AI Platform Pricing: Examples & Companies

What is it

AI platform pricing is how general-purpose AI platforms — model APIs, inference services, and multi-model hosting providers — package and charge for access to model capability. It is the pricing layer that sits beneath most of the AI software market: the frontier labs, the inference marketplaces, and the application platforms that resell or wrap model access all fall inside it.

This is the broadest product category in the corpus. 167 companies carry the ai-platforms product segment, spanning three rough strata. At the base are the frontier and open-model labs — OpenAI, Anthropic, DeepSeek, and Mistral AI — that expose per-token APIs. In the middle are inference platforms — Fireworks AI, Together AI, Groq, and Novita AI — that serve open-weight models and rent GPU capacity. At the application layer are platforms that embed models into a finished workflow — HeyGen for avatar video, Runway for generative video, Gumloop and Relevance AI for agent workflows.

Because those three strata serve fundamentally different buyers, the pricing diversity inside this one category is the widest in the corpus. The same segment contains companies billing per token, per credit, per seat, per action, and per resolution. A developer buying inference wants a public rate card and no minimum; an enterprise buying a legal-AI platform wants a per-seat quote; a prosumer buying video generation wants an all-inclusive monthly credit bundle. AI platform pricing is best understood not as one model but as a set of archetypes, each attached to a buyer profile.

The through-line is that the model itself is rarely the billing unit the customer sees. Tokens are the native cost driver of inference, but most application platforms abstract them away behind credits or seats so buyers can reason about spend without doing token math. Understanding a platform’s price means first identifying which archetype it belongs to, then reading how it translates underlying model cost into a unit its buyer will accept.

How it works

AI platform pricing resolves into five recurring archetypes. Which one a company picks is driven almost entirely by who it sells to.

Pure-usage APIs publish a rate per input and output token, metered at consumption with free credit to start and volume discounts at scale. The spread is dramatic: DeepSeek’s V4-Flash starts at $0.0028/1M tokens on a cache hit, Groq prices Llama 3.1 8B at $0.05/$0.08 in/out, and OpenAI exposes the GPT-5.x API from $0.20/1M tokens. Cached and batch modes commonly cut the headline rate by 50% — Fireworks AI and Together AI both apply a 50% batch discount and layer per-hour GPU rates (H100 around $6–$7/hr) on top of serverless token pricing.

Credit subscriptions convert model cost into a monthly pool the buyer can budget. A worked example: HeyGen runs Free $0 → Creator $29 → Pro $49 → Business $149 (+$20/seat), each tier bundling a credit allotment, with a separate $5 pay-as-you-go API wallet. Runway starts at $12/mo bundling monthly credits that meter video, image, and audio generation, plus a Max tier with 9,500 rolling credits. The credit is a deliberate abstraction — it lets the vendor change underlying model economics without renegotiating, and lets the buyer avoid token math.

Seat + credits hybrids anchor predictable revenue on a per-user fee, then meter the AI-intensive work against a pool. Cursor is the canonical case; Relevance AI pairs seat tiers (Free, Pro $19/mo, Team $234/mo) with Actions usage and no-markup pass-through Vendor Credits. Outcome-based pricing is the rarest and most aligned: Intercom charges $0.99 per Fin AI resolution on top of $29–$132/seat platform tiers, billing only when the AI actually closes a ticket.

Companies using this

The 167 companies below all carry the ai-platforms product segment. They range from frontier model labs and open-model inference platforms to application-layer generation tools and vertical enterprise platforms — sort and filter the table to compare pricing models, billing units, and free-tier availability across the category.

Patterns observed

Pure-usage token pricing is the category floor, and it keeps falling. Every frontier and open-model lab in the corpus exposes a per-token rate, and the low end is dropping fast — DeepSeek’s cache-hit rate reset the floor while Groq competes on latency rather than a headline discount. The competitive pressure is so intense that cached, batch, and priority tiers have become standard — Fireworks AI and Together AI both advertise 50%-off batch modes, which is now table stakes rather than a differentiator.

Application-layer platforms abstract tokens into credits. The higher up the stack a platform sits, the less its buyer sees raw tokens. HeyGen, Runway, and Ideogram (Free, Plus $15, Pro $42, Team $20/user) all sell credit-metered subscriptions where the credit maps to a generated output — a video, an image — not to a token count. This abstraction is what lets these platforms run mass-market $15–$49 price points while insulating margin from underlying model volatility.

Multi-track pricing on one page is increasingly common. Rather than pick a single archetype, the larger platforms run several in parallel. Mistral AI prices two distinct surfaces from one page: flat-rate Vibe assistant subscriptions ($0–$24.99/user/mo) and a pure per-token developer API from $0.10/1M. OpenAI runs consumer subscriptions (Go $8, Plus $20, Pro $100 and $200, Business $20/seat) alongside its token API. The platform that serves both a developer and a prosumer buyer increasingly maintains a distinct pricing track for each.

The $200 prosumer ceiling has hardened. A cluster of consumer-facing platforms — OpenAI (ChatGPT Pro at $200), Anthropic (Claude Max from $100), Cursor, Perplexity, Google, and You.com — added a roughly $200/month top tier above the standard ~$20 Pro plan. This tier monetizes power users who would otherwise cap against lower plan limits, and it has been stable since it emerged. Following how these plans layer usage on a subscription base is the same problem covered in the guide to usage invoicing and billing cycles.

Counterexamples & variants

Not every AI platform publishes a rate card. The archetypes above assume public pricing, but the vertical enterprise strata breaks that assumption. Harvey, the legal-AI platform, sells per seat via sales-led quotes with no public rate card and reported seat minimums — the opposite of the developer-facing transparency norm. Enterprise-knowledge and regulated-vertical platforms consistently gate pricing because their buyers procure top-down and their deals are negotiated, not self-served. If you are comparing a vertical platform against a developer API on price transparency alone, you are comparing two different go-to-market cultures, not two pricing philosophies.

Outcome-based pricing is aspirational for most of the category, not standard. Intercom’s $0.99-per-resolution Fin pricing is held up as the model to emulate, but it remains rare because it requires an attributable, discrete outcome the buyer and vendor both agree happened. Most AI platforms cannot cleanly define such an outcome — what is the “outcome” of a token API or a video generator? — so they fall back to consumption or seats. Outcome pricing is a variant available to workflow platforms with a countable end state, not a template the whole category can adopt.

Credit systems vary from fully transparent to deliberately opaque. The credit archetype hides a wide range of honesty. Relevance AI passes through Vendor Credits at no markup, telling the buyer exactly what the underlying model cost was. At the other extreme, many application platforms set an internal, undisclosed credit-to-cost ratio so the buyer cannot reverse-engineer margin. Two platforms can both say “credits” and mean opposite things about transparency — the variant matters more than the label, which is why choosing and reading the underlying metric is worth understanding in its own right (see the guide to choosing the right usage metric).

What this means for buyers vs vendors

For buyers

Identify the archetype before you compare prices. A per-token rate and a per-seat quote are not comparable numbers, and a $29 credit subscription may cost far more or far less than a metered API depending on your volume — model your own usage first. For developer APIs, use a token calculator such as the Anthropic pricing calculator or the Cursor pricing calculator to translate a headline rate into a monthly bill at your actual volume, and always check whether cached and batch modes apply to your workload — a 50% batch discount changes the math entirely. For credit-based platforms, ask what the credit maps to and whether it is passed through at cost. For vertical enterprise platforms that gate pricing, expect a negotiated seat quote and factor acquisition risk into any multi-year commitment: several corpus platforms, including OpenPipe, OpenMeter, and Tavily, were acquired inside an 18-month window, and gating risk rises after a deal closes. A rate-lock clause is the cheapest insurance available.

For vendors

Your pricing archetype should follow your buyer, not your architecture. If you sell to developers, a public per-token rate with generous free credit and clear cached/batch tiers is the price of entry — opacity reads as a red flag in this segment. If you sell to prosumers or teams, credits let you set a legible monthly price while insulating margin from model-cost swings, but you owe the buyer clarity on what a credit buys. If you sell top-down into an enterprise or regulated vertical, seats and sales-led quotes remain viable, and gated pricing is defensible where deals are genuinely negotiated. The strategic risk across all archetypes is model-cost deflation: with the token floor still falling, any platform whose margin depends on a fixed markup over inference cost should assume that markup compresses. The durable positions are abstraction (credits, outcomes) and differentiation (latency, vertical depth), not a thin resale spread. For the mechanics of metering and invoicing usage at scale, the guide to usage invoicing and billing cycles covers the operational layer.

Explore this theme in the knowledge graph →