Evaluation Cost Calculator

Estimate the tokens and cost of each DynamoEval evaluation type before you run it, then compare starting configurations for low, balanced, and high-risk use cases. These are estimates, not billing.

Evaluation

System Policy CompliancePolicy JailbreakStatic JailbreakAdaptive Jailbreak

Volume

Policies evaluated (1 to 5)

Compliant prompts (50 to 250)

Non-compliant prompts (50 to 250)

Perturbation types (1 to 4)

Variations per prompt (1 to 10)

Rewording is one of the types

Non-compliant prompts are split across the policies you select. Compliant prompts are sent to the target model once for each policy.

Models and Reasoning

Evaluation model (judge, generation, helper, attacker)GPT-6 Luna ($0.1 in / $0.5 out)GPT-5 mini ($0.25 in / $2 out)GPT-4.1 mini ($0.4 in / $1.6 out)GPT-5.4 mini ($0.75 in / $4.5 out)GPT-5.2 ($1.75 in / $14 out)GPT-5.6 Terra ($2 in / $12 out)GPT-4.1 ($2 in / $8 out)GPT-6.1 Sol ($2 in / $10 out)GPT-6 Astra ($10 in / $50 out)Custom prices

Reasoning budget per call on the evaluation model (a cap, in tokens)

Target model (the model being evaluated)Same model and reasoning as the evaluation modelA different modelDo not include (hosted elsewhere)

Estimated share of the budget actually used: 100%

The budget is a cap you set: the most hidden reasoning tokens one call can use. Enter 0 for reasoning off. DynamoEval reserves 2,500, 4,000, 8,000, or 16,000 for minimal, low, medium, and high effort.

The share is an estimate, not a setting. The model decides how much of the budget each call uses, and you cannot set it. Reasoning tokens billed = calls × budget × share. At 100% every call is assumed to use its full budget, which is an upper bound for capped calls. Judge and helper calls have no cap, so they could exceed it. After one measured run, enter the average reasoning used divided by the budget.

Assumptions

Low Expected High

These values have not been measured. Edit any of them below.

Edit assumptions

Base prompt length (tokens)

Policy text length (tokens)

Target response length (tokens)

System prompt on target (tokens)

Judge verdict length (tokens)

Generation output per call (tokens)

Generation passes

Judge checks re-asked (%)

System Policy Compliance

$17

Low to High assumptions: $11 to $26

Evaluation model

$12

Target model

$4.72

Calls

3,310

Input tokens

6.4M

Output tokens

397k

Reasoning tokens

0

Role	Calls	Input	Output	Reasoning	Cost
Target model	1,255	126k	321k	0	$4.72
Refusal judge	1,255	5.7M	13k	0	$10
Prompt generation	549	499k	38k	0	$1.41
Rewording	251	35k	25k	0	$0.41
Total	3,310	6.4M	397k	0	$17

• These are estimates, not billing. Prompt caching and batch discounts are not modeled, and results sit within about 10% of the cost guide at the same settings.

System Policy Compliance$17

Pricing Basis

List prices are USD per 1M tokens from OpenAI's model and pricing pages as of October 1, 2026. Azure Global Standard deployments match them, and Azure Data Zone deployments cost 10% to 20% more.

Figures come from the DynamoEval cost guide and match its tables within about 10% at the same settings. Values marked as assumptions have not been measured, so run one small evaluation and compare its token usage before relying on a larger estimate.