Contrast vs Splunk
Splunk prices traces per host or per span and keeps raw APM traces for 8 days by default. Contrast keeps each run in your workspace's own TraceDB database and bills for the data your queries scan. Asking what happened in what order takes one line of SQL.
Updated November 2026
At a glance
- 7.8×
- faster than ClickHouse
- 90%
- lower cost than Snowflake
- 0
- hosts or spans billed
Recorded races, October 2026. Splunk wasn't in the test.
How you're billed
List prices as each vendor publishes them, checked November 2026.
Contrast
- Billed on
- AI credits, data scanned and storage
- Price
- $10 per TB scanned
Plus $1 per 100 AI credits and $0.05 per GB-month, after a free monthly allowance.
Splunk
- Billed on
- Hosts or spans
- Price
- APM from $40 per host a month
Agent Observability from $8.50 per 10,000 spans a month.
Paying per span to find out why an agent failed?
An agent run is a string of calls, errors and retries. The questions worth asking are about that order.
Contrast
Point an OpenTelemetry exporter at Contrast and each trace becomes one run, with every model and tool call, its tokens and its status.
Splunk
Splunk APM accepts OTLP, Zipkin, Jaeger and SAPM, and keeps what arrives as spans and traces.
What happens in the two steps after a tool error
TraceDB keeps each run in time order, so paths() reads the next two steps directly. In SPL you sort every run backwards and stitch the paths together with streamstats.
SELECT path, count, entities
FROM paths(from => 'tool_error', steps => 2)
ORDER BY count DESC
LIMIT 10index=agent_traces
| sort 0 run_id, -_time
| streamstats current=f last(event) AS next_1 BY run_id
| streamstats current=f last(next_1) AS next_2 BY run_id
| where event="tool_error"
| fillnull value="(end)" next_1 next_2
| eval path=event." → ".next_1." → ".next_2
| stats count dc(run_id) AS entities BY path
| sort - count | head 10Contrast and Splunk, feature by feature
| Contrast | Splunk | |
|---|---|---|
Agent traces | ||
| OTLP over HTTP from any OpenTelemetry SDK | ContrastYes. Protobuf or JSON, or a Collector. | SplunkYes. OTLP, Zipkin, Jaeger and SAPM endpoints. |
| Each trace kept as one agent run | ContrastYes. Calls, timing, tokens and outcome. | SplunkYes. Workflows mapped from request to response. |
| Token and model on every model call | ContrastYes. Recorded on each chat call. | SplunkYes. Tokenomics tracks tokens by model. |
| Runs and product events in one engine | ContrastYes. TraceDB follows the warehouse tables you name. | SplunkPartly. Traces and events live in separate products. |
Querying runs | ||
| SQL over agent traces | ContrastYes. Standard SQL in TraceDB. | SplunkPartly. SPL, Splunk's own search language. |
| Built-in path, funnel and session functions | ContrastYes. paths(), followed_by(), sessions() and funnel(). | SplunkPartly. Assembled from sort and streamstats. |
| Saved groups for follow-up questions | ContrastYes. Save an answer, then query only it. | SplunkPartly. Through saved searches and lookups. |
Agents and billing | ||
| Hosted MCP server | ContrastYes. Its engine tools read TraceDB. | SplunkYes. MCP server for Observability Cloud data. |
| No per-host or per-span pricing | ContrastYes. AI credits, data scanned and storage. | SplunkNo. Per host, or per 10,000 spans. |
| Monthly spend cap | ContrastYes. Pay as you go stops at your budget. | SplunkPartly. Ingest filtering and Platform cost controls. |
Running alongside Splunk
- A workspace owner or admin starts TraceDB from Sources, under Agent traces. It takes a few minutes.
- Create an API key on the same page and download contrast-traces.env, which any OpenTelemetry SDK reads.
- Add an otlphttp exporter for https://api.embrasure.ai/v1/traces to your OpenTelemetry Collector, next to the one that already sends to Splunk.
- Name each run, then ask about runs in Contrast's chat or through the Contrast MCP server.
More comparisons
Speedups are from TraceDB races recorded on October 27, 2026 on six event questions over 100M events, each the median of 5 races; TraceDB was also 8.1× faster than Snowflake and 8.6× faster than DuckDB. TraceDB ran on 4 vCPUs of an 8-vCPU VM in AWS us-east-1, with ClickHouse and DuckDB each on the other 4; Snowflake ran on an X-Small warehouse in us-west-2. Splunk was not part of that test. The cost figure is the one the Contrast home page states for TraceDB against Snowflake. Splunk details are from its public documentation, checked November 2026: Splunk pricing, Send APM traces, Data retention in Splunk Observability Cloud, Splunk Agent Observability, About the search language, streamstats, What is the Splunk MCP Server?, Splunk MCP server for Observability Cloud. Product names are trademarks of their owners.

