What Will Claude Managed Agents Cost Per Accepted Job?
Use Anthropic’s published active-runtime and Claude Sonnet 5.5 token rates to price a pilot batch. Divide by accepted jobs and compare with the same workload on your current setup.
Find the runtime you can afford
Price a pilot batch against a fixed acceptance check, such as whether a generated reconciliation report accounts for every input row. Record active session minutes across every attempt, then count each accepted job once. For an illustrative batch of 100 accepted jobs, 600 active minutes cost $0.80 at the published runtime rate. With 2 million standard input tokens and 400,000 output tokens on Claude Sonnet 5.5, model usage adds $8.00. If measured ancillary charges add $3.00, the batch comes to $11.80, or $0.118 per accepted job. Against a $12.00 benchmark, only $0.20 remains, equivalent to 150 additional active minutes at unchanged token and ancillary costs. If a retry adds substantial context or tool output, split that increase from the runtime increase before deciding what to optimize.
Plans & pricing | Claude by AnthropicMeasure the worker with the same workload
Run the same redacted workload against the cloud environment and a self-hosted sandbox when both meet your access requirements. Keep the model, prompt, acceptance check, and concurrency target fixed. Exercise the cold-start path by scheduling an identical burst after an idle period. Record submission time, first tool result, and acceptance time for each job, then tally worker compute, storage, and operator time separately. Since Anthropic still runs the model and directs tool execution in the self-hosted setup, retain observed model charges in that comparison. If a local file must stay inside your network, return a synthetic marker or redacted summary as the tool result and inspect what the session receives. A worker that passes the timing trial can still fail the data boundary if a tool echoes file contents back to the model.
Choosing an agent environment - Claude Platform DocsTrack active time per completed job
Start with Anthropic's published Managed Agents rate of $0.08 per session-hour for active runtime. A hypothetical session active for all 720 hours of a 30-day month incurs $57.60 in runtime charges before model usage. If concurrent sessions contribute active time, add their hours together. For intermittent workloads, use the active runtime recorded during the pilot to estimate the monthly total.
Record submission time, completion time, and active runtime separately. Include failed attempts and retries in total spend, then divide by the number of jobs that produced an acceptable result. Compare the estimate against provider usage or billing for the same project and time window. A gap between elapsed time and active time is expected when the workload has pauses or idle periods.
Plans & pricing | Claude by AnthropicPrice a representative batch
Anthropic's pricing table lists Claude Sonnet 5.5 base rates of $2 per million input tokens and $10 per million output tokens. In an illustrative month, assume 300 completed jobs at four active minutes, 20,000 uncached input tokens, and 2,000 output tokens each. The example uses base rates and assumes zero paid tool calls, zero retries, and zero supplemental infrastructure costs.
The jobs add up to 20 active hours, six million input tokens, and 600,000 output tokens. That yields $1.60 for runtime, $12 for input, and $6 for output, or $19.60 total, about 6.5 cents per accepted job. Under these assumptions, trimming one active minute from every job saves $0.40 across the batch. If spend is high, inspect token use before spending engineering time on small runtime reductions.
Pricing - Claude Platform DocsPrice the sandbox separately
Anthropic offers a cloud environment and a self-hosted sandbox. With the cloud environment, Anthropic supplies the sandbox. With the self-hosted option, your worker executes tool calls on infrastructure you operate, while Claude still determines the next tool call on Anthropic's side. Include worker compute, storage, and maintenance in the self-hosted estimate. Check the current Managed Agents billing terms for the selected environment before budgeting any change to Anthropic's charges.
Use the same representative tool task for both environments when both are viable. Measure whether it completes and returns the expected result, then record what each environment costs to operate. If a required internal service is unreachable from the cloud environment, price the worker against the cost of making that service available through an approved path. Check what the tool returns to the model as part of the test.
Choosing an agent environment - Claude Platform DocsReconcile the first billing window
Before scaling, reopen Anthropic's pricing table and confirm the exact model name and per-million-token units. Check which cache reads, cache writes, tool charges, and deployment settings apply to the run. Anthropic's pricing docs point out that screenshot and zoom results from browser use are billed as image input. A browser-heavy job can therefore exceed a forecast based only on the initial prompt and final answer.
If the bill exceeds the pilot estimate, choose one expensive session and reconcile its runtime, input, output, and tool usage against the forecast. Check for repeated tool results and failed attempts before changing the model. Repeat the same representative task after each adjustment and require equivalent output quality. Once the session arithmetic matches the billed categories, update the monthly estimate with the observed workload mix.
Pricing - Claude Platform Docs