AI pricing is not one model — it is five
The reason AI vendor pricing feels impossible to compare is that vendors do not bill on the same axis. An LLM API charges per million tokens. A vector database charges per stored gigabyte plus reads and writes. A coding agent charges per seat with a usage quota. A voice provider charges per minute of speech. An observability tool charges per traced event or per seat. Two products that cost the same on paper can differ tenfold once your real workload meets the meter.
This overview is the category map. It does not re-run any single "X vs Y" rate comparison — the spoke pages below do that with verified numbers. Instead it teaches you to read each pricing model, then routes you to the right comparison for your sub-category: LLM APIs, vector databases, voice AI, coding agents, workflow automation, and observability.
The four pricing units to learn first
- Per-token — LLM APIs. Billed per million input and output tokens, with output usually 3-5x the input rate. Caching and context length swing the real cost far more than the headline rate.
- Per-GB plus operations — vector databases. A stored-data rate, then separate charges for reads, writes, and sometimes a serverless minimum. Index size, not query count, often dominates.
- Per-seat with a quota — coding agents and automation platforms. A flat monthly seat price buys a usage envelope; heavy users step up a tier or pay dollar-based overage.
- Per-minute — voice AI. Billed per minute of speech-to-text, text-to-speech, or full conversation, frequently stacked across providers in one call.
Get the unit right and the comparison becomes arithmetic. Get it wrong and you compare a sticker price to a usage envelope. The Pricing Model Picker helps you decide which unit fits the product you are building, and the AI Stack Cost Calculator rolls every unit into one monthly figure across three usage volumes.
LLM APIs — per-token
The per-token market moves fastest, so a comparison is only useful with a verified date attached. For the two largest providers, the Anthropic vs OpenAI API pricing breakdown lays the flagship and budget tiers side by side. When the only goal is the lowest rate for a given task, the cheapest LLM API guide ranks the current options. Because token prices fall on a schedule and features creep up, run the Model Price Drop Stress Test to see how a rate cut or a vendor increase moves your margin before you commit a stack to one provider.
Vector databases — per-GB plus operations
Vector pricing hides its real cost in index size and operation counts rather than the per-GB headline. The Pinecone vs Qdrant pricing comparison shows how a serverless and a self-hostable model diverge at scale, and the cheapest vector database guide ranks the field for a given embedding volume. To put your own numbers in, the Embeddings DB Cost calculator compares Pinecone, Postgres with pgvector, LanceDB, and Turbopuffer for your specific row count and dimensions.
The serverless model is where most RAG bills surprise people. If you are trying to read Pinecone serverless pricing — the read-unit, write-unit, and per-GB-month storage rates, plus the $50/month Standard plan minimum — the verified breakdown lays each line item against Qdrant's usage-based cluster billing. For the full vendor matrix in one place, the 2026 AI vendor pricing and TCO report aggregates every category's published rates with computed cost scenarios.
Voice AI — per-minute
Voice is the clearest per-minute category, and the trap is that a single call can stack speech-to-text, a model, and text-to-speech from three vendors. The Vapi vs Bland vs Retell pricing comparison breaks down the full-stack voice agents, while Deepgram vs AssemblyAI pricing covers the transcription layer underneath them. If voice is replacing a support function, the AI vs Human Support Cost calculator turns the per-minute rate into a monthly headcount comparison.
Coding agents — per-seat with a quota
Autonomous coding agents bill a flat seat price plus a usage quota, which makes their effective cost hard to read from the pricing page alone. The Devin vs Factory AI pricing comparison corrects the widely-quoted stale Devin numbers and lays the current tiers side by side. For the choice between building agent capability in-house and buying a managed one, the Build vs Buy Decision Engine weighs total cost of ownership against time to market.
Workflow automation — per-seat or per-run
Automation platforms split between per-run task pricing and per-seat agent pricing, and the line is blurring as agents replace static workflows. The Lindy vs Relay pricing comparison covers the newer agent-native platforms, where the meter is closer to a coding agent than a classic task tool. Pair it with the broader automation field to see where the per-run model still wins on a high-volume, low-judgment workflow.
Observability — per-event or per-seat
LLM observability bills either per traced event or per seat, and the difference decides whether a high-volume app or a small team is the expensive case. The Braintrust vs Phoenix vs Langfuse pricing comparison maps the event-versus-seat split across the major tools, so you can match the meter to whether your cost driver is traffic or team size.
Comparing across pricing models
The hardest comparisons cross meters — a per-token vendor against a per-seat one, or a self-host path against a managed serverless one. Two questions cut through it. First, what is the switching cost if the rate changes? The LLM Vendor Lock-In Cost calculator estimates the migration tax of being tied to one provider's pricing. Second, which feature is actually driving the bill? The AI Feature Attribution tool assigns cost to the features that incur it, so you price or cap the right one. When the build path looks too expensive, the Vibe-Code Platform Comparison contextualizes the managed no-code AI platforms against raw API cost.
How to read any AI pricing page
- Find the unit first — token, GB, seat, minute, or event. Everything else is noise until you know what you are being charged for.
- Read the overage, not the tier — the headline tier is a floor. The overage rate is what a real workload pays, and it is often buried below the pricing table.
- Date every number — AI rates move quarterly. A comparison without a verified date is a guess. Re-check the official page before committing.
- Separate the stack — one voice call or one RAG query can touch three vendors. Cost the whole path, not the cheapest single line.
- Model the price shock — assume your cheapest vendor raises rates or your model gets deprecated. The stress test shows whether your margin survives it.
Once a vendor is chosen, the next question is what it does to your unit economics. The AI Product Economics guide turns vendor cost into price floors, gross margin, and a defensible price for the product you sell.