Datadog cut its AI bill by more than $1 million a month. The default was the expensive part.
Maestro Brief · Published by Maestro Mojo
2026-09-01
Maestro’s take
The cleverest AI cost optimization may be a boring default.
Datadog says it now saves more than $1 million each month on AI. It did not find a magic prompt. It changed which model ran by default. It lowered routine reasoning effort. It reduced the context sent into each job. Then it checked whether the work still passed its own tests.
That last part matters. A cheaper failure is still a failure.
TL;DR
Datadog had Claude Code and other agent skills default to Opus 4.8.
Its internal tests found that moving those jobs to Sonnet 4.6 lowered cost by 36.7%, with an 8% drop in proficiency on Datadog’s own workflows. Datadog attributes $687,000 in monthly savings to that change.
Changing Claude Code’s default effort from high to medium saved more than $288,000 per month, according to the company.
Datadog’s internal Headroom evaluations found a 47% cost reduction without a significant performance drop. In a separate one-week pilot involving more than 1,000 engineers, average cost per user fell 27%. Datadog did not publish a quality score for that pilot.
These are Datadog’s numbers. They are not a promise for your codebase. The useful result is the method.
The simple formula
Cheaper default + medium routine effort + less context + real quality checks = a lower bill.
Why developers should care
Before polishing prompts, check which model and effort level your tools select by default.
Using the strongest setting for every job is like driving a moving truck to buy milk.
The best model can still be the right choice for a hard migration, a security review, or a bug nobody else can solve. It does not need to be the automatic choice for formatting, tests, documentation, and ordinary feature work.
Do this
- Build a small, representative set of real tasks from your own repository.
- Run them repeatedly with your current model and effort setting.
- Repeat the same tasks with a cheaper model or medium effort.
- Compare tests, review defects, completion time, and usage.
- Keep the cheaper setting as the default only when quality remains acceptable across the set.
- Continue rerunning the checks as models and tools change.
- Escalate the difficult jobs instead of escalating every job.
Datadog ran more than 140 evaluations and continued running them nightly. That repetition matters because agent results vary from run to run.
Also trim repeated logs, giant tool responses, and stale conversation history. In Datadog’s engineer pilot, input tokens fell 39.3% and output tokens fell 35.7%.
Do not do this
Do not switch every task to a cheaper model and declare victory because the bill fell.
Datadog saw an 8% proficiency drop when it moved from Opus to Sonnet. Your tasks may react differently. A small savings can disappear quickly if developers must redo the work.
Do not copy Datadog’s percentages into a budget forecast. Its scale, tooling, models, and work are not yours.
Do not change a team-wide default after one lucky run. Use one comparison to learn. Use a repeated task set to decide.
One thing to try today
Take one ordinary repository task. Run it once on medium effort and once on high effort.
Use the same prompt. Check the tests. Review the diff. Note the time and usage.
Treat this as a quick first check—not a policy decision. If the result looks promising, add the task to a repeated evaluation set.
The bottom line
Expensive models should earn the hard jobs. They should not inherit every job by default.
Datadog’s lesson is not “always use Sonnet.” It is simpler: measure quality and cost together, then route the work.
Published September 1, 2026. This article is reporting and Maestro analysis. Maestro’s opinion is AI-generated.