AI pricing becomes easier once you identify what the vendor is billing: a seat, tokens, a completed generation, delayed batch work, or infrastructure. Those units are not interchangeable.
The live Copilot plan snapshot above is the answer for buyers searching for a Copilot subscription. It is rendered from our maintained subscription dataset, including the current checked date, plan status, and published limits. The calculator and API table show the separate economics for software that calls models directly.
For a broader plan comparison, open the Microsoft Copilot subscription pricing page. Builders can also compare current OpenAI API pricing, Anthropic API pricing, Google AI pricing, and DeepSeek API pricing.
Disclosure: this article contains affiliate links. If you buy through them, AI Pricing Guru may earn a commission at no extra cost to you.
Start with the billing unit
| Pricing model | What is billed | Best fit | Main budget risk |
|---|---|---|---|
| Subscription | User or seat over time | Interactive work by people | Inactive seats and changing usage limits |
| Per-token API | Input, cached input, and output | Products, agents, and automation | Long context, verbose output, retries, and tool loops |
| Batch | Asynchronous API work | Backfills and offline processing | Silent scale and repeated failed jobs |
| Per-generation | Image, clip, audio minute, or another output unit | Creative and media workflows | Paying for attempts that are not accepted |
| Self-hosting | Hardware, energy, and operations | Stable high-volume inference | Low utilization and engineering overhead |
Many products combine these models. A coding assistant may sell a seat while routing requests to token-metered models. An enterprise plan may have a predictable invoice but still enforce model-specific quotas. A cloud deployment can add storage, networking, and observability to the model charge.
Write down the billing unit before comparing vendors. Otherwise a subscription allowance can be mistaken for API credit, or an open-weight model can be mistaken for free production infrastructure.
Subscription pricing: best for human workflows
Subscriptions work well when a person uses AI interactively for writing, analysis, research, design, or coding. A seat is easy to approve and predictable enough for departmental budgets.
The Copilot plans in the live table illustrate why the product name alone is not enough:
- Copilot Free is the entry point for basic consumer use.
- Microsoft 365 Premium bundles Copilot with individual productivity software and storage.
- Microsoft 365 Copilot Business is a per-seat product for organizations that need work-data grounding and administrative controls.
These are not prepaid API packages. A company building an AI feature should not convert a seat price into an assumed token allowance. Likewise, a free plan is not proof that a specific Business feature or trial entitlement is included. Verify the current plan record and provider terms for the account and region you intend to buy.
The right subscription metric is cost per active user who gets useful work done. Audit active seats, plan limits, model access, data controls, renewal terms, and whether the product fits the employee’s real workflow.
Per-token APIs: best for products and automation
API billing is the better fit when software sends requests on behalf of customers or internal processes. It lets a team measure cost per ticket, document, workflow, accepted patch, or other business outcome.
The basic formula is:
monthly API cost =
(input tokens ÷ billing unit × input rate)
+ (cached input tokens ÷ billing unit × cached rate)
+ (output tokens ÷ billing unit × output rate)Use the live table and calculator above instead of copying rates into a spreadsheet that will go stale. The most common forecasting mistakes are ignoring output length, resending long context, routing easy work to a premium model, and letting agents retry without a task budget.
Log input, cached input, and output separately. Then measure whether a more capable model reduces retries or human review enough to justify its higher token rate. Our API cost calculation guide has a fuller workload method.
Batch and per-generation pricing
Batch pricing exchanges latency for a lower or more efficient processing path. It fits document enrichment, evaluation runs, catalog cleanup, classification, and other jobs that can wait. Cap rows and estimated tokens before launch, sample outputs first, and count retries as their own cost bucket.
Per-generation pricing is common for images, video, and speech. The useful metric is not cost per attempt; it is cost per accepted asset. Quality settings, duration, resolution, revisions, and failed generations all affect the denominator.
For either model, connect billing records to an outcome count. A cheap batch that needs a full rerun or a low-cost image tool that produces many rejected drafts can be the expensive option.
Self-hosting and open models
Open weights remove or reduce the model-license charge, but they do not remove infrastructure cost. GPUs, inference servers, scaling, monitoring, security, upgrades, and engineering time remain part of the bill.
Self-hosting is strongest when demand is large and predictable enough to keep hardware busy. Hosted APIs usually win for small or spiky workloads because the provider absorbs idle capacity and operational complexity. A managed open-model platform can sit between those extremes.
If coding-agent usage is the workload, compare the GLM Coding Plan with direct API billing and self-hosting. Treat its plan quota, API rates, and any infrastructure route as separate purchasing models.
A practical buying decision
| Workload | Start with | Measure |
|---|---|---|
| One person using AI interactively | Subscription | Useful work per active seat |
| Customer-facing product feature | Per-token API | Cost per successful user outcome |
| Offline enrichment or backfill | Batch | Cost per accepted row or document |
| Image, video, or speech production | Per-generation | Cost per approved asset |
| Stable high-volume inference | Managed or self-hosted open model | Fully loaded cost per successful output |
| Mixed complexity | Model routing | Quality, retries, latency, and cost by task tier |
Start with the simplest billing model that exposes the cost you need to control. Then run a short pilot using real work. Track adoption for seats, tokens and retries for APIs, acceptance rates for media, and utilization for infrastructure.
The token calculator is the fastest way to test API volume. For the broader architecture choice, compare local AI, APIs, and subscriptions.
FAQ
Is a Copilot subscription the same as API access?
No. Copilot plans are user-facing products with their own features and limits. Software integrations should be budgeted against the relevant API or platform rate card unless Microsoft explicitly states that a plan includes that usage.
Is per-token pricing always cheaper than a subscription?
No. A subscription can be efficient for a heavily active person, while an API is easier to meter for software. Compare cost per successful outcome and include unused seats, retries, and output volume.
When does self-hosting become worth it?
Usually when demand is predictable, utilization is high, and the team can operate inference reliably. Small or uneven workloads often cost less on hosted APIs because there is no idle hardware to carry.
What should I check before buying an AI plan?
Check the current price, billing commitment, usage limits, included models, data controls, regional availability, and whether API use is separate. Recheck the maintained pricing page immediately before procurement.