Skip to content

Ostra Decision

Verdicts for CI and agents. One model, three testing calls: why a test failed, which tests a diff affects, and whether a selector will survive.

Benchmarks

The alternative to Ostra Decision is asking a coding model the same question on every failure. This is what that costs at CI volume.

Cost to triage failures
  • Ostra Decision
  • Coding models asked the same question
Failures triaged against total cost
Model02004006008001,000
Ostra Decision$0.00$0.0001$0.0002$0.0003$0.0004$0.0005
Ostra Pro$0.00$1.55$3.10$4.64$6.19$7.74
Claude Sonnet 5.5$0.00$4.60$9.20$13.80$18.40$23.00
Claude Opus 5.5$0.00$9.20$18.40$27.60$36.80$46.00
GPT-5.6 Terra$0.00$4.72$9.44$14.16$18.88$23.60
Grok 4.7$0.00$4.36$8.72$13.08$17.44$21.80

The lines are the total cost of 0 to 1,000 failed tests. Each coding model is sent 10,000 tokens of failure context and writes about 300 tokens of answer. Ostra Decision is priced per task, not per token: $0.50 per 1M tasks, and it returns a verdict. Other models at their published API list prices, read 2026-11-19: Claude Sonnet 5.5, Claude Opus 5.5, GPT-5.6 Terra, Grok 4.7. No caching or batch discounts on either side.

Capabilities

What goes in, what it works from, the three calls it makes and what comes out.

Input

  • A failure

    Why did it fail?

  • A diff

    What does it affect?

  • A selector

    Will it survive?

Context

  • Run history

    Triage is strongest with it

Tasks

  • Triage

    Bug, flaky test, environment or test defect

  • Test impact

    Run what matters first, skip the rest

  • Selector survival

    Catch brittle selectors early

One model, three tasksServerless

Output

  • A verdict

    With a confidence

  • A deferral

    When it is not confident

Available everywhere you work

Call it from your pipeline over REST or from your agent over MCP. The same key works in Studio and the CLI.

Studio

Pick Ostra as the model and watch the agent work a failing run, step by step.

Open Studio
checkout.spec.ts failed on main · 3 of the last 10 runs
 
✓ Read the run — video, steps, network, console
✓ Memory — [data-test=pay-now] renamed in v2.4.1
● Patch the selector and re-run the spec
 
- await page.click('[data-test=pay-now]')
+ await page.click('[data-test=pay-submit]')
 
Waiting for your approval before anything touches disk.

CLI

The same agent in your terminal and, headless, in your pipeline.

Install the CLI
❯ why did checkout fail on staging last night?
 
⏺ memory.search done
⏺ runs.get_failures done
 
The selector [data-test=pay-now] was renamed to
[data-test=pay-submit] in v2.4.1. Two specs still use the old one.
 
model ostra-pro · shop-e2e

MCP

One server for Cursor, Claude Code, Copilot and Codex. Ostra shows up as tools.

Connect an agent
// .mcp.json — Cursor, Claude Code, Copilot, Codex
{
"mcpServers": {
"ostra": {
"type": "http",
"url": "https://ostra.testrelic.ai/mcp",
"headers": { "Authorization": "Bearer $OSTRA_API_KEY" }
}
}
}

API

REST with streaming. One bearer key for every model and every task.

Read the rates
$ curl https://ostra.testrelic.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OSTRA_API_KEY" \
-d '{"model":"ostra-pro","stream":true,
"messages":[{"role":"user",
"content":"Fix the flaky checkout spec"}]}'
 
data: {"choices":[{"delta":{"content":"The flake comes from"}}]}
data: {"choices":[{"delta":{"content":" a renamed selector"}}]}

Try Ostra Decision now.

Frequently asked questions

Ostra Decision is one model that makes three testing calls: why a test failed, which tests a diff affects, and whether a selector will survive the next release. It returns a verdict a pipeline or an agent can act on. It is in beta.

Through the REST API at ostra.testrelic.ai/api/v1, choosing the task with the task field on /predict, and over MCP, where each task is its own tool. The same key also works in Studio and the CLI.

Send a POST to https://ostra.testrelic.ai/api/v1/predict, choosing the task with the task field: triage, test_impact or selector_survival. Authenticate with your API key as a bearer token. Over MCP, each task is its own tool.

It takes a failure, a diff, or a selector, with whatever run history you have. It returns a verdict with a confidence, or an explicit deferral.

One. Every task goes to the same model id, ostra-lite. It runs serverless, so there is nothing to provision.

$0.50 per 1M tasks for triage, $2.00 per 1M tasks for test impact, $0.25 per 1M locators for selector survival, from a prepaid balance. A triage or a test-impact request is one task whatever its size; a selector check is billed for each locator in it, so a call with 500 locators counts as 500. Top-ups start at $5. You can pay as you go with no plan, and Ostra usage is also included in Growth, Pro and Enterprise.

No. You can pay as you go from a prepaid balance, with no plan. Growth, Pro and Enterprise also include Ostra usage, which is metered and resets automatically: Pro has 10x Growth's, and Enterprise is unlimited.

A request with no balance is refused until you add credit. You can add $5 to $999 per payment. To limit spend, each key can carry a spend cap, and the account has a monthly spend limit.

Ostra Decision is priced per task, not per token, so triaging 1,000 failures costs about $0.0005. Asking Claude Sonnet 5.5, Claude Opus 5.5, GPT-5.6 Terra and Grok 4.7 the same question, with 10,000 tokens of failure context each, costs $21.80 to $46.00 at their published list prices, and each also writes about 300 tokens of answer; Ostra Decision returns a verdict. We have not published a quality comparison against coding models.

No. It defers when it is not confident. Failure triage gives a verdict on about four in ten failures and leaves the rest to a person or to Ostra Pro. Triage is strongest when a failure comes with run history; with only an error message it answers far less often.

It is in beta. Scores and behaviour change between releases; the current release is v.beta.5.

Use Ostra Pro when you need the reasoning or the fix, or want a model to write tests. Ostra Decision returns a verdict, not an explanation, so send the failure to Ostra Pro for that.