What a token means
A token is a small piece of text, such as a word or part of a word. As a rough guide, 1,000 tokens represent about 750 English words. Code, numbers and other languages can produce different counts. Use measured usage when preparing a production budget.
Material sent to the model: instructions, documents, conversation history and tool results.
Material generated by the model. Billed generation may include reasoning as well as the visible response.
Repeated context a provider can reuse at a lower read rate. Cache creation and storage may have separate charges.
The amount of information a model can work with at once. It is a capacity limit, not free allowance.
Subscriptions and API usage
A staff subscription buys access to an application under its plan limits. An API lets your own software call a model and is typically metered by usage. Budget separately for staff seats, custom integrations and any agent platform fees; check what the proposed contract actually includes.
Definitions and billing categories are described in the provider documentation: OpenAI token guide ↗.
Published token prices
All amounts below are US dollars per one million tokens. These are direct provider API rates for standard processing and the lower context tier where relevant. Taxes, currency conversion, enterprise discounts and additional tool fees are excluded. The models do not have identical capabilities.
| Provider | Model | Input | Cache read | Output |
|---|---|---|---|---|
| OpenAI | GPT-6 Luna | $0.1 | $0.01 | $0.5 |
| OpenAI | GPT-6.1 Sol | $2 | $0.1 | $10 |
| OpenAI | GPT-6 Astra | $10 | $1 | $50 |
| Anthropic | Claude Haiku 4.5 | $1 | $0.1 | $5 |
| Anthropic | Claude Sonnet 5.5 | $2 | $0.2 | $10 |
| Anthropic | Claude Opus 5.5 | $4 | $0.2 | $20 |
| Anthropic | Claude Fable 5.1 | $10 | $0.25 | $50 |
| Gemini 3.5 Flash-Lite | $0.3 | $0.03 | $2.5 | |
| Gemini 3.8 Flash promotional | $0.75 | $0.075 | $3.75 | |
| Gemini 3.1 Pro Preview | $2 | $0.2 | $12 |
Pricing conditions: Google lists the Gemini 3.8 Flash promotion through 31 December 2026, after which input and output rates rise to $1.50 and $7.50 per million tokens. Gemini 3.1 Pro pricing increases above 200,000 input tokens and Google lists cache storage charges.
OpenAI publishes separate long-context rates. Its displayed standard GPT-6.1 Sol cache-write rate is $2.50 per million tokens, distinct from the $0.10 cache-read rate. Anthropic separates cache writes by duration: Sonnet 5.5 five-minute writes cost $2.50 per million tokens and one-hour writes cost $4.00, compared with $0.20 for cache reads.
How to read the table: a lower token rate does not guarantee a lower cost per completed task. Compare observed results on the same business tasks, including retries, escalation and human correction.
OpenAI API pricing ↗ · Anthropic API pricing ↗ · Google Gemini API pricing ↗
Turn token rates into a budget
For an uncached request, divide input tokens by one million and multiply by the input rate. Do the same for billed output tokens, then add cache and tool charges separately when they apply.
At 10,000 tasks, that example is $300 in model charges. The illustrative monthly scenarios below use 10,000 tasks, with no caching or extra tool charges.
| Model | Document task | Agent workflow |
|---|---|---|
| GPT-6 Luna | $15.00 | $150.00 |
| GPT-6.1 Sol | $300.00 | $3,000.00 |
| GPT-6 Astra | $1,500.00 | $15,000.00 |
| Claude Haiku 4.5 | $150.00 | $1,500.00 |
| Claude Sonnet 5.5 | $300.00 | $3,000.00 |
| Claude Opus 5.5 | $600.00 | $6,000.00 |
| Claude Fable 5.1 | $1,500.00 | $15,000.00 |
| Gemini 3.5 Flash-Lite | $55.00 | $550.00 |
| Gemini 3.8 Flash promotional | $112.50 | $1,125.00 |
| Gemini 3.1 Pro Preview | $320.00 | $3,200.00 |
Document task: 10,000 input + 1,000 output tokens. Agent workflow: 100,000 input + 10,000 output tokens across its steps. These are price calculations, not performance benchmarks.
The harness can multiply the bill
A harness is the software around a model that manages context, tools, permissions and the sequence of work. It influences how many model calls run, what each call reads, whether another agent is involved and when the process stops.
Resending a 20,000-token policy at each of ten steps creates 200,000 input tokens per workflow. At 10,000 workflows and $2 per million input tokens, that policy alone costs $4,000 without caching. Retrieving only relevant passages or reusing eligible cached context can reduce that component.
Retrieve relevant passages instead of replaying entire histories and files.
Set budgets for turns, retries, tools, elapsed time and delegated agents.
Reuse repeated context and unchanged work when freshness rules allow it.
Measure correction time alongside model quality and API spend.
Also budget for application seats, paid searches, hosting, data retrieval, integration work, monitoring and human review. Keep one-off implementation costs separate from recurring running costs. For a UK budget, use an explicit planning exchange rate and account for taxes separately.
A practical evaluation checklist
- Choose a frequent workflow and define what counts as an accepted result. Include difficult cases.
- Test providers on the same tasks. Record tokens, tool fees, elapsed time, errors, retries and review minutes.
- Compare total pilot cost with accepted results. Include failed attempts in the spending total.
- Set spending alerts and limits on steps, retries and delegated agents. Define actions that need approval.
- Model normal and busy months. Check high-cost outliers, promotion expiry dates and demand changes.
The decision to makeSelect the lowest total operating cost that meets the required quality and turnaround time. Start with a controlled pilot, then expand once cost per accepted result and review workload are understood.
Official sources checked 30 September 2026. Recheck rates before procurement or a production launch. Subscriptions, API usage, staff seats, custom integrations, hosting, retrieval, monitoring and taxes should be budgeted separately.
[1] OpenAI tokens ↗ · [2] OpenAI pricing ↗ · [3] Anthropic pricing ↗ · [4] Google pricing ↗