 31, 2026 and shut down on November 30, 2026; OpenAI recommends migrating to Promptfoo. The open-source openai/evals repository still exists.
Run an LLM-judged evaluation with confidence intervals, a red-team pass and a prompt-tuning run on your agent. Free to start, no card needed, then pay as you go.