Every major AI provider runs the same two billing universes side by side. One sells you a bucket of usage for a flat monthly fee; the other charges you for exactly what you burn, per token. Here's how the big five line up on both.
You pay a fixed monthly fee. Tokens are still counted under the hood, but you never see them — you experience limits as messages inside rolling windows + a weekly/daily cap.
Like an all-you-can-eat buffet with a "please don't camp the table" policy.
The API bills per million tokens, split into input + output. Team/Enterprise tiers add pooled seats, admin, and compliance — with raw API usage billed separately.
Like a utility meter: you pay for every kilowatt that flows.
| Dimension | Consumerthe $0–$300/mo chat apps | Enterprise / APIraw API · Team · Enterprise |
|---|---|---|
| What you're billed on | A flat subscription. Usage = number/length of messages, abstracted into a "usage" allowance. | Actual input + output tokens. Seats + (optionally) metered API spend on Team/Enterprise. |
| Do you see tokens? | No. Tokens run silently behind the limit. Long chats & attachments just drain your window faster. | Yes — itemized. You can count tokens before sending and see them on every response. |
| How limits work | Rolling session windows + daily/weekly caps. Resets on a clock, not a wallet. | Rate limits (tokens/min, requests/min) by tier — but no hard "you're cut off" wall; you pay for what you use. |
| Who runs this model | Anthropic, OpenAI, Google, xAI all sell consumer apps. AWS doesn't — it's API/enterprise only. | Everyone, plus AWS Bedrock & Google Vertex resell other labs' models as managed, governed endpoints. |
| Cost predictability | Totally fixed — your bill can never surprise you. At most you pause briefly until the window refreshes. | Variable. Scales linearly with volume; great for automation, but the bill moves with usage. |
| Best for | Individuals, daily hands-on work, coding assistants in the app, classroom/personal use. | Apps, pipelines, agents, high volume, org-wide rollout, governance & audit needs. |
| Cost levers | pick a smaller modeltrim contextbuy extra usage at API rates | prompt cachingbatch API −50%right-size model |
| Provider / app | Free | Standardeveryday tier | Premiumpower / pro |
|---|---|---|---|
| ClaudeAnthropic | $0 | Pro $20/mo | Max $100 (5×) · $200 (20×) |
| ChatGPTOpenAI | $0 | Plus $20/mo | Pro $100 / $200/mo |
| GeminiGoogle | $0 | AI Pro $20/mo | AI Ultra $100 / $200/mo |
| GrokxAI | $0 | SuperGrok $30/mo | Heavy $300/mo |
| AWSAmazon | — | no consumer chat plan | enterprise / API only (Bedrock · Amazon Q) |
The market has settled on a near-universal ~$20/mo standard tier (SuperGrok is the outlier at $30). Premium tiers buy more usage of the same flagship models, not different ones. All standard/premium tiers count tokens silently — the price buys you a bigger, faster-refilling bucket.
| Provider | Flagship model | Input$ / M tok | Output$ / M tok | Cheaper / fast tier |
|---|---|---|---|---|
| Anthropic | Claude Opus4.7 / 4.8 |
$5.00 | $25.00 | Sonnet $3 / $15 · Haiku $1 / $5 |
| OpenAI | GPT-5.5current flagship |
$5.00 | $30.00 | GPT-5.4 $2.50 / $15 |
Gemini 2.5 Pro≤200K context |
$1.25 | $10.00 | 2.5 Flash $0.30 / $2.50 | |
| xAI | Grok 4.3current flagship |
$1.25 | $2.50 | Grok 4.1 Fast $0.20 / $0.50 |
| AWS | Amazon Nova Proon Bedrock |
$0.80 | $3.20 | + resells Claude, Llama &c. (≈20–35% markup) |
| DeepSeek | DeepSeek-V4V4-flash tier |
$0.14 | $0.28 | cache hit −98% · (V3/R1 names retiring Jul 2026) |
| Mistral | Mistral LargeEuropean, open-ish |
$2.00 | $6.00 | Small ~$0.20 / M |
| Meta | Llama 4open weights |
≈$0.15–0.50 | ≈$0.50–0.90 | self-host = electricity only (host-set rates) |
Per million tokens (MTok); output is the expensive side everywhere — typically 2×–6× input. Batch ≈ −50% and prompt caching (cached reads −90%) are offered across providers and stack. AWS is different: it makes the Nova models but mostly operates as a marketplace (Bedrock) hosting other labs' models behind one governed, IAM-controlled endpoint — convenient, at a markup. DeepSeek and open-weight Llama models undercut the Western flagships by ~5–20×, and self-hosted Llama costs only electricity — the cheap end is a real competitive force. Note: newer flagships (e.g. Opus 4.7+, GPT-5.5) can emit more tokens for the same task, raising effective cost even at the same rate card; very long contexts (>200K) often bill at a higher tier.
Feeding an image in (a screenshot, chart, or mockup) costs input tokens on every provider — roughly proportional to its pixel area:
Generating an image is a separate product. Claude doesn't generate images at all — it only reads them. OpenAI, Google, and xAI do, and bill it as its own line item (per-image, or in a tokens-per-image scheme — e.g. roughly $0.01–$0.19 per image depending on model and resolution) rather than as normal text output. On consumer plans, both reading and generating just eat into your usage window with no visible line item.
Worked example — a 1,500 × 1,000 px screenshot ≈ (1,500 × 1,000) ÷ 750 ≈ 2,000 input tokens. At Gemini 2.5 Pro input ($1.25/MTok) that's ~$0.0025; on any $20 chat plan it's "free" but quietly shrinks your usage window.
Sources — Anthropic API pricing & vision docs (platform.claude.com); DeepSeek (api-docs.deepseek.com — V4-flash $0.14/$0.28, cache hit −98%; legacy V3/R1 names deprecated Jul 2026); Mistral (mistral.ai/pricing — Large $2/$6); Meta Llama 4 via hosts (Together, Groq, Bedrock — ≈$0.15–0.90, open weights); OpenAI API pricing (developers.openai.com — GPT-5.5 $5/$30, GPT-5.4 $2.50/$15); Google Gemini API pricing (ai.google.dev — 2.5 Pro $1.25/$10, Flash $0.30/$2.50); xAI docs (docs.x.ai — Grok 4.3 $1.25/$2.50, 4.1 Fast $0.20/$0.50); AWS Bedrock pricing (aws.amazon.com/bedrock/pricing — Nova Pro $0.80/$3.20). Consumer tiers: each provider's plans page. Treat all figures as current-as-of June 2026 and subject to change.