Free tools Windows power users keep installed
One-click scans. No signup required.
OpenAI API cost depends on the model, the amount of input and output it processes, and any tools or modalities attached to the request. There is no single API price. Many models quote separate rates for input, cached input, cache writes, and output, while web search, file search, voice, images, audio, and containers can use different units or additional charges. Your organization’s rate limits and spending controls are separate from those prices.
The figures and labels below reflect OpenAI documentation available on October 13, 2026. Model availability, prices, tiers, and limits can change, so check the live pricing page and your organization’s Settings > Organization > Limits page before budgeting or deploying.
As an Amazon Associate I earn from qualifying purchases.
What you pay for when using the OpenAI API
For a token-priced request, the bill is driven by the model and by usage categories. Input is the content you send; output is what the model generates. Depending on the model and endpoint, OpenAI can also list cached-input and cache-write rates. Reasoning tokens generated internally by reasoning models count as output usage even when they are not displayed in the final answer.
A useful cost model is:
(input tokens ÷ 1,000,000 × input rate) + (cached-input tokens ÷ 1,000,000 × cached-input rate) + (cache-write tokens ÷ 1,000,000 × cache-write rate) + (output tokens ÷ 1,000,000 × output rate)
#1 Best Overall
Add any separate tool or modality charges. Not every endpoint uses every category, and the applicable rate depends on the selected model, context length, and service mode. Rates are normally quoted per one million tokens, so divide each token count by 1,000,000 before multiplying.
Input, output, and cached input
- Input tokens: your prompt, conversation history, system instructions, tool definitions, schemas, files, and other supplied content.
- Output tokens: generated text and, for reasoning models, internal reasoning usage counted by the API.
- Cached input: reusable input served through prompt caching. It can have a lower rate than ordinary input when the model and workload support it.
- Cache writes: tokens written into a cache can have their own charge.
Because output rates can be much higher than input rates on a given model, a long generated report may cost more than a much larger prompt. Conversely, a huge repeated system prompt may make caching valuable. The pricing table for the model you select is the authority for the actual rates.
Tools and modalities are not always token-only
Some capabilities add charges that are not calculated simply as text tokens. OpenAI’s 2026 pricing documentation lists these examples:
| Capability | Published charge | Qualification |
|---|---|---|
| Web search | $10.00 per 1,000 calls | OpenAI, 2026; search-content tokens are also billed at the selected model’s rates. |
| File Search storage | $0.10 per GB per day | OpenAI, 2026; 1 GB is free. |
| File Search tool calls | $2.50 per 1,000 calls | OpenAI, 2026; applies to the Responses API. |
| GPT-Live 1 voice sessions | $0.05 per minute | OpenAI, 2026; billed by the second without rounding up to a whole minute. Backend model and tool usage is separate. |
Images, audio, transcription, containers, and other tools can use their own model rates, call charges, storage fees, or time-based units. Include those items in a budget instead of assuming every API operation has one universal token price.
Rank #2
- Used Book in Good Condition
Why a short visible answer can still use many tokens
A token is not the same as a word. OpenAI describes a token as a character, part of a word, a whole word, punctuation, or (in some cases) a space. Tokenization depends on the model’s encoding, spelling, spaces, and language. For English, “about four characters” or “roughly three-quarters of a word” is only a rough estimate; it is not a billing calculation, and other languages can tokenize very differently.
Hidden and structural usage
The text you see in the final answer is only part of a request’s accounting. A request may also include:
- Conversation history and system instructions repeated on every call.
- Tool descriptions, function schemas, and structured-output definitions.
- Images, files, or other supported inputs.
- Reasoning tokens produced internally by a reasoning model.
- Search-result content or other tool-returned material inserted into the model context.
That is why estimating cost from the visible answer’s character count often fails. Two models can tokenize identical text differently, and they can generate different amounts of output or internal reasoning. A lower advertised price per million tokens does not automatically produce a lower cost for your complete task.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →How to count before sending
- Use OpenAI’s Tokenizer to inspect representative text interactively.
- For plain text in code, use the
tiktokenlibrary with the encoding associated with the target model. - For a complete Responses API input, use OpenAI’s input-token counting API. It accounts for supported formatting and inputs such as images, files, tools, and conversations, which plain-text counting cannot fully represent.
- After production calls, compare the estimate with the response’s usage object and your dashboard data.
How to compare model prices meaningfully
Do not rank models by one input-rate column. Compare the dimensions that determine your workload’s end-to-end bill.
| Comparison question | Why it changes cost |
|---|---|
| Which model and family? | Each model can have different input, cached-input, cache-write, and output rates. |
| What is the input/output mix? | Output is often priced differently from input, so a chatty or report-generating task can have a different cost profile from a classification task. |
| Can the request use cached input? | Repeated prompts may qualify for a cached-input rate, while creating the cache can have a separate write charge. |
| Is the context short or long? | Pricing documentation can separate short-context and long-context rates. Use the column matching the request. |
| Which service mode applies? | Service modes can have distinct pricing and availability. Confirm the mode selected by your request. |
| Are tools or other modalities involved? | Search, storage, calls, audio, images, transcription, and containers may add non-token charges. |
| How much reasoning is used? | Internal reasoning tokens count toward usage even if the final response is brief. |
The reliable comparison is a representative workload: send the same realistic inputs to candidate models, record input, cached-input, output, and reasoning usage, add tool charges, and compare the resulting task cost. A per-million-token headline is not a substitute for that measurement.
How usage appears in API responses and the dashboard
Response usage fields
The field names depend on the endpoint:
| Endpoint | Main usage fields |
|---|---|
| Chat Completions | prompt_tokens, completion_tokens, and total_tokens |
| Responses | input_tokens, output_tokens, and total_tokens |
Depending on the model and endpoint, usage details can also expose cached-input and reasoning-token counts. Log the complete usage object, model name, endpoint, tool calls, and request identifier for cost analysis. Never infer usage solely from the displayed answer.
Usage Dashboard timing
The Usage Dashboard reports activity by billing period and project. Its displayed data uses UTC, so a day boundary may not match your local calendar. Align your internal reports to UTC when reconciling dashboard totals with application logs.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesRate limits: what they control and where to see them
Pricing answers “how much each unit costs.” Rate limits answer “how quickly your organization or project may send requests or tokens.” They are separate controls. Limits can apply to requests per minute or day, tokens per minute or day, images per minute, and audio minutes per minute for some streaming models. The first applicable limit you reach can stop a request.
Organization and project scope
Limits are generally defined at organization and project levels rather than as a personal allowance. They vary by model, account tier, endpoint, and sometimes share capacity among models. A model-and-tier example in documentation is not a promise of your current throughput.
To see the configuration that applies to you, open Settings > Organization > Limits in the API platform. Use that page for current model RPM and TPM values instead of copying a tier example from an older article.
A specific endpoint example
OpenAI documents a limit of 300 requests per minute per vector store ID for the specified vector-store file endpoints. This is an endpoint-specific limit, not a general API request quota and not a guarantee for other vector-store operations.
Usage tiers, monthly approval, and spend controls
Three controls are often confused:
- Rate limits: short-term throughput such as requests or tokens per minute.
- Approved monthly usage: the organization’s permitted monthly API usage level.
- Spend controls: alerts and hard limits configured for an organization or project.
OpenAI documentation gives these usage-limit examples for 2026: $100 per month for Free, $500 for Build, $5,000 for Launch, and $200,000 for Grow. They are documentation examples, not a promise of your account’s cap. The Free tier requires an allowed geography, while paid-tier qualification is based on total credit purchases.
Best Value
A spend alert notifies you without stopping traffic. A hard spend limit can cause affected API requests to return HTTP 429. Check the live organization settings for the value and scope that actually apply to your projects.
Diagnosing a 429 error
OpenAI states that a 429 response can indicate a temporary rate limit, an exhausted prepaid balance, or a spending or usage limit. Read the error code and message before deciding to retry.
Temporary request or token limit
- Check whether the response includes a
Retry-Aftervalue. - Wait at least the stated period before sending another request.
- If the header is missing or invalid, retry with exponential backoff and jitter.
- Cap both the retry count and total retry duration.
- Reduce concurrency, request size, or generation length if you repeatedly reach the same RPM or TPM ceiling.
Unsuccessful requests can count toward per-minute limits. A tight loop that resends the same failing request may therefore prolong the outage.
Prepaid balance, usage, or spend limit
Retrying does not restore access when the account has exhausted prepaid credits, reached an approved usage limit, or hit a project or organization hard spend limit. Follow the message: add the required balance, request or qualify for a higher usage limit, or change the configured cap. A hard limit is a billing control, not a transient network condition.
How to prevent recurring 429s
- Use a queue that enforces your documented RPM and TPM ceilings.
- Apply exponential backoff with random jitter rather than fixed-interval retries.
- Track tokens as well as request counts; a few large prompts can exhaust TPM while RPM remains available.
- Separate workloads by project when your organization’s controls allow it.
- Alert on balance, usage, and spend separately so a billing cap is not mistaken for throttling.
A practical budgeting workflow
- Define the workload: count calls, average and worst-case input, expected output, tools, files, images, and languages.
- Select candidate models: record the exact model, context class, and service mode.
- Measure token usage: use the Tokenizer or counting API for estimates, then capture real response usage.
- Separate categories: calculate ordinary input, cached input, cache writes, output, and reasoning usage independently.
- Add non-token charges: include search calls, file storage, tool calls, audio time, image operations, or other applicable units.
- Model peaks: estimate concurrent requests and tokens per minute, not only monthly averages.
- Set controls: configure alerts and hard limits deliberately, knowing that hard limits can produce 429 responses.
- Reconcile monthly: compare UTC dashboard totals with application logs and investigate differences by model, project, and endpoint.
API billing is separate from ChatGPT subscriptions
ChatGPT subscription pricing and API usage are different products and billing systems. An individual or business subscription does not establish a particular API token price, monthly allowance, or rate limit. For API budgeting, use the model pricing table, response usage fields, Usage Dashboard, and organization Limits page.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
