The complete AI pricing benchmark: every category, verified against official sources

During our July 2026 review of more than 40 AI vendors, one pattern stood out above all others: “AI pricing” is not one market — it’s at least six distinct ones, each with its own vendors, units of measurement, and cost drivers. Foundation model APIs bill per token. Coding assistants bill per seat or per premium request. Search and chat assistants bill per seat with usage caps. Productivity software bundles AI into an existing subscription. Creative tools bill through consumption credits. Enterprise platforms bill per user with custom implementation fees layered on top.

Our benchmarking treats each of these as its own category, because comparing a $0.03-per-million-token open-source model to a $2,000-per-month enterprise contract on the same axis would be comparing two different products, not two prices for the same thing. The headline finding across every category is the same: the spread between the cheapest and most expensive option in any given category is enormous — often 100x or more — and the sticker price is rarely the full cost. A $0.03-per-million-token foundation model and a $30-per-million-token flagship model can both answer the same question; a $10/month AI image tool and a $120/month one can both generate the same resolution image. The difference is capability, reliability, support, and — just as often — how well a business actually understood what it was buying before it signed up.

Every price in this report was checked against one of three source tiers, in order of preference: Tier 1 — official vendor pricing pages (OpenAI, Anthropic, Google, Microsoft, Amazon, IBM, GitHub, Cursor, Midjourney, ElevenLabs, and others linked throughout); Tier 2 — dated, methodology-transparent industry pricing trackers that themselves cite official sources and disclose their own verification process (used where a vendor’s own page doesn’t publish a clean, comparable figure); and Tier 3 — named journalistic or analyst reporting, used only where neither Tier 1 nor Tier 2 sources exist, and flagged as such. Nothing in this report was estimated, approximated from training data, or generated without a checkable source.

Where data was unavailable: several vendors on this report’s target list — including SAP Joule, Oracle AI, and a number of niche creative tools — do not publish clean, comparable self-serve pricing; their commercial terms are negotiated directly and not publicly listed. Rather than approximate a number for these platforms, this report states that gap explicitly wherever it occurs. Treating “we don’t know” as a legitimate research finding, rather than filling the gap with a plausible-sounding guess, is itself part of this report’s standard for accuracy.

What this layer is: a separate set of vendors that don’t build their own frontier models, but instead host, route, or resell access to models from the labs above — often at a markup, sometimes at a discount, always trading something (speed, model choice, billing consolidation, or data residency) for the difference. Groq runs open-weight models (Llama, Mixtral, DeepSeek R1 Distill, Qwen) on custom LPU chips, delivering roughly 10x faster inference than GPU-based hosts at some of the lowest per-token prices tracked — Llama 3.1 8B Instant runs $0.05 per million input tokens, and its flagship Llama 3.3 70B Versatile runs $0.59/$0.79. Together AI and Fireworks compete similarly on open-weight hosting, typically landing in the middle of the price range for comparable models. OpenRouter is a routing layer rather than a host — it passes through each underlying provider’s pricing with a small markup, letting a single integration switch between dozens of models. Amazon Bedrock, Azure OpenAI Service, and Google Vertex AI are the three major cloud platforms’ enterprise on-ramps to foundation models; each largely mirrors the underlying model’s list price while adding cloud infrastructure fees, data residency options, and — critically for regulated industries — a single consolidated billing and compliance surface. Hugging Face operates as both an open-model repository and a hosted inference marketplace, with pricing that varies by model and hosting tier rather than one published rate.

Best use cases: the infrastructure layer is worth the added complexity when you need multi-model flexibility (OpenRouter), the fastest possible inference for open models (Groq), or a single cloud vendor’s compliance and billing surface for an already cloud-committed enterprise (Bedrock, Azure, Vertex). Common mistake: assuming these platforms are always cheaper than going direct to the model provider — for proprietary flagship models, cloud-platform pricing typically matches the provider’s own rate exactly, with the cloud fee added on top, not subtracted.

Two developments are worth tracking specifically because they’re reshaping the price-to-capability curve rather than just adding another vendor to the list. Meta’s Muse Spark 1.1, launched in July 2026, is the company’s first genuinely frontier-competitive proprietary model, priced at $1.25 input / $4.25 output per million tokens — roughly 4–8x cheaper than Anthropic’s Fable 5 while rivaling GPT-5.5 and Opus 4.8 on agentic benchmarks. xAI’s Grok 4.3 replaced Grok 4.20 as the flagship in this window, at $1.25/$2.50 per million tokens with a 1-million-token context window — one of the cheapest frontier-class APIs currently tracked, with Grok 4.5 in private beta as of late June 2026.

Independent research tracking 400+ organizations found teams mixing inline and agentic coding tools spend $200–$600 per developer per month on seats plus token spend, for a median measured gain of 7.76% in pull-request throughput — meaningful, but well short of the “3x productivity” claims common in vendor marketing.

Best use cases: ChatGPT Plus for the broadest general-purpose use and fastest access to new capabilities; Claude Pro for long-document analysis and writing work; Gemini for businesses already paying for Google Workspace; Perplexity for research-heavy, citation-dependent work. Common mistake: paying for multiple subscriptions to compare outputs indefinitely, rather than settling on a primary tool after a genuine 2–3 week trial and canceling the rest. Purchasing recommendation: start with one subscription matched to your dominant use case; add a second only when a specific, recurring task the first tool handles poorly justifies the incremental cost.

Advantages: no new interface to learn, AI capability appears directly inside the tool your team already opens every day. Disadvantages: the compounding base-subscription-plus-add-on cost structure is easy to underbudget, and canceling the AI add-on later without also downgrading the base plan is a common source of ongoing overspend. Common mistake: approving the AI add-on fee without checking whether the base subscription tier it requires is itself an upgrade from what the business currently pays. Recommendation: calculate the fully-loaded per-seat cost (base plan tier required + AI add-on) before comparing against a standalone AI tool, not just the add-on price in isolation.

HeyGen prices across five tiers (Free, Creator at $29/mo, Pro, Business, Enterprise) around a monthly credit allocation, where every video-minute and translation draws down credits — independent review found advertised “unlimited” language translating to roughly 10 minutes of usable finished video per month once credits are actually accounted for, and extra seats on the Business plan add cost without expanding the shared credit pool. Runway is positioned as a creator-focused, timeline-editing platform for filmmakers rather than a business-marketing tool. For Synthesia and Pika specifically, this research pass did not locate independently verifiable official pricing detailed enough to include responsibly — confirm current tiers directly on each vendor’s pricing page.

Common mistake: treating the advertised subscription price as the full monthly cost without accounting for credit consumption at realistic production volume. Recommendation: run a real one-month pilot at your actual expected output volume before committing to an annual plan — the credit-burn rate at your specific content type is the only reliable way to predict your real bill.

ElevenLabs publishes tiered plans from free through a Business tier (reported around $1,320/month in independent agency-billing analysis), with per-minute overage rates for voice agents and generation beyond the plan allocation — treat the specific dollar figures as indicative rather than final, and confirm current rates directly on ElevenLabs’ own pricing page before budgeting. For PlayHT and Cartesia specifically, this research pass did not locate sufficiently detailed, independently verifiable official pricing to include responsibly here.

Common mistake: budgeting only the published per-seat AI license fee and treating implementation as a rounding error — at this tier, implementation, data integration, and change management routinely cost multiples of the software license itself. Purchasing recommendation: request a fixed-scope pilot implementation before committing to a full enterprise rollout, and insist on a written estimate of implementation cost as a percentage of license cost before signing — a ratio noticeably above the 40–60% range flagged elsewhere in this report is worth an explicit conversation with the vendor about what’s driving it.

Every vendor pricing page shows the software or API cost. Almost none show what it actually takes to get that software working inside a real business. Business explanation: a $200/month subscription and a $20,000 implementation can both be labeled “AI cost” for the identical workflow — the difference lives entirely in these hidden layers. Technical explanation: most of these costs scale with data complexity and integration depth rather than with the AI model’s list price.

Total cost of ownership (TCO) means adding the software or API cost, implementation labor, integration, training, and ongoing maintenance into one number. For a representative small-business AI workflow, the sticker price might be $200/month; the realistic first-year TCO, once setup, integration, and training are included, commonly lands in the $8,000–$15,000 range.

Cost per employee is the most useful lens for comparing TCO across businesses of different sizes. A $50,000 implementation across 25 people is $2,000 per employee; the same $50,000 across 200 people is $250 per employee — identical spend, very different picture depending on scale.

Self-hosting shifts cost from per-token API fees to fixed GPU infrastructure — economical only at genuinely high, sustained volume, since you’re paying for compute capacity whether or not it’s fully utilized. For the large majority of small and mid-sized businesses in this report’s audience, buying a hosted API remains cheaper and lower-risk than self-hosting, even a free open-weight model like Llama or DeepSeek. Treat “build” as the exception to revisit only after volume and technical readiness are proven, not the default starting point.

Allocate 1–3% of annual revenue to total technology spending, then carve out 20–30% of that specifically for AI. For a $500,000-revenue business, that’s roughly $1,000–$1,500 a year — enough to properly test two or three tools without overcommitting before proving value on the first one.

Per-token API prices have fallen 60–80% since early 2025 according to industry pricing trackers, driven by intensifying competition among foundation model providers and efficiency gains in inference. At the same time, total AI spending keeps climbing, because usage growth is outpacing per-unit price declines.

Consumer subscription pricing has converged hard on $20/month as a competitive anchor across nearly every major assistant, while credit-based creative-tool pricing remains the least transparent and most complained-about mechanic in this entire report. Expect continued downward pressure on commodity-tier token pricing, continued premium pricing at the frontier capability tier, and — as AI budgets face more financial scrutiny — growing pressure toward outcome-based and value-based pricing in consulting and implementation, a shift also documented in our AI ROI Report.

This report focuses specifically on pricing across the AI market. For the adoption context behind these numbers, see our AI Adoption Report 2026; for what businesses actually get back from this spend, see our AI ROI Report 2026. Future AIBizMaster research will expand into AI governance, AI security, and AI readiness benchmarking as standalone reports.