720 hours of Claude agent runtime is $57.60

$57.60. That is hosted Claude Managed Agents runtime for a session that stays running through all 720 hours of a 30-day month, at the $0.08 per session-hour on Anthropic's pricing page. Idle time on that meter is $0. Anthropic's own uncached Opus 5.5 hour is $0.58, and $0.58 × 720 is $417.60. The cached hour on the same page is $0.428, and $0.428 × 720 is $308.16.

The rest of this note asks which other published clocks can even stay on for those 720 hours, and what they cost when they do. Self-hosted Claude Managed Agents, the November 12 private beta, is the case with no public runtime price. The product explanation is Self-hosted Claude Managed Agents has no public price. The 10-minute AgentCore comparison, and the V1 rate card, stay on AgentCore runtime V2 is now the default. Coding-agent seats and credits stay on A Devin ACU has no public price.

The month is one number. The clocks are not.

Seven hundred twenty hours is 30 × 24. November 2026 has 30 days. A platform that bills only while work is in flight will charge less than 720 hours for an agent that spends the night waiting. A platform that bills the allocation until you pause it will charge all 720 if you forget the pause. The table uses 720 billed hours so the rates can be compared. The assumption column says when a real session will not last that long.

Source: aiagentsinfo.com. Rates from the pages in the references, each read November 13, 2026. Runtime cells are this note's multiplication, shown in the next section. The Opus hour is 50,000 input tokens and 15,000 output tokens, uncached, repeated every hour. Token rates: Opus 5.5 at $4 and $20 per million, Gemini 4 Carbon at $2 and $10 through November 30, gpt-5.6-sol short context at $4 and $20. The Sol rate is promotional through November 30, and the page does not print December 1.

E2B's own default is a different shape, 2 vCPU and 512 MiB. At the FAQ's per-second rates that hour is $0.1089, and 720 hours is $78.41. The FAQ's rounded "about $0.109" times 720 is $78.48. The default is cheaper per hour than a 2 GiB box only because the memory is a quarter of a gibibyte. It is the wrong row if the agent needs the 2 GB the other lines use.

OpenAI's 16 GB container is $0.48 per 20-minute session. Across 2,160 of those blocks it is $1,036.80 of runtime, and $1,396.80 with the Sol token shape. The sandbox page's large size is 16 GB and 4 vCPU. The pricing page also lists 64 GB at $1.92 per 20 minutes. The sandbox size table does not have a 64 GB row, so this note does not turn that price into an Agents API sandbox.

How each runtime dollar was produced

Every runtime cell is a published rate times 720, except the OpenAI cells, which are the same rate times 2,160 blocks of 20 minutes. Those two counts are the same month. 720 hours contain 2,160 blocks of 20 minutes.

The per-minute reading of OpenAI's footnote lands on the same month. A 4 GB session at $0.12 per 20 minutes is $0.006 per minute. 720 × 60 × 0.006 is also $259.20. The 5-minute minimum would change a short session. It does not change a session that fills every minute of the month. The page still does not define which sessions are "eligible" for the per-minute reading. Both readings are on the page. For this month they agree.

Claude's cached cell is the pricing page's own split, not a guess. Forty thousand of the 50,000 input tokens are cache reads at 0.05× the $4 input price. Per hour that is $0.04 of fresh input, $0.008 of cache reads, $0.30 of output, and $0.08 of runtime. The page prints the sum, $0.428. This note multiplies the printed sum.

AgentCore's $116.21 will not appear as one line on one session. The quotas page says a microVM session's maximum duration is 8 hours, and the default idle timeout is 15 minutes of inactivity. 720 ÷ 8 is 90. An agent that is actually busy for a month is a chain of sessions, with a new microVM after each cap, and the cold start between them is not in the $116.21. CPU that drops to zero during I/O is also not in it. AWS's published support-agent example on the pricing page is the opposite shape, 60 seconds of CPU inside 10 minutes, at $0.006703. That example, and the V1 rates, are worked on the V2 runtime guide. This note does not repeat them.

Google's $74.16 has the same kind of asterisk. The Agent Platform page says that for Runtime, idle time spent waiting for the next prompt between turns is not billed. An agent that waits all night and works for an hour bills the hour. The $74.16 is the bill only when 1 vCPU and 2 GiB are the billed resources for all 720 hours. The Carbon token add-on is a second price list. gemini-4-carbon on the Gemini API is $2 and $10 per million tokens through November 30, 2026, and $4 and $20 starting December 1. Output includes thinking tokens. This note's 15,000 output tokens are the whole output line. If the model thinks past that, the invoice is higher, and the page says Carbon is not on the free tier. Thirty-six million input tokens at $2 is $72. 10.8 million output tokens at $10 is $108. Together with $74.16 that is $254.16 for a November run of this shape, and $434.16 once the December rates are in force.

Who prints no runtime hour

Four products people mean when they say "an agent that stays up" do not give you a number to multiply by 720.

Source: aiagentsinfo.com. Devin from devin.ai/pricing, HTTP 200 at 07:47 UTC on November 13, 2026. Cursor from the Cloud Agents and models docs, HTTP 200 the same minute. Codex from the help article retrieved earlier that morning. Self-hosted Claude from the pricing page.

Cursor's $136.80 is 36 million input tokens at $2 plus 10.8 million output tokens at $6. It is the list rate for Grok 4.7. The docs say Cloud Agents draw the plan's pool first. A week that still has included usage will bill less than the list rate. The $30 Premium seat is the right to start one. It is not the 720-hour meter, and Plus at $8 does not include Cloud Agents at all.

Devin's $20 Pro seat is the same kind of number. It buys a quota, and the pricing page says cost per unit of extra usage is API pricing. It does not print the rate, and it does not print an hour. Turning $20 into a cost per always-on day requires a quota size the page does not give. The billing-unit census is where the ACU gap is documented, from the docs fetch on November 12. This note does not re-derive it.

What this means if you are pricing a persistent agent

Price the clock you will actually leave running.

If the agent works in bursts and the platform bills running time, the month is the sum of the bursts. Hosted Claude at $0.08, and Google Agent Compute at $0.085 per billed vCPU-hour plus memory, are that kind of clock. A session you leave idle on Claude adds no runtime. A prompt gap on Google Agent Runtime adds no Agent Compute. The $57.60 and the $74.16 are what you pay when the clock does not stop.

If the agent sits in a sandbox you allocated, you pay until you pause or kill it. E2B is explicit. Daytona prints the same two compute rates and does not say what a stopped sandbox costs, so this note will not call Daytona free when stopped. OpenAI's hosted sandbox can disappear an hour after activity and keep-alives stop, which caps a forgotten bill, and the page does not say the keep-alives themselves are free.

If the platform kills the session on a timer, "24/7" means a harness that starts the next session. AgentCore's default is 15 minutes idle and 8 hours absolute, on microVMs. The Instances path is the one AWS's own pricing example uses for a job that runs for hours in one persistent session, and that example's box is $233.05 across 720 billed hours, on an 8 vCPU instance, before the model.

If the vendor will not print the hour, do not invent one from the seat. Self-hosted Claude, Devin, Cursor Cloud Agents, and Codex Cloud are in that group this morning. The cost calculator will price the tokens. It will not price the runtime, which is the gap this table is for.

The figures in the tables are licensed CC BY 4.0. Credit AI Agents Info if you reuse them.