Why engineering teams choose Last9 vs Honeycomb
Both are OpenTelemetry-native and both are built for high-cardinality data. Honeycomb is organised around wide events, explores traces fast, and bills per event, with sampling as the main way to hold that bill down. Last9 keeps every trace, covers metrics, logs, and traces on one licence, and runs its AI agent on your model provider.
100+ native integrations, OpenTelemetry-first. A Last9 engineer runs the move with your team.
Named a Gartner® Cool Vendor in AI for SRE and Observability, cited for its unified telemetry platform.
The Last9 agent runs on your model provider, with no per-investigation charge, in your own cloud account.
No meter counting every span, and no plan tier to unlock your own cloud. The infrastructure runs on your own cloud bill.
At a glance
Last9 vs Honeycomb at a glance
The questions engineering teams ask when they compare the two.
Does it have an MCP server for coding agents?
Yes, hosted, since March 2025, for any MCP client. An agent can act through it: add drop rules and edit dashboards
Yes, and a broad one: read tools across schema, queries, BubbleUp, traces, boards, triggers, and SLOs, plus write tools that create and update boards, triggers, SLOs, and recipients. None of them reach ingest rules, and there are no delete tools
Does it have an AI SRE agent?
Yes. The Last9 agent investigates in the app, in Slack, and in your IDE, and builds a diagnosis dashboard from what it finds
Yes. Canvas is an AI-guided workspace inside the product, and Agent Skills give Claude Code and Cursor knowledge of instrumentation and investigation. Automated investigations are early access
Can the AI use our model provider, in our own cloud?
Yes. The Last9 agent runs on your model provider or inference endpoint, in your own cloud account, with no per-investigation charge
Not documented. Honeycomb Intelligence runs in Honeycomb's own service, with no published option to point it at your own model provider
Can it monitor LLM apps and coding agents?
Yes. Agents Monitoring reports conversations, tokens, cost, latency, and success rate from OpenTelemetry GenAI telemetry, with a separate view for coding agents
For LLM apps, yes: Agent Timeline lays out a conversation across traces, with LLM calls, tool calls, failures, and token use, on every plan. Honeycomb documents no view of what your coding agents do and cost
Is trace data sampled?
No. Every trace is kept. You decide what to drop or aggregate, with rules you preview against live data
Usually, and it is the main cost lever. Refinery does trace-aware tail-based sampling so you keep every error trace and a fraction of the rest. It is a cluster your team sizes and runs
Can I filter telemetry before I pay to store it?
Yes, from one place. The Control Plane drops, remaps, redacts, forwards, and aggregates metrics, logs, and traces at ingest, from the UI
Yes, in several places you operate. Refinery samples, collector processors filter, and the Telemetry Pipeline add-on on Enterprise filters, samples, redacts, and extracts metrics from logs
Does it cover metrics, logs, and traces equally?
Yes. Metrics, logs, and traces are first-class, with PromQL and LogQL over them and APM built from the same OpenTelemetry data
Traces and wide events are the centre of the product. Logs are stored as events against the same allotment, and metrics became generally available in March 2026
How is it priced?
One licence. The infrastructure runs on your own cloud bill, at your negotiated rates
Published plans, billed per event, where every span and log record counts and a 150-span trace is 150 events. A free plan and a published Pro rate, with unlimited seats, and a quote above the Pro ceiling
Can it run in our own cloud account?
Yes. Last9 runs in your own cloud account, on AWS or GCP in any region, and manages it for you
Yes. Private Cloud runs the same platform in your own AWS account, operated by Honeycomb or by your team, and is listed on the Enterprise plan
| Capability | Last9 | Honeycomb |
|---|---|---|
| Does it have an MCP server for coding agents? | Yes, hosted, since March 2025, for any MCP client. An agent can act through it: add drop rules and edit dashboards | Yes, and a broad one: read tools across schema, queries, BubbleUp, traces, boards, triggers, and SLOs, plus write tools that create and update boards, triggers, SLOs, and recipients. None of them reach ingest rules, and there are no delete tools |
| Does it have an AI SRE agent? | Yes. The Last9 agent investigates in the app, in Slack, and in your IDE, and builds a diagnosis dashboard from what it finds | Yes. Canvas is an AI-guided workspace inside the product, and Agent Skills give Claude Code and Cursor knowledge of instrumentation and investigation. Automated investigations are early access |
| Can the AI use our model provider, in our own cloud? | Yes. The Last9 agent runs on your model provider or inference endpoint, in your own cloud account, with no per-investigation charge | Not documented. Honeycomb Intelligence runs in Honeycomb's own service, with no published option to point it at your own model provider |
| Can it monitor LLM apps and coding agents? | Yes. Agents Monitoring reports conversations, tokens, cost, latency, and success rate from OpenTelemetry GenAI telemetry, with a separate view for coding agents | For LLM apps, yes: Agent Timeline lays out a conversation across traces, with LLM calls, tool calls, failures, and token use, on every plan. Honeycomb documents no view of what your coding agents do and cost |
| Is trace data sampled? | No. Every trace is kept. You decide what to drop or aggregate, with rules you preview against live data | Usually, and it is the main cost lever. Refinery does trace-aware tail-based sampling so you keep every error trace and a fraction of the rest. It is a cluster your team sizes and runs |
| Can I filter telemetry before I pay to store it? | Yes, from one place. The Control Plane drops, remaps, redacts, forwards, and aggregates metrics, logs, and traces at ingest, from the UI | Yes, in several places you operate. Refinery samples, collector processors filter, and the Telemetry Pipeline add-on on Enterprise filters, samples, redacts, and extracts metrics from logs |
| Does it cover metrics, logs, and traces equally? | Yes. Metrics, logs, and traces are first-class, with PromQL and LogQL over them and APM built from the same OpenTelemetry data | Traces and wide events are the centre of the product. Logs are stored as events against the same allotment, and metrics became generally available in March 2026 |
| How is it priced? | One licence. The infrastructure runs on your own cloud bill, at your negotiated rates | Published plans, billed per event, where every span and log record counts and a 150-span trace is 150 events. A free plan and a published Pro rate, with unlimited seats, and a quote above the Pro ceiling |
| Can it run in our own cloud account? | Yes. Last9 runs in your own cloud account, on AWS or GCP in any region, and manages it for you | Yes. Private Cloud runs the same platform in your own AWS account, operated by Honeycomb or by your team, and is listed on the Enterprise plan |
Differences
Where Last9 and Honeycomb differ
Three places the two platforms take a different approach, with the tradeoffs stated.
An AI agent that investigates, on your model
-
One assistant in the app, in Slack, and in your IDE through MCP. It investigates, builds a diagnosis dashboard, and hands you the link.
-
The production agent, Last9, responds to pages with context from Slack, Linear, and PagerDuty. It runs on your model provider or inference endpoint and picks the model for each step. Customer data is never used for training.
-
An agent can do more than read. Through MCP it can add a drop rule and stop ingesting the noisy source it found.
Honeycomb
Honeycomb has invested heavily here. Its MCP server is broad, with write tools for boards, triggers, SLOs, and recipients, and it can hand a question to Canvas. Agent Skills teach Claude Code and Cursor how to instrument and investigate, and Agent Timeline shows conversations, tool calls, and token use.
Tradeoff
The difference is scope and where it runs, not whether either has an agent. Honeycomb's write tools stop at boards, triggers, SLOs, and recipients; Last9's reach drop rules, so an agent can change what is collected. Several of Honeycomb's newest AI features are early access. Compare both on your own incidents.
Every trace, instead of a sample of them
-
Last9 applies no sampling. Every trace stays available for an investigation, and you decide what to drop or aggregate with rules you write and preview against live data.
-
Drop, remap, redact, forward, and aggregate metrics, logs, and traces at ingest, from the UI, in one place rather than a separate mechanism per signal.
Honeycomb
Refinery is open source, trace-aware, and good at what it does: keep every error trace, keep a fraction of the healthy ones, and weight query results by sample rate so counts stay representative. Events sampled out before they arrive do not count against your allotment.
Tradeoff
Sampling is a real answer to volume, and weighting keeps the aggregates honest. The cost lands later, on the odd request nobody predicted: a trace sampled away is not there when you go looking for it. Refinery is also a cluster your team sizes, runs, and tunes.
High cardinality, counted differently
-
Labels like customer_id or tenant stay on the metric, with no hard cardinality limit and no charge per label or metric name. 20M series per metric per day by default, with higher limits on request.
Shared
This is not a row where one of us is built for high cardinality and the other is not. Honeycomb was founded on the idea, states no cap on unique values per column, and counts metrics as data points rather than attribute combinations. Both platforms expect per-customer and per-tenant labels.
Honeycomb
Events-based metrics work differently from time series: each distinct combination of attributes produces more events, and every event counts toward the monthly allotment.
Matters most for
What differs is the unit you pay in. Honeycomb bills by event, so span depth and request rate set the tier. Work out your events per month before comparing either platform on price.
Your cloud, and one licence
-
Last9 runs single-tenant in your own cloud account, on AWS or GCP in any region, and manages it for you, including the AI. Deployment, upgrades, and on-call are ours.
-
Every feature is on one licence: distributed tracing, profiling, custom metrics, custom dashboards, alerting, and RBAC, with forward deployed engineers on your account.
Honeycomb
Private Cloud runs the same platform in your own AWS account, either operated by Honeycomb or by your team, single-tenant or multi-tenant. It is listed on the Enterprise plan, and the SaaS service runs in a US or EU location.
Matters most for
This is the closest deployment match of any platform we get compared to, so it is not the reason to switch. Worth asking both vendors which regions are supported and what plan it needs.
Which fits you
Which one fits your team?
Choose Last9 if
- You want every trace kept, rather than a sampling policy deciding which ones survive
- You want metrics, logs, and traces as equals, with PromQL and LogQL over all of them
- You do not want the bill to move with how many spans each request produces
- You want an AI agent that investigates on your own model provider, and an MCP server whose write access reaches drop rules
- You want one place to drop, remap, and aggregate at ingest, without running a sampling proxy yourself
Choose Honeycomb if
- Your work centres on exploring wide events and traces, and BubbleUp is how your team debugs
- You want a free plan and a published rate you can model before talking to anyone
- You already run Refinery, or you want trace-aware tail-based sampling that keeps every error trace
- Agent Timeline, Agent Skills, and write tools for boards, triggers, and SLOs cover what you need from AI
Migration
Moving off Honeycomb
What the switch involves, and what Last9 does for you.
Forward deployed engineers
A Last9 engineer runs the move with your team, including your boards, triggers, and SLOs.
Talk to an engineerKeep your instrumentation
Honeycomb is OpenTelemetry-native and so is Last9, so services already sending OTLP need only a new exporter endpoint. Services still on the older Beeline libraries move to OpenTelemetry first.
See the integrationsBoards and triggers get rebuilt
Queries written in Honeycomb's Query Builder do not convert mechanically, because the data model is a different shape. Boards, triggers, and SLOs are rebuilt as PromQL and LogQL, and the Last9 agent does that work with your team.
Run both
Add a second OTLP exporter and send the same data to both platforms while you validate. Keep Refinery in front of Honeycomb until the cutover, then retire it, and compare what full-fidelity traces look like next to the sampled ones.
Objections
The objections we hear most
Each one is a fair reason to hesitate.
Honeycomb pioneered high-cardinality observability. Is that not its strength over Last9?
It is a genuine strength, and this is not a row where one of us handles high cardinality and the other does not. Honeycomb states no cap on unique values per column and counts metrics as data points rather than attribute combinations. Last9 supports 20M series per metric per day by default, with higher limits on request. The difference is the unit you pay in and how much raw data you keep, not whether the labels work.
BubbleUp and the Query Builder are excellent for finding what is different about a slow request.
They are, and teams who spend their day exploring traces rate them highly. Last9 takes a different route: metrics, logs, and traces together, with PromQL and LogQL, a query builder, and an agent that investigates in the app, in Slack, and in your IDE. Which fits depends on whether your debugging starts from wide events or from metrics and logs side by side.
Honeycomb has a free plan and publishes its Pro rate. Last9 does not.
That is true and worth weighing. Honeycomb bills per event, so a 150-span trace is 150 events, and request rate and span depth decide your tier rather than GB. Above the Pro ceiling it is a quote, as Last9 is. Work out your monthly event count first, because that number, not the rate card, is what decides the comparison.
Honeycomb Private Cloud already runs in our own AWS account.
It does, and this is the closest deployment match of anything we are compared to. Honeycomb runs the same platform there, operated by its team or yours. Last9 runs single-tenant in your own account on AWS or GCP and manages it, including the AI. Honeycomb lists Private Cloud on its Enterprise plan, so ask both vendors what tier and which regions apply to you.
Is Last9 proven at our scale?
Last9 regularly handles 500M+ events a minute and 60M+ concurrent users, and has monitored 12 of the world's 20 largest streaming events. It is SOC 2 Type II certified and PCI ready, with SSO and access controls. Teams across fintech, healthcare, gaming, media, and commerce run it in production, including Replit, Pine Labs, Tata 1mg, and Housing.com.
67%
Average reduction in observability costs
2.5x
More telemetry retained without sampling
100+
Native integrations, OpenTelemetry-first
24x7
Support from forward deployed engineers
Customers
What do Last9 customers say?
With Last9, we just eliminated the toil. No more worrying about failing dashboards or alerts — we can finally use our metrics. It just works.
Last9 allowed us to offload the operational overhead of scaling so we could focus on business alerting rather than infrastructure management.
Frequently asked questions
Ask AI about this comparison
Do not take our word for it. Send the question to an assistant and check what it finds.
Can't find the answer to your question?
Talk to our support team for help.
Start observing for free. No lock-in.
OpenTelemetry · Prometheus
Just update your config. Start seeing data on Last9 in seconds.
Datadog · New Relic · Others
We've got you covered. Bring over your dashboards & alerts in one click.
Built on Open Standards
100+ integrations. OTel native, works with your existing stack.
