LangGraph Platform vs Temporal Cloud vs Inngest: Real Durable AI Workflow Cost per 1 Million Runs in 2026
A million runs sounds like a useful shopping list. It leaves out the part that decides the bill: what happens inside each run, where your code runs, and how much history you keep. A quick agent that makes two calls is a different purchase from one that waits for approval, retries a failing tool and saves every trace.
Start by drawing one real workflow. Temporal Cloud counts server-side Actions; Inngest counts function runs and step executions; managed LangGraph deployment, now documented as LangSmith Deployment, charges for requested resources over time. You cannot turn five steps into the same five paid units across all three. There is no defensible universal “37 times cheaper” result.
Checked 1 November 2026. Prices are published USD rates before tax. The examples are independently calculated illustrations, not load tests or account quotes. Model and tool charges remain separate.
What are you actually buying?
| Route | What the platform bill covers | What still needs a budget |
|---|---|---|
| LangGraph on managed LangSmith Deployment | A hosted agent runtime and state infrastructure, with deployment resource billing and a paid access plan. | Seats, tracing, model calls, external tools and any extra deployment capacity. |
| Temporal Cloud | The managed Temporal Service, including durable workflow history and orchestration. | Your application Workers, their hosting and operations, models and tools, plus Cloud storage and support. |
| Inngest Cloud | Managed orchestration for your functions, with execution allowances and separate plan limits. | The host running your function code, models and tools, and any extra capacity or observability. |
Temporal's deployment guide puts your Workflows, Activities and Workers on your infrastructure. Inngest's hosting guide likewise calls functions wherever your app runs. An unlimited number of registered serverless workers does not make their hosting free. This is why comparing either orchestration fee with a hosted LangGraph runtime alone misses part of the purchase.
One explicit workflow, three different meters
Imagine a support assistant that reads a record, asks a model what to do, looks up a permitted answer, asks the model to draft a reply, then saves that draft. That is five operations. For the baseline, every operation succeeds on its first attempt. Each of one million monthly requests starts exactly one workflow. There are no child workflows, approval waits, timers, queries, heartbeats, schedules or extra events.
| Implementation assumed | Baseline quantity | What changes it |
|---|---|---|
| Temporal: one Workflow start and five regular Activities | 1,000,000 × (1 + 5) = 6,000,000 Actions | Retries and other server-side operations; replay itself is not another billed execution. |
Inngest: one function run and five step.run() calls | 1,000,000 × (1 + 5) = 6,000,000 executions | The actual steps, invocations and retry behavior. Do not reuse this count for a different implementation. |
| LangGraph: five nodes in a managed deployment | No fixed resource-hours follow from the run count. | Requested CPU, memory, deployment type, running hours and replicas. Measure the capacity needed. |
The two six-million counts happen to match in this deliberately simple case. They do not establish equivalent performance or reliability. Temporal's Actions reference also meters timers, signals, queries and Activity retries. Inngest's pricing calculator and FAQ explicitly count a function plus its steps. Neither promises that a real agent always fits this pattern.
Temporal: add storage and support to the Actions
For the current Pay-As-You-Go Developer route, the price page lists $50 per million Actions, no base monthly fee, and Developer support at 10% of usage. Active history costs $0.042/GB-hour; retained history costs $0.00105/GB-hour. A GB-hour means a gigabyte kept for an hour.
Our six million Actions cost $300. Before storage, adding support makes $330. For a separate illustrative storage assumption, keep an average 1 GB active and 40 GB retained throughout a 720-hour month: 720 × $0.042 + 40 × 720 × $0.00105 = $60.48. Actions plus that storage are $360.48; support adds $36.048, making $396.53 after rounding the final total. Your Worker hosting and model/tool bills still sit outside it.
The storage amounts are assumptions, not included Developer allowances or a measurement of this workflow. Payload size, completion time and retention change them. The official pricing documentation keeps Developer at $50 per million at every volume; progressive discounts apply to higher support plans. Do not apply the next tier to every Action just because you pass five million.
Business support has a $500/month minimum or 10% of consumption, whichever is greater, with its own included usage. It is a different purchase. Temporal's Developer announcement says existing self-service customers can switch in Billing Center. Check your actual plan rather than assume an older account has already moved. Trial credits can reduce an eligible first bill; they are not a recurring discount in these figures.
Inngest: six million executions, then check the plan
The current public plan table shows Pro at $99/month with one million executions, then a headline $50 per additional million. Applying that rate to all five million extra executions gives $99 + 5 × $50 = $349/month. This is a headline-rate planning scenario, not a guaranteed invoice: the FAQ describes progressive overage discounts without giving all the tier thresholds.
There is another useful wrinkle. With one million runs and five steps, the public calculator recommends Business at $499/month, which includes ten million executions. We checked that result with its other inputs left at their displayed defaults: five seats, three concurrent steps, three workers, 5,000 scores, 50,000 queue depth and 20 realtime connections. It does not display a $349 Pro quote for that selection. Confirm which plan and rate schedule your account can use before committing; the cheaper arithmetic alone does not settle it.
Pro includes 100 concurrent steps, 15 seats and seven-day trace/log history. Events, queue depth, worker count and span-data ingestion have separate allowances; extra capacity can cost more. Business changes those allowances too. “Six million executions” is only one dimension, and neither $349 nor $499 includes your external function host or model calls.
If you already use checkpointing, read the usage-limit notice. It describes an execution-counting correction effective October 20, 2026 and an October/November overage cap tied to September charges for affected accounts. That account-specific transition can change an actual bill; we have not assumed eligibility or used it as a permanent rate.
LangGraph: price a deployment, then its traces
The LangChain price page lists Plus at $39 per seat/month and includes one Small Serverless deployment. That is access to a particular hosted configuration, not a promise that it can handle a million requests. LangChain recommends Dedicated for customer-facing agents needing consistent uptime, autoscaling and high availability.
The deployment billing guide prices requested runtime CPU at 0.0675 LSU/vCPU-hour and memory at 0.009 LSU/GiB-hour; one LSU costs $1. Dedicated database CPU and memory are 0.177 and 0.025 LSU per resource-hour. Shared Serverless database, database disk and managed Redis are not separately billed there.
| GCP configuration illustration | 720-hour deployment charge | With one $39 Plus seat, before tracing and models |
|---|---|---|
| The included Small Serverless deployment | $0 within that included deployment | $39 |
| An additional billable Small Serverless: 0.8 vCPU, 2 GiB | 720 × (0.8 × $0.0675 + 2 × $0.009) = $51.84 | $90.84 |
| Dedicated Small: minimum runtime 3 vCPU/6 GiB, database 1 vCPU/4 GiB | 720 × ($0.2565 + $0.277) = $384.12 | $423.12 |
These are configuration budgets, not prices for processing one million runs. Serverless has one replica and cold starts; Dedicated adds runtime cost as it scales. Billable Serverless time continues until scale-down, not just until the last request. The documentation describes resource pricing replacing per-run/uptime pricing on October 1, 2026; check transition terms for an existing account.
Tracing can be a much bigger line than a small deployment. Plus includes 10,000 base traces per month for the organization, not per seat. At the displayed 0.005 LSU per extra trace, saving one base trace for each of one million requests adds (1,000,000 − 10,000) × $0.005 = $4,950. Add that to the one-seat Dedicated Small illustration and you reach $5,373.12 before models, tools and extra replicas. This still does not certify enough capacity for the workload.
A trace can contain multiple nested operations; five nodes do not automatically mean five separately charged traces. Sampling changes the count. Retention upgrades and evaluation features can add charges too: the tracing guide explains base and extended billing. Decide what you need to debug and retain before assuming every request must use the same tracing policy.
What happens when an API starts failing?
Suppose 100,000 Temporal Activities each need one additional attempt, with everything else unchanged. That adds 100,000 Actions: $5 in usage and $0.50 in Developer support. The storage-assumption total rises from $396.53 to $402.03. It can also add model/tool charges and Worker time; the $5.50 is only the orchestration/support difference.
For Inngest, count what its meter records for your retry and checkpointing setup. For LangGraph, measure the extra occupied resources and traces. Neither gets an invented flat “retry surcharge” here. Waiting for an approval also changes the exercise: count the actual timer/event operations, time limits and retained state instead of multiplying the successful five-step example.
How to choose without buying the wrong saving
If you already have a LangGraph agent and want its managed runtime, test a suitable deployment and record resource hours, replicas and trace volume. If you want Temporal Cloud, add the Worker infrastructure you will operate and choose the support response you need. If your app already serves event-driven functions, check whether Inngest fits that host and whether execution, concurrency and retention limits match the job.
The LangGraph library is MIT-licensed; using it in your own application is a separate route from buying managed LangSmith Deployment. A cloud seat price is not automatically a library license fee. The managed LangSmith self-hosted/hybrid offerings have their own Enterprise terms.
Self-hosting is another project budget, not a free escape hatch. Allow for databases, backups, monitoring, upgrades and engineering time. Check the exact server release license separately from an SDK's license: Inngest's current server license file includes SSPL terms and an Apache future-license notice. Do not assume every current component is simply Apache 2.0, or invent a universal run-count threshold where operating it yourself pays off.
Before signing up, run a permitted representative test containing a normal job, a failed tool call and an approval wait. Record completed jobs, latency, billable events, resources, history and model usage. Then price the same acceptable output and operating requirements. The decision is the complete bill and the work your team must own; these examples do not establish a speed, reliability or quality winner.