Generative AI licensing is the set of commercial and legal models under which vendors grant the right to use large language models and the assistants built on them. Four models recur across vendors. Assistant products are sold per user, often as an add-on that requires a base licence: Microsoft 365 Copilot is a User SL that must sit on a listed prerequisite Microsoft 365 or Office 365 licence.[1] Model access through an API is metered in tokens, as with OpenAI’s prices “per 1M tokens” for input, cached input and output.[2] Large customers can buy committed capacity, such as Google’s Provisioned Throughput in generative AI scale units.[3] Model weights can also be licensed for self-hosting, under terms that range from the Apache License 2.0 to Meta’s Llama community licences.[4][5]
Most vendors combine several of these models. Anthropic’s Enterprise plan charges “a fixed seat fee and separate usage charges”,[6] and every GitHub Copilot plan carries a monthly allowance of AI credits on top of the seat.[7] This article compares the seven vendors covered in the LICENSEWARE Wiki: OpenAI, Anthropic, Meta, Mistral AI, Microsoft, GitHub and Google. The general theory of term and metered licensing is in Subscription and consumption licensing.
History
The dated vendor documents show the models forming between 2023 and 2026. OpenAI states that “As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models”.[8] Meta released the Llama 2 Community License Agreement on 18 July 2023, with the 700 million monthly active user condition that later Llama versions keep.[9][5] Mistral AI introduced its Non-Production License with Codestral on 29 May 2024.[10] In February 2025 the Open Source Initiative stated that the Llama 3.x community licences do not meet the Open Source Definition.[11]
Units have also changed. On 1 September 2025 Microsoft changed the common currency for Copilot Studio agents from messages to Copilot Credits.[12] Anthropic’s Priority Tier capacity commitments “are no longer available for purchase”, and existing commitments run to their end date.[13] Google renamed the Vertex AI Platform as Gemini Enterprise Agent Platform.[14]
Scope and definitions
- Token. The metering unit for model APIs. Vendors price input and output tokens separately: Anthropic defines “MTok” as million tokens,[15] and Google’s output price includes thinking tokens.[16]
- Credit. A vendor-defined unit that converts usage into a common currency. Microsoft calls Copilot Credits “the unit that measures agent usage”,[17] and GitHub sets 1 AI credit at $0.01 USD.[7]
- Seat. A per-user subscription. GitHub defines a Copilot seat as a licence assigned to a unique user account through a Copilot Business or Copilot Enterprise plan.[18]
- Committed capacity. Throughput reserved for a term, such as OpenAI Scale Tier “token units” of input and output tokens per minute for one model snapshot.[19]
- Open-weight licence. A licence to use, modify and redistribute model weights. Only some meet the Open Source Definition, whose ten criteria include free redistribution and permission for derived works.[20]
How it works
Per-seat assistants and prerequisites
Assistant plans are priced per user per month. ChatGPT Business sells Standard and Premium seats, and OpenAI charges “for all standard and premium seats in your workspace, whether or not they’ve been assigned”.[21] Claude Team charges for the members on the team at the start of each billing cycle.[22] GitHub bills organizations for assigned Copilot seats at the end of each monthly cycle.[23] Mistral’s Vibe Team plan and Google’s Gemini Enterprise are also priced per user or per seat per month.[24][25]
Some assistants depend on a base licence. Microsoft Product Terms state that purchases of listed licences “also require the purchase of a Prerequisite License”, and list the Microsoft 365, Office 365 and standalone plans that qualify for Microsoft 365 Copilot.[1] Copilot Enterprise is designed for enterprises using GitHub Enterprise Cloud.[26] A Gemini Enterprise licence is specific to a Google Cloud project and location.[27] Seat entitlement can also bundle usage: Microsoft 365 Copilot users can use employee-facing agents at no charge under conditions, while agents used by people without a Copilot licence draw Copilot Credits.[17]
Token and credit metering
API prices are set per million tokens and differ for input, output and cached input. Anthropic prices a 5-minute cache write at 1.25x the base input price and a cache read at 0.1x in most cases.[15] Mistral states that cached input costs up to 90% less.[24] Asynchronous batch processing is discounted by 50% at OpenAI, Anthropic and Mistral,[28][15][24] and Google runs Batch and Flex at half the Standard rate on the models shown.[16] Regional processing can add a surcharge, for example Anthropic’s 1.1x multiplier for US-only inference.[15]
Credits and pools sit between seats and tokens. OpenAI Enterprise and Edu customers on credit-based agreements “purchase a shared credit pool at the contract level”.[29] GitHub adds each Business or Enterprise seat’s included AI credits to a pool at the billing entity level, forfeits unused credits and resets the allowance on the first day of each calendar month.[7] Microsoft Copilot Studio capacity is pooled across the tenant, enforced monthly and does not carry over, and a capacity pack holds 25,000 Copilot Credits per month.[17][30] Google’s Gemini API Paid tier requires a prepayment of at least $5 and serves requests only while the Prepay balance is positive.[31]
Vendors reserve the right to change rates. OpenAI price changes take effect “fourteen days after they are posted”,[32] and Anthropic rate updates take effect at the earlier of 30 days after posting or notice to the customer.[33]
Committed capacity
Committed offers trade a term commitment for guaranteed throughput or a discount. Google sells Provisioned Throughput in GSUs for 1 week, 1 month, 3 months or 1 year, and sizes it by converting queries per second into tokens per second and rounding up to the model’s minimum increment.[3] OpenAI’s Scale Tier token units are bought for a minimum of 30 days, and usage above the entitlement is billed at pay-as-you-go rates.[19] Microsoft offers a Copilot Credit Pre-Purchase Plan, “a one-year, prepaid option for Copilot Credits”.[12] Anthropic’s sales-assisted Enterprise plans list tiered incentives on committed spend,[34] and a ChatGPT Enterprise agreement may include a spending commitment across ChatGPT Enterprise and the API.[35]
Open-weight model licences
Open-weight licences are not metered. Meta’s Llama licences grant “a non-exclusive, worldwide, non-transferable and royalty-free limited license”.[5] Their commercial condition is a threshold: if the licensee’s products had more than 700 million monthly active users in the calendar month before the version’s release date, it must request a licence “which Meta may grant to you in its sole discretion”.[5] The licences do not say how monthly active users are counted.[5] Redistributors must display “Built with Llama”, and derivative models must begin their names with “Llama”.[5]
Mistral AI releases some models under the Apache License 2.0 and others under a modified MIT licence whose users with monthly revenue above USD 20M need a commercial licence or must use Mistral Studio.[4] The Apache License 2.0 is on the OSI list of approved licences.[36] Redistribution under it requires a copy of the licence, marking of modified files and retention of notices.[37] Mistral’s Research License limits use to non-profit, non-commercial research,[38] and its Non-Production License limits use to testing, research, personal or evaluation purposes in non-production environments.[39] For a SAM team, a self-hosted model has no seat or token meter, but its licence conditions must still be tracked against the deployment.
Acceptable use and data terms
Every vendor attaches an acceptable use policy, but its legal effect differs. The Llama licences make compliance with the Acceptable Use Policy a licence condition, so a breach can end the licence;[5] the Llama 4 policy also withholds rights for multimodal models from EU-domiciled individuals and companies.[40] OpenAI states that breaking its Usage policies “may mean you lose access”.[41] Anthropic’s Usage Policy applies “including via any authorized resellers or passthrough access”, and violations may lead Anthropic to “throttle, suspend, or terminate” access.[42] Mistral’s Usage Policy does not apply to products on customer infrastructure or to its open-source models.[43] Commercial terms also restrict competitive use: OpenAI bars using Output to develop competing models,[32] and Anthropic bars using its services to train competing AI models.[33]
Business terms generally exclude customer content from training. OpenAI “will not use Customer Content to develop or improve the Services” unless the customer agrees,[32] Anthropic “may not train models on Customer Content from Services”,[33] and Google Cloud will not use Customer Data to train or fine-tune models without permission.[44] Microsoft states that Copilot prompts and responses are not used to train foundation LLMs.[45] GitHub’s Generative AI Services Terms apply the same default to Business and Enterprise bought directly.[46] Free and individual tiers differ: Google may use content from unpaid Gemini API services to improve its products,[47] and OpenAI, Anthropic and GitHub individual terms permit training unless the user opts out.[48][49][50]
Vendor comparison
| Vendor | Unit | How counted | Commitment options |
|---|---|---|---|
| OpenAI | Standard seat, Premium seat; input and output tokens[21][2] | All purchased Business seats billed per cycle; API per 1M input, cached input and output tokens[21][2] | Monthly or annual Business billing; Enterprise credit pool or token-based; Scale Tier, 30-day minimum[28][29][19] |
| Anthropic | Standard and Enterprise seat; base input and output tokens[51][15] | Team members at start of cycle; Enterprise usage at API rates; tokens per MTok[22][6][15] | Annual Enterprise seats; committed-spend incentives; Priority Tier closed to new purchase[6][34][13] |
| Meta | Monthly active users (licence condition, not a price)[5] | Over 700 million in month before release date, including affiliates[5] | None published; licence above threshold at Meta’s discretion[5] |
| Mistral AI | Input and output per M tokens; user per month[24] | Input and output counted separately; Vibe Team per user[24] | Monthly, or fixed-term annual or multi-annual with early termination fee[52] |
| Microsoft | User SL; Copilot Credit[1][17] | Add-on per user on a prerequisite; credits per agent feature, pooled per tenant[1][17] | Annual Copilot commitment, paid yearly or monthly; 25,000-credit packs; one-year pre-purchase; Azure pay-as-you-go[53][30][12] |
| GitHub | Copilot seat; GitHub AI Credit[23][7] | Assigned seats at cycle end; tokens by model converted to credits, pooled[23][7] | Monthly per granted seat; pay-as-you-go additional usage with budgets[23][7] |
| Gemini input token; GSU; Gemini Enterprise seat[16][3][25] | Per 1M input and output tokens; Agent Platform charges only requests that return 200; per seat per month[16][3][25] | $5 minimum prepay; Provisioned Throughput 1 week to 1 year; Code Assist 12-month term[31][3][54] |
Practice in software asset management
Enterprise agreements
Generative AI terms usually sit inside a master agreement with an order form. The OpenAI Services Agreement ranks the Order Form above the Agreement and the OpenAI Policies,[32] Anthropic charges fees at its Model Pricing Page rates “unless otherwise agreed by the parties”,[33] and Mistral applies its pricing page unless an Order Form says otherwise.[55] The purchase channel can change the governing terms. GitHub Copilot bought through Microsoft falls under Microsoft’s Product Terms rather than GitHub’s own terms,[56] and Claude models on Amazon Bedrock and Google Cloud are invoiced by the cloud provider.[15]
Tracking seats and consumption
An effective license position for a generative AI vendor has two parts: a seat count reconciled as in named user licensing, and a consumption balance compared with the credit pool or commitment. ISO/IEC 19770-1:2012 stated that its SAM processes applied to software as a service.[57] The U.S. Government Accountability Office has recommended that agencies compare license inventories with purchases regularly.[58] The vendors publish the controls that make this possible. GitHub lets administrators set budgets per user, cost center, organization and enterprise, and enables additional usage by default unless the paid usage policy is disabled.[7] OpenAI API rate limits are set at the organization and project level, not the user level,[59] and Anthropic sets them at the organization level.[60] Gemini API rate limits apply per project rather than per API key.[61]
Individual accounts and shadow AI
Staff can buy individual AI plans outside the organization’s agreements, a pattern sometimes called shadow AI. The licensing question is which terms govern that use. GitHub states that Copilot Free is intended for individual use only, and cancels a personal plan automatically, with a prorated refund, when the user is assigned an organization seat.[23] Individual ChatGPT and Claude plans run under consumer terms, not the business agreements.[48][49] Because data-use terms differ by tier, as described above, identifying these accounts is also a data-protection task.
Chargeback
Consumption meters can be allocated to cost owners. Microsoft Copilot pay-as-you-go uses billing policies that link an Azure subscription to a set of users, with an optional budget.[62] GitHub’s budgets can be set per cost center.[7] Gemini API tiers and caps are set at the billing account level,[31] and Gemini Enterprise licences are distributed to projects and locations.[27]
Relationship to other license models
Per-seat AI plans follow the named-user pattern, and token or credit balances follow the consumption pattern described in Subscription and consumption licensing. Open-weight licences resemble other software licenses for self-hosted code, except that the Llama threshold is measured in users of the licensee’s products rather than in deployments.[5] Microsoft 365 Copilot reuses the User SL meter of other Microsoft Online Services.[1]
Criticism and challenges
The vendor documents show several difficulties for licensees. Units are not comparable: OpenAI says a previous credit total “is not a direct estimate of the same activity’s dollar cost”.[35] Prepaid capacity can be lost, as with GitHub’s forfeited AI credits and Microsoft’s non-carryover capacity.[7][17] Published rates can change at short notice, fourteen days after posting in OpenAI’s case, and a signed Order Form ranks above the published terms.[32] Licence status is also disputed: the OSI rejects the description of Llama licences as open source,[11] while Mistral’s pricing FAQ and Help Center describe Apache 2.0 self-hosting differently.[24][4]
Out of scope
This article does not cover image, audio and tool prices, full per-model price tables, discounts, accounting treatment, or AI regulation. Vendor detail, including catalog rows, is in the seven vendor articles linked above.