Four Vendors, Four Meters: AI Pricing Compared
Anthropic, OpenAI, Google and Microsoft all sell the same kind of capability, and none of them bills it the same way. One sells a seat with usage on top, one sells tokens per model and tier, one bundles AI into an office suite, and one runs three separate meters for one assistant brand. A side-by-side table is the first thing a renewal team asks for and the hardest thing to build honestly.
This is that table, with its limits stated. Every price below is a published list price taken from the vendor's own page and accessed on 23 October 2026. Negotiated discounts, credits and enterprise agreements are not included, because none of them is public. We will refresh the table each quarter and date every change.
If the units themselves are new to you, start with Tokens, Explained for the People Who Pay the AI Bill.
API prices per million tokens
Standard tier, prompts under each vendor's long-context threshold, US dollars per million tokens. The flagship, mid and small grouping is ours, not the vendors'.
| Vendor | Tier | Model | Input | Cached input | Output |
|---|---|---|---|---|---|
| Anthropic | Flagship | Claude Opus 5.5 | $4 | $0.20 | $20 |
| Anthropic | Mid | Claude Sonnet 5.5 | $2 | $0.10 | $10 |
| Anthropic | Small | Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 |
| OpenAI | Flagship | gpt-6-astra | $10 | $1 | $50 |
| OpenAI | Mid | gpt-6.1-sol | $2 | $0.10 | $10 |
| OpenAI | Small | gpt-6-luna | $0.10 | $0.01 | $0.50 |
| Flagship | Gemini 4 Argon | $2 promotional ($4 from 1 Dec 2026, when the promotion ends) | $0.10 ($0.20) | $10 ($20) | |
| Mid | Gemini 3.8 Flash | $0.75 ($1.50 from 1 Jan 2027) | $0.075 ($0.15) | $3.75 ($7.50) | |
| Small | Gemini 3.5 Flash-Lite | $0.30 | $0.03 | $2.50 |
List prices, accessed 23 October 2026. Anthropic's cached figure is the cache read rate; writing to its cache costs 1.25 times input for five minutes or 2 times for one hour. Google adds a storage charge per hour for cached Flash and Flash-Lite content. Haiku 5.5 prompts over 100,000 tokens cost $0.50 in and $2.50 out.
Microsoft is not in this table on purpose. Microsoft Foundry sells OpenAI and other models per token at rates that vary by region and deployment type, and it sells Claude through the Azure Marketplace in Claude Consumption Units converted from Anthropic's own rates, according to Anthropic's pricing page. A single Microsoft row would hide more than it shows.
Discounts and surcharges that apply everywhere
| Lever | Anthropic | OpenAI | |
|---|---|---|---|
| Batch | 50% off input and output | 50% off (Batch and Flex) | 50% off |
| Faster service | Fast mode on Opus 5.5: $8 in, $40 out | Fast tier: 2 times standard | Priority: $3.60 in, $18 out on Argon today |
| Data location | US-only inference: 1.1 times | Not compared here | Not compared here |
| Web search | $10 per 1,000 searches | $10 per 1,000 calls | 5,000 free a month, then $14 per 1,000 |
List prices, accessed 23 October 2026, from the three pricing pages linked above. The batch discount is the one lever every vendor offers on the same terms, and the one most often left unused.
Seats where a price is published
| Vendor | Plan | Published price | What sits on top |
|---|---|---|---|
| Anthropic | Team, Standard seat | $20 per seat per month annual, $25 monthly | Up to $500 a month of pooled API credits; 2 to 150 seats |
| Anthropic | Team, Premium seat | $100 per seat per month annual, $125 monthly | As above |
| Anthropic | Enterprise | $20 per seat per month, billed annually, 20-seat minimum | All usage billed at API rates |
| Microsoft | Microsoft 365 Copilot | $30 per user per month, paid yearly | Copilot Credits for agents beyond the licensed-user allowance |
| Microsoft | Microsoft 365 Copilot Business | $18 per user per month annual (promotion, 1 July to 31 December 2026; list $21), $25.20 monthly | As above |
| Microsoft | Copilot Studio capacity pack | $200 per month for 25,000 Copilot Credits | Pay-as-you-go through an Azure subscription; Commit Units save up to 20% |
List prices, accessed 23 October 2026.
Seats against meters, vendor by vendor
- Anthropic has split the two cleanly. The Enterprise seat buys access and administration; every token is billed at API rates. We covered the change in Anthropic Split the Enterprise Seat From the Meter.
- OpenAI publishes its API rates in full. Its ChatGPT business plans are a different matter: its help center describes flexible, credit-based pricing for Enterprise, Edu and Business, but the pricing pages could not be read by automated access when we checked and the Enterprise price is not published. We leave those numbers out rather than repeat a reseller's.
- Google sells the Gemini API per token and bundles Gemini features into Google Workspace plans. Workspace prices vary by country and currency, and we could not confirm a US-dollar figure from the official page, so they are not listed. Gemini Enterprise had no working official pricing page when we checked.
- Microsoft runs three meters under one brand: Copilot seats, Copilot Credits for agents, and Foundry tokens or provisioned throughput. The seat price is public; what the credits and Foundry deployments cost depends on how agents and applications are built.
The pattern is a hybrid: a seat that buys the right to use, and a meter that bills the use. How quickly the model under a brand can change is the subject of The Model Is the Commodity. Microsoft Just Proved It. The meter tends to follow.
Where each vendor reports usage
A price list is half of a reconciliation. The other half is the usage record, and each vendor keeps it somewhere different:
| Vendor | Usage and cost source | What it can break down |
|---|---|---|
| Anthropic | Usage and Cost Admin API | Usage by API key, workspace, model and service tier; daily cost in USD; requires an admin key |
| OpenAI | Organization usage and costs endpoints | Usage by model, project and key; daily costs by project and line item |
| Cloud Billing reports for the billing account behind the API | Cost by project and SKU | |
| Microsoft | Microsoft Graph Copilot usage report, Power Platform admin center, Azure Cost Management | Last Copilot activity per licensed user and app; Copilot Credit consumption by environment; Foundry spend by subscription |
Claude Enterprise seats on claude.ai report through a separate analytics interface from the API, so an organization that uses both has two Anthropic sources before it has looked at any other vendor.
What none of them publishes
- Your price. Volume discounts, committed-spend agreements and credit pools are negotiated and confidential.
- ChatGPT Enterprise. No public list price.
- Microsoft provisioned throughput rates in this post. Microsoft's documentation explains how provisioned throughput units are billed; the dollar rate depends on region and deployment type, and we have not quoted it.
- Microsoft 365 E7. Trade press has reported a price; we have not found it on a Microsoft page.
- Price per task. Every vendor publishes a price per token, none a price per answer. Reasoning tokens, tokenizer differences and cache hit rates make the same task cost differently on two models with identical list prices.
That last point is why a table like this one is a starting position, not a forecast. AI Vendor Lock-In Turns Usage Into a License Metric explains why the consumption record matters more than the rate card when you try to move.
The question this leaves open
Four vendors, four meters, and at least six places where usage is recorded. The prices compare in a table. The usage does not, until someone puts seats, meters and keys side by side for the same organization. Where should that view live, and who owns it?
Software Inventory Manager (SIM) inventories your software. Related reading: Anthropic Split the Enterprise Seat From the Meter, AI Vendor Lock-In Turns Usage Into a License Metric and The Model Is the Commodity. Microsoft Just Proved It.
