Understanding AI tokens
What they are, how billing works, and current costs across major providers
April 2026 · API pricing reference
01 - What is a token?
Hello
token
ization
is
complex
= 5 tokens · each colour = 1 token
~4
characters per token
750
words per 1,000 tokens
1.5
A4 pages per 1,000 tokens
02 - How billing works
Your prompt
→
Input tokens counted
→
AI responds
→
Output tokens counted
→
Invoice
Output tokens typically cost 3–5× more than input tokens. Always budget for both.
03 - Price comparison (USD per 1 million tokens · Input / Output)
These are public API rates - standard pay-as-you-go prices available to anyone. Corporate and enterprise contracts are negotiated separately and are typically 20–50% lower than the figures shown. See section 05 for the full breakdown.
Anthropic Claude
Safety-focused · long context
Best for
Document analysis, coding, compliance & long-context tasks
Haiku 4.5fast
$1 / $5
Sonnet 4.6balanced
$3 / $15
Opus 4.6flagship
$5 / $25
Opus 4.7latest
$5 / $25
OpenAI GPT
Broad ecosystem · multimodal
Best for
General-purpose apps, vision, audio & existing Microsoft integrations
GPT-4o minifast
$0.15 / $0.60
GPT-4obalanced
$2.50 / $10
GPT-4.1flagship
$2 / $8
Google Gemini
Most affordable · free tier
Best for
High-volume, cost-sensitive workloads & Google Workspace users
Flash-Lite 2.5fast
$0.10 / $0.40
Flash 3.0balanced
$0.50 / $3
Pro 2.5flagship
$1.25 / $10
Pro 3.1latest
$2 / $12
Perplexity Sonar
Search-augmented AI
Best for
Real-time research, fact-checking & cited answers from the web
Sonarfast
$1 / $1
Reasoning Probalanced
$2 / $8
Sonar Proflagship
$3 / $15
+ $0.005 per web search call. Includes live web results by default.
04 - How to reduce costs
−90%
Prompt caching
Reuse repeated system prompts or documents at 10% of standard input cost
−50%
Batch processing
Send non-urgent jobs asynchronously - all providers offer a 50% batch discount
−60%
Right-size your model
Use fast/cheap models for simple tasks, flagship only when quality demands it
05 - Public API vs enterprise / corporate pricing
| Public API Pay-as-you-go |
Enterprise / Corporate Negotiated contract |
|
|---|---|---|
| Who it's for | Developers, startups, individuals - anyone with a credit card | Mid-to-large organisations with significant or committed AI spend |
| Pricing | Fixed, published rates - same for everyone | Negotiated and confidential - typically 20–50% below list price |
| Discounts | Batch processing (−50%) and prompt caching (−90%) only | Volume discounts, committed-use agreements, multi-year deals |
| Contract | None - stop anytime, no minimums | Annual or multi-year commitment, guaranteed minimum spend |
| Cloud route | Direct from provider (Anthropic, OpenAI, Google, Perplexity) | Via AWS, Azure, or Google Cloud - apply existing enterprise agreements |
| Support | Self-serve documentation and community forums | Dedicated account manager, priority SLA, onboarding assistance |
| Data & privacy | Standard terms of service - data may be used to improve models | Custom DPA, data residency options, opt-out from model training |
| When to use | Prototyping, low-to-medium volume, variable workloads | Production at scale (>$50k/year spend), compliance-sensitive use cases |
Procurement tip: If your organisation expects meaningful AI spend, contact each provider's enterprise sales team directly. Real corporate rates are never published. For organisations already on AWS, Azure, or Google Cloud enterprise agreements, accessing AI models through those platforms can unlock existing committed-spend discounts with no new contract needed.