DigitalOcean Agent Droplets are discounted prepaid balances, not flat unlimited-agent plans. The $50 Pro fee funds a $50 monthly subscription balance and discounts eligible Managed Agents usage by 15%. The $200 Team fee funds a $200 balance and discounts eligible usage by 20%. That lets Pro cover up to $58.82 and Team cover up to $250 of eligible usage at standard list rates. After the balance runs out, usage continues at standard rates from other available credits or prepaid balances.
For a small business, the practical choice is simple: use direct pay-as-you-go funding when usage is light, irregular, or dominated by Claude or GPT. Consider Pro when eligible monthly usage reliably exceeds about $50, and Team when it reliably exceeds about $200. Neither plan eliminates metered charges for compute, memory, storage, egress, tools, or inference.
DigitalOcean Managed Agents and Agent Droplets are in public preview. Treat them as an engineering platform with change risk, not a finished no-code employee service.
Agent Droplets fund metered Managed Agents usage. They do not turn the underlying meters into unlimited service.
DigitalOcean Agent Droplets pricing at a glance
The paid plans exchange a monthly commitment for a limited discount on eligible usage. Prices and rules below were verified from DigitalOcean documentation and preview terms on October 22, 2026.
| Funding option | Upfront amount | Discount | Best fit | Important catch |
|---|---|---|---|---|
| New-customer credit | $5 promotional credit | None | A first session or tiny proof of concept | New DigitalOcean customers only; it is funding, not a recurring free tier |
| Pro | $50/month | 15% on eligible usage while the subscription balance remains | Steady eligible usage above roughly $50/month | Covers at most $58.82 of eligible list-price usage; unused balance generally expires |
| Team | $200/month | 20% on eligible usage while the subscription balance remains | Steady eligible usage above roughly $200/month | Covers at most $250 of eligible list-price usage; unused balance generally expires |
| Inference and Agents balance | Add funds as needed | None | Irregular or uncertain usage | No subscription discount, but funds do not expire and other DigitalOcean products do not draw from it |
| Account prepayment balance | Add funds as needed | None | One shared prepaid pool across DigitalOcean | Other DigitalOcean products can reduce the money available to agents |
The Pro plan's maximum monthly saving is $8.82: $58.82 of list-price eligible usage costs $50 after the 15% discount. Team's maximum monthly saving is $50: $250 of list-price eligible usage costs $200 after the 20% discount. Buying a plan below its cash break-even point does not create savings; it creates an expiring balance.
Plan fees are billed in advance, exclude tax, renew automatically, and are non-refundable under the preview terms. Unused subscription balance normally does not roll over. DigitalOcean documents two exceptions: unused balance from the first billing month can roll into the second, and a remaining balance carries forward when upgrading mid-month.
What the Pro and Team fees actually buy
Pro and Team change the payment rate for eligible resources; they do not add agents, seats, compute, models, or tools. DigitalOcean explicitly describes Agent Droplets as a pricing and billing layer on Managed Agents.
Both plans advertise unlimited agents and seats, but that language is subject to fair use. DigitalOcean may impose dynamic usage, rate, throttling, or concurrency limits. Harness Runtime also documents a team limit of up to 100 simultaneous sessions depending on resource tier and account history. “Unlimited agents” therefore means there is no per-agent license meter—not unlimited concurrent compute or unlimited consumption.
The subscription balance pays for Managed Agents usage in this order:
- Agent Droplets subscription balance;
- promotional credits;
- Inference and Agents balance;
- account prepayment balance.
Eligible charges are discounted only while subscription balance remains. Other Managed Agents charges, including third-party model inference, consume the balance at standard rates. That matters because a Claude- or GPT-heavy workload can drain the prepaid balance without receiving the 15% or 20% model discount.
The full Managed Agents cost stack
A realistic budget needs separate lines for runtime, storage, network, tools, and model inference. The plan fee is only the funding source.
| Cost layer | Current list rate or billing rule | Discounted by Agent Droplets? |
|---|---|---|
| Harness Runtime CPU | $0.044 per vCPU-hour of actual CPU consumed | Yes |
| Harness Runtime memory | $0.0095 per GB-hour allocated over time | Yes |
| Session storage | $0.05 per GiB-month held over time | Yes |
| Public internet egress | $0.01 per GiB | Yes |
| Snapshots and checkpoints | $0.05 per GiB-month retained | Yes |
| Custom sandbox templates | $0.05 per GiB-month of stored image size | Yes |
| Standard Action Gateway invocation | $0.10 per 1,000 calls | Yes |
| Exa web search through Action Gateway | $10.10 per 1,000 calls, including gateway invocation | Yes |
| Exa web fetch through Action Gateway | $3.10 per 1,000 calls, including gateway invocation | Yes |
| DigitalOcean-hosted model inference | Model-specific token rates | Yes |
| Third-party Claude or GPT inference | Current standard model rate | No |
| External APIs, SaaS plans, review labor, engineering | Provider or internal cost | No |
Harness Runtime bills active compute per second. CPU cost can fall to zero while an agent waits and consumes no CPU, but memory remains a time-based allocation charge while the session is active. DigitalOcean publishes full-allocation hourly reference prices of $0.0535 for XSmall, $0.107 for Small, $0.126 for the default Medium, $0.252 for Large, and $1.008 for XLarge. Actual CPU charges can be lower than those full-utilization examples.
Storage deserves its own alarm. Session volumes, checkpoints, snapshots, and custom templates are billed over time. Retained checkpoints continue accruing storage charges while a session is paused—even if prepaid balances are empty.
Which model charges receive the discount
The model boundary is hosting, not brand prestige. DigitalOcean-hosted models are eligible for the Agent Droplets discount; third-party commercial models are not.
Current eligible examples named by DigitalOcean include Kimi K3, GLM-5.3, and TypeSafe Jev. The live model catalog listed these standard rates on October 22, 2026:
| DigitalOcean-hosted model | Input / 1M tokens | Output / 1M tokens | Cached input / 1M tokens |
|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | $0.30 |
| GLM-5.3 | $1.40 | $4.40 | $0.26 |
| GLM-5.3 Flash | $0.15 | $0.50 | $0.03 |
TypeSafe Jev was listed at $0.042 per 1 million tokens through Serverless Inference, with no output price shown in the pricing table. A Jev call made through Action Gateway also incurs the $0.10-per-1,000 gateway invocation charge.
Claude and GPT remain standard-rate usage, including when accessed through DigitalOcean Serverless Inference. For example, the catalog listed Claude Sonnet 5.5 at $2 input and $10 output per million tokens, but Agent Droplets does not reduce those rates. If you bring your own Anthropic or OpenAI key, that external provider bills you directly instead of DigitalOcean.
The discount applies to eligible DigitalOcean-hosted infrastructure, tools, and models. Claude and GPT stay on the standard-rate path.
Model availability and eligibility can change during public preview. Pin the model ID in production, monitor the model support policy, and recheck the live pricing table before committing to a monthly plan.
What happens when the subscription balance runs out
The discount ends before the workload necessarily stops. When the Agent Droplets balance is exhausted, Managed Agents usage continues at standard rates against promotional credits, then the Inference and Agents balance, then account prepayment.
If every applicable balance reaches zero, DigitalOcean pauses active agent sessions and preserves completed work in a snapshot until funds become available. Some Action Gateway tools that do not require prepayment may not be blocked by the gateway balance check, but prepaid runtime, inference, and paid tools are interrupted.
This is safer than quietly creating an unlimited invoice, but it is not a complete cost-control system:
- there is no per-session or per-product spend limit for Harness Runtime or Action Gateway;
- a looping or compromised agent can drain prepaid balances quickly;
- account auto-reload can turn an interruption guard into an automatic spending path;
- other DigitalOcean resources can drain the general account prepayment balance;
- snapshots and checkpoints can continue accruing storage after compute pauses.
Use a dedicated Inference and Agents balance when you want agent costs isolated from ordinary DigitalOcean resources. Use auto-reload only with a deliberately low refill amount and an alert below the reload threshold.
Worked small-business cost scenarios
Cost per accepted outcome is more useful than cost per session or agent. These scenarios are models based on published list rates—not hands-on benchmarks, quotes, or promises about workload performance. They exclude tax, engineering, review labor, external SaaS charges, retries beyond the stated assumptions, and negotiated pricing.
Document processing: pay as you go is cheaper
Assume a monthly document workflow uses GLM-5.3 Flash, 20 million input tokens, 4 million output tokens, 120 default-Medium full-allocation hours, 50 GiB-month of session storage, 10 GiB-month of checkpoints, 20 GiB of egress, and 20,000 standard tool calls.
| Component | Modeled cost |
|---|---|
| Model inference | $5.00 |
| Medium runtime | $15.12 |
| Session storage and checkpoints | $3.00 |
| Egress | $0.20 |
| Standard Action Gateway calls | $2.00 |
| Pay-as-you-go total | $25.32 |
If 12,000 documents enter the workflow and reviewers accept 10,800, the cost is $0.0023 per accepted document. Pro would discount the metered usage to $21.52 but still charge the $50 monthly fee, making the effective cash cost $0.0046 per accepted document. At this volume, direct funding is the better buy.
Recurring web research: Team can beat direct funding
Assume 40 monthly research briefs use Kimi K3, 25 million input tokens, 4 million output tokens, 300 Medium full-allocation hours, 5,000 web searches, 10,000 web fetches, 100 GiB-month of session storage, 50 GiB-month of checkpoints, and 50 GiB of egress.
| Component | Standard-rate cost |
|---|---|
| Model inference | $135.00 |
| Medium runtime | $37.80 |
| Exa searches and fetches | $81.50 |
| Session storage and checkpoints | $7.50 |
| Egress | $0.50 |
| Eligible list-price usage | $262.30 |
Direct pay as you go costs $262.30. Pro saves at most $8.82, producing a $253.48 total. Team's $200 balance covers the first $250 of eligible list-price usage after the 20% discount; the remaining $12.30 bills at standard rates, for $212.30 total. If 36 of 40 briefs are accepted, the cost is $5.90 per accepted brief on Team versus $7.29 with direct funding.
The scenario assumes every item is eligible. Adding Claude or GPT changes the balance drawdown and reduces the realized benefit.
Coding and operations agent: third-party inference dominates
Assume a coding and operations agent uses Claude Sonnet 5.5 through DigitalOcean for 50 million input and 10 million output tokens. It also uses 250 Large full-allocation hours, 100 GiB-month of session storage, 50 GiB-month of checkpoints, 100 GiB egress, 50,000 standard tool calls, 2,000 searches, and 5,000 fetches.
| Component | Modeled cost |
|---|---|
| Claude Sonnet 5.5 inference at standard rate | $200.00 |
| Large runtime | $63.00 |
| Storage and checkpoints | $7.50 |
| Egress | $1.00 |
| Standard tools, search, and fetch | $40.70 |
| Pay-as-you-go total | $312.20 |
If reviewers accept 360 of 400 completed tasks, the cost is $0.87 per accepted task before labor. Only $112.20 of the modeled total is eligible for a plan discount. Because the $200 of Claude usage can consume subscription balance at standard rates, a Team plan does not guarantee a 20% reduction on the whole bill—or even on all eligible usage. Separate usage by model and funding source before buying the plan.
When Agent Droplets fit—and when they do not
Agent Droplets fit steady Managed Agents workloads that use DigitalOcean-hosted resources and benefit from pause, resume, checkpoints, and governed tools. They are less compelling when inference is mainly Claude or GPT, usage is sporadic, or the team does not want preview risk.
Choose Pro or Team when:
- eligible monthly usage consistently clears the plan fee;
- one team wants predictable prepaid funding and modest savings;
- work needs persistent sessions, sandboxed code, browser activity, or Action Gateway;
- the organization can operate budgets, approvals, monitoring, and rollback;
- DigitalOcean-hosted models are acceptable for a meaningful share of inference.
Stay pay as you go when:
- usage changes sharply month to month;
- an early pilot costs less than $50 in eligible usage;
- Claude, GPT, or external APIs dominate the bill;
- unused monthly balance is likely to expire;
- the team needs clean measurement before choosing a commitment.
Avoid the platform for now when a preview service without an SLA is unacceptable, private VPC connectivity is mandatory, or the organization lacks engineering ownership for agent permissions and failures.
Agent Droplets versus self-managed and modular alternatives
The right alternative depends on which operating burden you want to own. Vendor comparison claims were not treated as independent fact; the table below uses current first-party pricing and product documentation plus editorial fit judgment.
| Option | Better when | Cost posture | Main tradeoff |
|---|---|---|---|
| DigitalOcean Agent Droplets | You want Managed Agents with persistent sandboxes, governed tools, and a modest eligible-usage discount | $50 or $200 prepaid commitment plus metered usage | Public-preview terms, expiring balance, and no discount on Claude/GPT |
| Managed Agents pay as you go | You are measuring a pilot or usage is irregular | Metered list rates funded from non-expiring agent balance | No 15% or 20% discount |
| Self-managed DigitalOcean Droplet | A stable always-on worker is simple enough for your team to operate | Bundled VMs start at $4/month; v5 resources bill per second without a monthly cap | You own patching, isolation, scaling, queues, secrets, observability, and recovery |
| AWS Bedrock AgentCore | Your stack is already AWS-heavy or you want independently priced runtime, gateway, memory, browser, code interpreter, identity, and observability | Consumption-based with no minimum commitment; model and related AWS charges are separate | More services and meters to design, govern, and reconcile |
| Separate sandbox, model, and tool vendors | You need best-of-breed components or portability | Each provider bills its own native unit | Integration, identity, telemetry, and incident ownership return to your team |
A $4 self-managed Droplet is not a like-for-like substitute for an isolated managed sandbox. It is simply cheap VM capacity. Compare total ownership—including security updates, queue recovery, monitoring, backups, and engineer time—not the smallest server sticker price.
AWS AgentCore is also not one comparable flat agent price. Its modules have separate consumption rates, while foundation-model charges and services such as CloudWatch can sit beside them. Evaluate one identical workload across both platforms instead of copying either vendor's savings claims.
Pause and resume can reduce compute—not storage
Pausing is useful when an agent waits for a human or external event, but it is not a complete stop-billing button. Paused sessions stop compute charges while preserving workspace state and conversation history. Session volumes, retained snapshots, and checkpoints can continue accruing storage charges.
Harness Runtime defaults to a 15-minute idle suspend, while the LangGraph guest has a 10-minute session idle timeout. Each session has a monthly active-compute cap of 744 hours. Use shorter idle suspension for approval-heavy work, then measure whether cold resumes affect task latency or failure rate.
For recurring workflows:
- pause immediately when waiting for human approval;
- remove obsolete checkpoints on a retention schedule;
- delete finished sessions after preserving required evidence;
- alert on active hours, allocated memory time, storage growth, and balance drawdown;
- separate long-lived state from disposable execution sandboxes.
Least privilege, budget alarms, and rollback
A safe small-business pilot starts with reversible work, read-only access, and a hard human boundary around consequential actions. Harness Runtime supports default tool actions of allow, ask, or deny, explicit shell and file rules, policy validation, injected secrets, and outbound host allowlists.
Use these controls from day one:
- grant read-only credentials first and issue separate credentials per environment;
- deny unneeded shell commands, file paths, tools, and network hosts;
- require human approval for customer messages, payments, deletions, account changes, production deployments, and public publishing;
- set spend alerts below the Pro or Team balance, below every auto-reload threshold, and on unusual tool-call velocity;
- cancel runs that repeat the same tool or error beyond a defined count;
- checkpoint before consequential changes, then test rollback before production;
- record model ID, environment version, tool calls, balances, accepted outcome, reviewer override, and final state for every run.
A production agent needs a control loop: restrict access, watch spend, require approval, pause safely, and prove rollback.
Do not rely on balance exhaustion as the primary kill switch. A prepaid balance limits one funding path, but auto-reload, promotional credits, and account prepayment can keep usage running. Application-level run limits and explicit cancellation logic are still necessary.
Public-preview risk changes the buying decision
Agent Droplets should be treated as a reversible pilot purchase until the preview matures. DigitalOcean's preview terms allow it to change plan tiers, fees, discount rates, eligible resources, eligible models, or the balance mechanic. The preview is provided as-is and as-available, with no SLA and no service credits.
Current product limitations include no webhook notifications for run lifecycle events, no live preview of a running app, no private VPC resource connectivity or VPC peering, and no VPC-scoped model keys. DigitalOcean may also apply fair-use throttling or concurrency limits.
Before production, define an exit path:
- keep agent instructions, tool schemas, evaluation data, and source code outside a single hosted session;
- make artifacts exportable from session storage;
- document how to move the workload to a self-managed Droplet, AWS AgentCore, or another sandbox provider;
- retain a deterministic or human fallback for critical work;
- review preview terms and model eligibility before every renewal.
Verdict: useful discount, not unlimited-agent pricing
DigitalOcean Agent Droplets can be a good fit for small technical teams with steady, eligible Managed Agents usage—but the plans are easy to misread. Pro is a $50 prepaid commitment with up to $8.82 in monthly savings. Team is a $200 commitment with up to $50 in savings. Neither plan includes unlimited compute, tools, storage, egress, or inference.
Start with the $5 new-customer credit when eligible, then use the Inference and Agents balance for a measured pilot. Track cost per accepted outcome and the share of usage that is actually discountable. Move to Pro only when eligible usage reliably clears $50 per month; move to Team only when it reliably clears $200 and the extra commitment will not expire unused.
If Claude or GPT dominates the workload, price those tokens at standard rates first. The plan's headline percentage is not the percentage reduction on your total bill.
Methodology and limitations
This guide is a vendor-document-verified editorial synthesis, not a hands-on benchmark. DigitalOcean plan fees, discounts, balance order, expiry rules, eligible resources, runtime rates, model rates, tool-call rates, preview status, fair-use language, and limits were checked against current official documentation and legal terms on October 22, 2026. AWS AgentCore and self-managed Droplet comparisons were checked against their current official pricing pages. Vendor comparisons and workload estimates were not treated as independent performance proof.
The worked examples are transparent models built from published list rates. Actual CPU consumption, memory allocation, token mix, cache use, tool loops, session retention, acceptance rate, tax, engineering time, and negotiated pricing will change real costs. Google Search Console and GA4 evidence was unavailable, so this guide makes no first-party traffic, conversion, or audience-performance claims.
Primary sources:
- DigitalOcean Harness Runtime pricing
- How to pay for Harness Runtime
- Agent Droplets pricing
- Agent Droplets public-preview terms
- Harness Runtime limits
- Harness Runtime features
- Action Gateway pricing
- DigitalOcean Inference pricing
- DigitalOcean Droplet pricing
- Amazon Bedrock AgentCore pricing
FAQ
Is DigitalOcean Agent Droplets pricing unlimited?
No. Pro and Team allow unlimited agents and seats subject to fair use, but compute, memory, storage, egress, tools, and inference remain metered. The monthly fee funds a balance and discounts eligible usage until that balance is exhausted.
How much usage does the $50 Pro plan include?
The $50 subscription balance covers up to $58.82 of eligible usage measured at standard list prices because that usage receives a 15% discount. Third-party model usage consumes balance at standard rates and can reduce the eligible usage covered.
How much usage does the $200 Team plan include?
The $200 subscription balance covers up to $250 of eligible usage at standard list prices because eligible usage receives a 20% discount. It is not a $200 unlimited-use bundle.
Do Pro and Team discount Claude or GPT?
No. Anthropic Claude and OpenAI GPT are third-party models and remain at standard rates, including through DigitalOcean Serverless Inference. DigitalOcean-hosted models such as Kimi K3 and GLM-5.3 are eligible while subscription balance remains.
What happens when the balance reaches zero?
Usage moves to promotional credits, then the Inference and Agents balance, then account prepayment at standard rates. When all applicable balances are empty, active sessions pause and completed work is preserved in a snapshot until funding returns.
Does unused Agent Droplets balance roll over?
Usually not. Unused subscription balance is generally forfeited at month-end. DigitalOcean documents a first-to-second billing-month rollover and carry-forward during a mid-month plan upgrade.
Is Agent Droplets production-ready?
It may support production-like pilots, but Agent Droplets and Managed Agents are in public preview. The preview has no SLA, can change, and currently has networking and lifecycle-notification limitations. Use approval gates, rollback, and a portable fallback.