Security operations is where enterprise AI agents got real budgets first, because the work was already measurable: alerts in, investigations out. Every major security vendor now sells an AI analyst that triages, investigates and responds. Ask any of them what one investigated alert costs, and you will get five different answers, because the vendors picked five different units: reserved compute capacity, opaque tokens, credits, bundled quotas, and per-execution pricing. None of them convert into each other. This post maps the meters, does the arithmetic that is actually possible, and lists what a buyer can verify before signing.

Microsoft: a meter that runs on time, not on prompts
Security Copilot pricing is capacity-based. You provision Security Compute Units (SCUs) at $4 per SCU per hour, with a minimum of one, and consumption beyond your provisioned number bills at $6 per overage SCU. The unit is a block of compute reserved for an hour, which means the clock runs whether anyone prompts or not: one provisioned SCU, left on 24/7 for a 30-day month, costs $2,880 before a single alert is looked at.
Microsoft's billing documentation illustrates consumption with teaching numbers: a typed prompt at 3.0 SCUs, an incident summary in Defender at 0.5 SCU, a promptbook at 3.7 SCUs. Microsoft does not publish per-operation rates; real consumption is only visible in a tenant's usage dashboard. The documented worked example: in an hour where a tenant with 4 provisioned SCUs consumes 7.2, it pays $16 for the provisioned capacity plus $19.20 for the 3.2 overage SCUs, $35.20 for that one hour. Overage runs 50 percent above the provisioned rate, and provisioned capacity is billed whether the hour was busy or not.
The bigger commercial move is bundling. With the July 2026 licensing update, Microsoft 365 E5 and E7 customers receive 400 SCUs per month for every 1,000 user licenses, capped at 10,000 SCUs per month. A 400-seat E5 tenant gets 160 SCUs; a 100-seat tenant gets 40. At Microsoft's illustrative 3 SCUs per prompt, the 40-SCU allowance buys about a dozen typed prompts a month, and Microsoft says use beyond the included pool will eventually be throttled with 30 days notice, with a pay-as-you-go option at $6 per SCU. For E5 shops the assistant is effectively free at the margin until it is silently not.
Google: tokens with no published exchange rate
Google Security Operations meters its autonomous agents in Security Tokens, a single consumption unit introduced for the Agentic SOC. Several details are unusual and worth reading carefully. Tokens are charged only on successful completion of an agent run; a failed investigation consumes nothing. Assistive chat and summaries consume nothing. Only the generally available autonomous agents, currently the Triage and Investigation Agent, draw down the balance.
Entitlement is tiered by contract size. Enterprise Plus and Google Unified Security customers get a daily allotment that resets at midnight and never rolls over: 10 million tokens per day under $1M annual contract value, 20 million at $1M to $5M, 60 million above $5M. Enterprise tier gets zero included tokens. Beyond the allotment you buy an annual token subscription as an add-on SKU, and Google explicitly states it cannot be bought standalone. Paid consumption only began metering on August 31, 2026.
What Google does not publish is the per-token price on a public sheet (packages themselves are quote-only), still less a tokens-per-investigation figure: the docs say consumption depends on log density and threat complexity and "the true number comes from your environment." A 10-million-daily-token allotment is impressive until you know what an investigation costs, which you can only learn from the 90-day evaluation window. For scale, Google's other token meter, Model Armor in Security Command Center, prices at $0.10 per million tokens; nothing says SecOps tokens share that rate, and buyers should assume they do not.
CrowdStrike and SentinelOne: credits nobody has priced
CrowdStrike Charlotte AI runs on AI credits. Qualifying customers can opt into 50 credits per month that renew monthly, and per the promotion terms qualification means licensing at least one of six modules (Falcon Insight XDR, Cloud Security, Exposure Management, Next-Gen SIEM, Adversary Intelligence Premium, or Falcon for IT). Anything beyond the free 50 is arranged with an account representative. As of November 2026, no public page lists a per-credit price, and the Agentic SOAR pricing page ties the credit allotment for Next-Gen SIEM customers to data ingestion volume, meaning the AI meter is priced off the SIEM meter. The Falcon platform itself does publish device tiers, $59.99 to $184.99 per device per year across Go, Pro and Enterprise, but none of the three lists Charlotte AI as included.
SentinelOne did the same dance in June 2026: it opened Purple AI Agentic Investigation, its zero-click autonomous investigation feature, to all customers and introduced Singularity Credits as a unified currency for AI work on the platform. During the complimentary trial credits are consumed but not charged, credits are purchasable after trial through partners and direct billing, and per-credit prices are not published.
Splunk and Elastic: a quota meter and the one honest meter
Splunk's Triage agent documentation shows a third model: the limit is entitlement volume, not a price line at all. The tenant runs about 200 investigations per day, each with up to 20 triage revisions, and the agent bills nothing separately beyond the Enterprise Security subscription (the fullest agentic set requires the ES Premier edition, per Splunk's product page). For a team under roughly 200 findings a day, the AI is a flat-rate bundle; above it, the meter does not raise your bill, it denies coverage.
Elastic is the only major vendor that priced its security AI the way an LLM API is priced. The serverless security pricing page lists the Elastic Managed LLM at $4.50 per million input tokens and $21 per million output tokens, plus per-execution meters for Workflows ($0.0108 after 10,000 free executions) and Agent Builder ($0.025 after 10,000 free), effective from May 1, 2026. You can estimate a bill before buying, which is more than can be said for the SCU, token and credit meters above.
The standalone AI SOC vendors: quote, per endpoint, per employee
The cross-stack AI SOC platforms sell autonomy directly rather than bundling it into a SIEM or EDR: Dropzone sizes pricing on employee count, Intezer on endpoints, and most others publish nothing. Vendor-side category estimates run from $50K to over $1M annually, a marketing page's range with a marketing page's precision.
Why the five meters do not convert

Strip the branding and the meters measure three different objects. Microsoft sells reserved capacity: you pay for compute-hours, the same abstraction as a GPU rental. Google, CrowdStrike and SentinelOne sell work-done in a unit they define internally, charged on completion, with no published exchange rate; that is closer to a toll booth than a utility. Splunk and Elastic sell entitlements and executions attached to platform consumption. Only Elastic's per-million-token price is comparable to anything in the LLM API world, and even that measures model inference, not investigations.
The practical consequence: there is currently no way to compare the cost of one triaged alert across these platforms from public information. Anyone who publishes a number like "AI triage costs $0.12 per alert" is projecting one vendor's meter onto everyone else's.
What a buyer can actually verify
Four checks survive the unit soup, and they come from the vendors' own fine print:
- Treat bundled allowances as throttles, not budgets. 400 SCUs per 1,000 E5 licenses is about 160 prompts per 400 users per month at Microsoft's illustrative rate. Ask what happens at exhaustion: Microsoft says throttle with 30 days notice; Google says included-token agents stop starting new runs but running agents complete.
- Run the eval window before committing. Google ships a 90-day consumption-evaluation period and shows simulated token profiles for pre-GA agents. CrowdStrike offers a four-week trial. Microsoft's usage dashboard is the only source of real per-operation numbers.
- Set the ceilings on day one. Google caps paid token consumption per tenant only if paid consumption is explicitly activated per tenant, and admins can configure daily ceilings. Microsoft overage is opt-in with a set limit. Unactivated protections are the most expensive default in this market.
- Refuse illustrative math in forecasts. The 3 SCU prompt and the 0.5 SCU summary are teaching scenarios in Microsoft's own documentation. A business case built on vendor examples will not match the first invoice, which is precisely why every meter here resets, ratchets, or expires: none of these units roll over.
The agentic SOC is the enterprise AI category with the most honest use case and the least honest price lists. The meters are the product for now, and the only way to shop them is inside a trial.
Sources
- Microsoft Security Copilot pricing page
- Microsoft Learn: Security Compute Units and capacity
- 2Data: Security Copilot SCU model after the July 2026 E5 change
- Google Cloud documentation: Agentic SOC Security Tokens pricing and billing
- Google Security Command Center pricing (Model Armor token rates)
- VirtualMetric: Google SecOps pricing guide
- CrowdStrike Charlotte AI product page and Agentic SOAR pricing page
- CrowdStrike AgentWorks hackathon terms (free-credit promotion eligibility)
- Flamingo.run: Charlotte AI credits, MCP and pricing breakdown
- SentinelOne press release, June 17, 2026: Purple AI Agentic Investigation and Singularity Credits
- MSSP Alert: Singularity Credits coverage
- Splunk help: Setting up the Triage and Response agents
- Splunk Enterprise Security product page (Essentials vs Premier)
- Elastic Cloud Serverless pricing for Elastic Security
- Cotool: Built-in security AI vs AI SOC platforms comparison
- Exaforce: Top AI SOC platforms cost estimates
