Datadog LLM Observability Pricing: $14,125 at 100M Spans vs $3,819 on Braintrust

Datadog LLM Observability pricing reaches $14,125 a month at 100M spans on the agent workload we priced, using its cheapest published rate. Braintrust Pro comes to $3,819 on the same volume. The gap comes from the meter. Datadog charges per model call, Braintrust charges per byte, and the official LangSmith and Arize pages we opened do not print an overage rate at all.

Is Datadog LLM Observability expensive? At low volume it is competitive, and at high volume it is the costliest published rate we could model. On a 10-span agent run with 4 model calls, it matches Braintrust at 1M spans and costs 3.7 times as much at 100M spans. Your payload size can reverse that result.

Datadog LLM Observability pricing meters model calls, not agent steps

Datadog's bill depends on one count: how many times your application calls a model. Its documentation states that the product is metered and billed on the number of LLM spans ingested. It defines an LLM span as a single request to an LLM provider, and warns that one agent workflow can produce several.

Datadog's list page now labels this product Agent Observability, while buyers still search for the older name. Both names point at the same meter. For a view of how much of Datadog's own revenue depends on this customer type, see the breakdown of Datadog's AI-native revenue concentration.

The list page prices four retention bands

The first 100,000 LLM spans cost $160 a month on annual billing, $200 month-to-month and $240 on-demand. A footnote adds that 40,000 LLM spans are free. Beyond the first 100,000, each further 10,000 is priced by how long you keep it.

On annual billing the rate is $3.50 at 15 days, $5 at 30 days, $6.50 at 60 days and $7.50 at 90 days. On-demand runs from $5 to $9 over those same four bands. The retention choice therefore changes the unit price of your whole volume, which matters more than any volume discount you are likely to win.

Why tool spans change the comparison

A typical agent trace is mostly not model calls. Retrieval, tool execution and orchestration all produce spans that a span-counting vendor would charge for. Datadog's documented meter counts only the LLM requests, and the list page carries no line for the other span kinds.

So "per span" undersells what Datadog is actually selling. It is closer to per-model-call pricing, and the fair comparison measures LLM spans against total spans. How much that ratio swings run to run is the subject of the analysis of agent token cost variance.

The agent workload behind every LLM observability cost figure

Every cost in this post comes from one invented but fully disclosed workload, because no vendor publishes a comparable one. The rates are real and the workload is ours. Change any assumption below and the totals move, which is exactly why they are listed.

These are the author's planning assumptions and are not measurements of any real system. Volumes below are total spans, so 1M spans means 100,000 runs, 400,000 LLM spans and 7.2 GB.

The three headline volumes sound like one number and act like four. A volume of 10M spans is 1M traces, 4M LLM spans and 72 GB of data on this workload. Each vendor reads a different one, so a quote expressed in only one unit tells you very little.

I would refuse any internal budget that states a single volume figure. Ask for the run count, the spans per run, the model calls per run and the kilobytes per run. Those four measurements convert every price list in this category.

LLM observability cost at 1M, 10M and 100M spans

Datadog and Braintrust are the only two vendors whose public pages price this workload at all three volumes. They start level and end far apart. LangSmith is shown as a formula because its trace rate is not printed, and Arize's published plan ends well before the first column.

Calculated by the author from official list rates. The term r is dollars per 1,000 base traces, which LangChain's pages do not state. The Datadog 30-day rows apply the 30-day rate only above the first 100,000 LLM spans, because the list page does not say whether the included block is uplifted. Doing so would add $15.

Cost per million spans shows the shape in one figure. Datadog annual falls from $265 at 1M to $141.25 at 100M, while Braintrust falls from $255.60 to $38.19. Volume multiplies by 100 between the two ends, Datadog's bill multiplies by 53 and Braintrust's by 15.

The reason is arithmetic. Datadog's rate per LLM span never declines on the list page, so the bill converges on a straight line. Braintrust's flat fee and its low byte rate keep its unit cost falling for longer, on this payload at least.

Where Datadog's per-span pricing explodes

Datadog's bill grows fastest when you add retention or model calls, because both act on every span you send. A flat rate looks safe in a demo and compounds under real traffic. Those two levers matter more than the choice between annual and on-demand.

Retention multiplies the entire volume

Moving this workload from 15-day to 90-day retention on annual billing takes the 100M-span bill from $14,125 to $30,085. That is 2.1 times the cost for the same traffic. Teams usually want that longer window for the reason covered in the piece on agent drift in production: slow behavioural change needs a quarter of history.

Braintrust handles the same request differently, charging $0.50 per GB per month after the first 30 days. Holding 2 extra months of a steady 720 GB intake would add about $720, for roughly $4,539. My position is that retention should be priced before the trial starts, since this is where the two meters differ most.

Model-heavy agents pay the full rate

Datadog's bill is directly proportional to LLM share above the included block. The table below holds total volume at 100M spans and changes only the mix, adjusting the Braintrust payload to match.

Calculated by the author. Scores are held at 1 million in every row, and Braintrust data overage is $3 per GB above 5 GB.

Braintrust pricing turns on bytes per run, and the bytes can flip the result

Braintrust's cost advantage survives only while your runs stay small. Its FAQ defines processed data as the total bytes ingested across logs, experiments and datasets, including inputs, outputs, prompts, metadata and spans. A context-heavy agent sends far more of those bytes than our 72 KB run.

The break-even payload

At 100M spans, Braintrust matches Datadog's 15-day annual bill at about 415 KB per run. Against Datadog's 30-day rate, the matching retention tier, the break-even rises to about 615 KB. The 1M-span crossover is far lower, at about 103 KB per run.

That 415 KB is not exotic. Holding other spans at 2 KB, it implies about 101 KB per LLM span, or roughly 25,000 tokens of prompt and completion at our 4-byte assumption. An agent replaying a long conversation or a retrieved document set can carry that today.

Scores are the second meter, and there is no ceiling

Each recorded score counts once toward the monthly allowance, and Pro charges $1.50 per 1,000 after 50,000. At 100M spans, scoring just 10% of runs costs $1,425, or 37% of the Braintrust bill. Scoring every run would take that line to about $14,925, which alone exceeds the Datadog total. Teams building an eval suite should budget that meter deliberately.

Two rules make this harder to control. Processed data is measured at ingestion, so deleting logs later does not reduce it. Braintrust's FAQ also says it does not offer hard spending limits that prevent usage. Any sampling or truncation has to happen in your code before the bytes leave.

LangSmith pricing prints the seat and omits the trace rate

LangSmith is the one vendor here where you can see the team cost but not the traffic cost. LangChain's pricing page lists Plus at $39 per seat with 10,000 base traces, then "pay-as-you-go". Its billing documentation refers readers back to that page for the excess rate.

Third-party trackers do report a rate, and they disagree on its history. Our sourcing rules bar a figure with only that support, so it is not printed here. For an overview of each tool's billing unit and workflow, see the comparison of LangSmith, Braintrust and Arize.

Where seats lose to usage pricing

Seat cost dominates at low volume. Eight Plus seats cost $312 a month, which is more than Datadog's entire 1M-span annual bill of $265 before LangSmith charges a single overage trace. One $39 seat buys the same money as 111,429 Datadog LLM spans at the annual rate.

So a large team with modest traffic should not buy on a per-seat plan unless it values the seat workflow directly. Break-even rates make the question concrete. LangSmith undercuts Datadog annual only below about $1.23 per 1,000 traces at 10M spans and $1.38 at 100M. To undercut Braintrust, it would need about $0.22 and $0.35.

Where trace pricing beats span pricing

Trace billing rewards depth. A trace is a single execution that can contain many steps, so long agent loops are cheap per step. If the same 100M spans arrive as 50-span traces, the workload is 2M traces, and the Datadog break-even rate rises to about $6.94 per 1,000.

Retention has its own trap. LangSmith's documentation says that when an automation rule matches any run in a trace, all runs in that trace are upgraded to extended retention. Extended retention is 180 days by default, and the fee is again not printed. Check this rule before you attach an automation to a broad filter.

Arize pricing stops at 50K spans on the public page

Arize publishes a cheap entry price and no usable price above it. Arize's pricing page lists AX Pro at $50 a month with 50,000 spans, 10 GB of ingestion and 30-day retention, and its usage documentation shows the same allowances. Neither page prints an overage rate.

Our workload hits that wall at once. At 1M spans the AX Pro allowance covers 5% of span volume, even though the 7.2 GB fits inside the data allowance. The two higher volumes run past both limits, at 7.2 and 72 times the 10 GB. Enterprise is custom, so this is a sales conversation at every volume in the table.

Phoenix moves the cost to your own team

The alternative route is Phoenix, which Arize describes as open-source and local-first, with self-hosting under the Elastic License 2.0. The page prints no platform price. At 100M spans you would be running storage, query capacity and on-call for 720 GB of new data each month.

I would not treat that as free. For a team that already operates an OpenTelemetry stack it may be the cheapest line here, and for a team that does not it is a hiring decision. Neither case belongs in a list-price table, which is why Phoenix is absent from ours.

Where this cost model is weakest

The largest weakness is that the workload is invented. A 10-span, 72 KB run is a plausible shape, and no public dataset confirms it is typical. If your runs are longer, heavier or more model-bound, the ranking above can change, and the sensitivity table shows it moving.

List prices are the second weakness. Anyone reaching 100M spans is likely to negotiate, and Datadog customers often buy this product inside a wider contract. A Datadog account with a large existing commitment could land well below our $14,125. Our table is a ceiling for planning, and it should not be read as a forecast.

The comparison also favours Datadog on retention, on purpose. Our headline pairs its cheapest 15-day tier with Braintrust's 30-day tier. The matching 30-day comparison is $20,110 against $3,819. We chose the conservative pairing so that the headline gap would not depend on the unfavourable reading.

Judge costs may also differ. We could not confirm from Datadog's documentation whether managed evaluation calls count as billable LLM spans. If they do, the 10% scoring assumption adds about 2.5% to Datadog's LLM span volume. Finally, feature fit is outside this model: Datadog's value for a team that already runs its APM may justify paying 3.7 times more. Before you sign, read the concessions buyers should request on usage-based contracts.

Frequently asked questions

Where to start

Instrument one production-like agent run today and write down four numbers: total spans, LLM spans, kilobytes and scores. Then multiply them into this post's table using your actual monthly run count. The ranking you get is worth more than ours, and the exercise takes an afternoon. Keep the output next to the review of AI FinOps tooling gaps, so that observability appears as its own budget line.

Next, ask LangChain and Arize for their overage rates in writing, and ask Datadog whether judge calls bill as LLM spans. A vendor that cannot answer those questions before the contract will not answer them more quickly on an invoice. Prices in this post were last verified on 8 November 2026, and we will re-check them within 90 days.

Related reading