AI API Cost Calculator: Compare GPT, Claude, Gemini & Grok

A privacy-first planning tool for developers and serious AI users · Single-request workflow reviewed October 5, 2026 · Workflow budget added October 10, 2026 · Provider rates have separate review dates

API pricing looks simple until input tokens, output tokens, cached context, and request volume start multiplying. Use this page to estimate a realistic monthly cost before choosing a model or committing to a provider.

The short answer

Use Token counter for a quick, local estimate of a prompt’s input size.
Use Monthly API cost when you know your average prompt, response length, and monthly request count.
Use Subscription vs API when you want to compare metered usage with a fixed monthly plan.
Plan a multi-step job in Workflow budget, with separate models and billed calls for each stage.
Already have usage records? Calculate cost per accepted result, including money spent on retries and an optional value for review time.
These are planning estimates. Confirm the provider’s current pricing, limits, taxes, and product terms before spending money.

HiseHub AI Cost Lab

Estimate AI API costs before the bill arrives

Count a prompt locally, model a monthly workload, and compare GPT, Claude, Gemini, and Grok API costs. The calculator is an estimate—not a provider invoice—and it never asks for an API key.

Model / billing route Input tokens / request Output tokens / request Requests / month Cached input share

Use average tokens, not characters. For GPT-6.1 Sol and GPT-6 Sol/Luna, more than 272,000 input tokens applies the long-context rate to the full request; cache share means cache reads only.

Grok 4.7 uses full-request long-context rates at 200,000 input tokens or more, before subtracting cache hits. Global and US options use the same model; the US endpoint adds 10% to all token rates.

Grok search fees are not included. Web Search counts calls; X Search counts fetched posts and profiles, including repeats. See the search-fee units and worked example. This targeted search-billing note was checked October 8, 2026; token-price review dates below remain separate.

Enter total billed output, including thinking/reasoning. Gemini 2.5 Pro switches above 200,000 input tokens (not at exactly 200,000). Gemini cache storage is a separate charge; it is not inferred from cache-hit share.

Load a worked example

Plan an AI task with several model calls

One task is the whole job, not one API call. Separate extraction, retries and synthesis when their models or prompt sizes differ. Enter all billed input and output per call, including resent history, tool definitions and billed thinking. This is a text-token plan, not a quality test or an enforced spending cap.

Tasks / month

Stage 1

Stage name (optional) Stage model / billing route Stage input tokens / call Stage output tokens / call Stage calls / task, including retries Stage cache-read share (%)

Other USD costs / task (optional) Fixed USD costs / month (optional)

Optional costs can include tools, runtime, cache writes/storage or hosting. Use amounts already in USD and count each charge once. Blank means unknown, not free. Enter 0 only for a known zero. Tax, special tiers and unentered costs remain outside the subtotal.

Examples are invented workloads, not provider runs, measured retry rates or equal-quality comparisons. Change them to match your actual logs.

Enter a monthly task count and each stage. Nothing is sent to a model or saved by this calculator.

Long-input prices and context/output limits are checked per call, before monthly multiplication. Cache-read share does not model cache writes or retention. Runtime and concurrency are not inferred from token counts. For provider budget limits and a worked111-call example, read the workflow cost guide. After real jobs finish, use the cost-per-accepted-result worksheet.

Paste a prompt or text (processed in this browser)

Model / route for a one-request estimate Expected output tokens

Model / billing route Input tokens / request Output tokens / request Requests / month Subscription price / month

Cached input share

This compares metered API spend with a fixed subscription price. A subscription’s usage caps, included features, and product licensing still need a separate check.

How to use this estimate

Useful for planning, not for pretending every provider bills the same way

Input and output prices are separate. Long answers can cost much more than short ones even when the prompt is unchanged.
Cache estimates are optional. Enter the share of input tokens served from cache only when the provider and model publish a cache rate.
Token counts vary by tokenizer. The browser counter is deliberately labelled an estimate; use the provider’s own token endpoint for a production invoice check.
Tools and special tiers can add cost. Batch, priority, reasoning, image, video, cache-write, and server-side tool charges may not be represented in a simple text estimate.

Workload costs compare documentation-reviewed billing entries only. Unreviewed benchmark-cache rows remain separate headline references, without monthly-cost or break-even quotes. The cache was synchronized October 27, 2026 06:37 UTC; that synchronization does not re-review official prices. Sources and actual review dates are shown below. These are manual documentation snapshots, not live API/account tests. The 72-hour cache freshness guard still pauses money estimates if the cache is stale.

Google’s published Gemini 3.8 Flash promotion ends December 31, 2026. This page uses the documented rate schedule based on the server’s UTC date, not an ongoing price collector. Reload around the transition and check Google’s current pricing before budgeting. A retained explicit cache adds retained tokens × retained hours × the storage rate / 1,000,000. Cache-hit share does not specify retention time and does not guarantee a cache hit.

At exactly 200,000 Grok input tokens, this estimate follows xAI’s main pricing table (≥200K); abbreviated model and regional prose says “above/exceed.” Check the invoice at that boundary. US inference does not guarantee US handling of every file or tool; cache reuse across endpoints is not guaranteed. Regional endpoint details.

Using Grok? Separate API billing from Grok 4.7 Fast and consumer access, or compare SuperGrok with X Premium+.

Worked examples: compare the Opus 5.5 / Sol / Luna cost cases. Need the consumer-plan side? Compare AI subscription prices by country, use the AI Plan Finder, or review AI model capability and value.

Sources, rates and boundaries

Official billing rates used by this calculator

USD per million billed text tokens, paid Standard tier. These are token subtotals, not total invoices or capability rankings. Cached-input rates apply to actual cache reads; writes, explicit storage, tools and tax are separate.

Model / route	Input / cache read / output	Conditions and official sources
GPT-6.1 Sol (API billing)gpt-6.1-sol	$2 / $0.1 / $10 Above 272,000 input: $4 / $0.2 / $15	OpenAI Standard text-token rates. More than 272K input tokens prices the full request at the long-context rates. Cache reads cost 5% of the input rate. Reviewed 2026-10-01 · Official pricing
GPT-6 Sol (API billing)gpt-6-sol	$2 / $0.2 / $10 Above 272,000 input: $4 / $0.4 / $15	OpenAI Standard text-token rates. More than 272K input tokens prices the full request at the long-context rates. Reviewed 2026-10-01 · Official pricing
GPT-6 Luna (API billing)gpt-6-luna	$0.1 / $0.01 / $0.5 Above 272,000 input: $0.2 / $0.02 / $0.75	OpenAI Standard text-token rates. More than 272K input tokens prices the full request at the long-context rates. Reviewed 2026-10-01 · Official pricing
GPT-6 Astra (API billing)gpt-6-astra	$10 / $1 / $50 Above 272,000 input: $20 / $2 / $75	OpenAI Standard text-token rates. More than 272K input prices the entire request at long-context rates. Include billed reasoning in output; cache writes and special processing tiers are separate. Reviewed 2026-10-05 · Official pricing · Model limits / access
Claude Haiku 5.5 (API billing)claude-haiku-5-5	$0.1 / $0.01 / $0.5 Above 100,000 input: $0.5 / $0.05 / $2.5	Claude API Standard text-token rates. Above 100,000 total prompt tokens, use the higher rates for the request, including output and cache reads. 5-minute writes shown; 1-hour writes are $0.20 short / $1 long per 1M. Include billed thinking in output and recount older Haiku prompts with the new tokenizer. Batch, clouds and tools differ. Reviewed 2026-10-08 · Official pricing · Model limits / access
Claude Sonnet 5.5 (API billing)claude-sonnet-5-5	$2 / $0.1 / $10	Claude API Standard text-token rates. Cache reads fell to $0.10 per million on October 7, 2026; base input/output and cache-write rates did not change. 5-minute writes shown. Cloud-channel, Batch, 1-hour caching and tools differ; include billed reasoning in output. Reviewed 2026-10-08 · Official pricing
Claude Opus 5.5 (API billing)claude-opus-5-5	$4 / $0.2 / $20	Claude API Standard text-token rates. Provider, cloud-channel, batch, and fast-mode pricing can differ. Reviewed 2026-10-01 · Official pricing
Gemini 3.8 Flash (API billing)gemini-3.8-flash	$0.75 / $0.075 / $3.75	Gemini Developer API paid Standard text-token rates. Output includes billed thinking. Cache reads do not include explicit cache storage; grounding, other modalities and special tiers are separate. Explicit cache storage: $0.5 per 1M token-hours (excluded). Current promotion through 2026-12-31; from 2027-01-01: $1.5 / $0.15 / $7.5; storage $1. Reviewed 2026-10-05 · Official pricing · Model limits / access
Gemini 3.5 Flash-Lite (API billing)gemini-3.5-flash-lite	$0.3 / $0.03 / $2.5	Gemini Developer API paid Standard text-token rates. Include billed thinking in output. Explicit cache storage and grounding are not included in the token subtotal. Explicit cache storage: $1 per 1M token-hours (excluded). Reviewed 2026-10-05 · Official pricing · Model limits / access
Gemini 2.5 Pro (legacy API access)gemini-2.5-pro	$1.25 / $0.125 / $10 Above 200,000 input: $2.5 / $0.25 / $15	Legacy access: Google limits 2.5 models to users who actively used them previously. More than 200K input applies long rates to the whole request. Include billed thinking in output; explicit cache storage is separate. Explicit cache storage: $4.5 per 1M token-hours (excluded). Reviewed 2026-10-05 · Official pricing · Model limits / access
Grok 4.7 — global APIgrok-4.7	$2 / $0.5 / $6 At least 200,000 input: $4 / $1 / $12	Standard API text-token estimate, not Grok 4.7 Fast. Include billed reasoning in output. Follow the main pricing table at exactly 200,000 input tokens; verify boundary charges against the invoice. Hosted search fees are excluded; X Search counts fetched posts and profiles, not calls. Reviewed 2026-10-05 · Official pricing
Grok 4.7 — US API (+10%)grok-4.7	$2.2 / $0.55 / $6.6 At least 200,000 input: $4.4 / $1.1 / $13.2	Same grok-4.7 API model, a different billing route—not a capability upgrade. Include billed reasoning in output. Standard rates only; not Grok 4.7 Fast. Hosted search fees are excluded; X Search counts fetched posts and profiles, not calls. Reviewed 2026-10-05 · Official pricing

From token prices to usable work

Calculate cost per accepted AI result

A cheap API call is not necessarily a cheap finished task. Use your own spending and task records to compare two workflows, including retries, rejected answers and review effort. This worksheet does not use model rankings or the price catalogue above.

Define the unit before you test. Count each original task once, with at most one accepted final result per task. Set the acceptance rule first and use the same task set, period and rule for both workflows. API calls include all calls in a multi-step workflow and retries; they are not the task count.

Currency for all amounts Review-time value per hour (optional)

No currency conversion is performed. Enter amounts already in the selected currency. Time valuation is an optional planning assumption, not an additional invoice charge; do not count wages twice if they are included in other cash spend.

Workflow A

Original tasks tested Accepted final results Total API & tool spend Other allocated cash spend

Include billed retries, discarded answers and tool charges in the API total. Other cash can include allocated hosting or search costs, but not charges already counted in the API total. Enter an explicit 0 where your records show no spend.

Optional: API calls and review time

API calls, including retries Total review & editing minutes

Workflow B

Original tasks tested Accepted final results Total API & tool spend Other allocated cash spend

Optional: API calls and review time

API calls, including retries Total review & editing minutes

Enter both workflow records. Blank amounts are unknown, not zero. Nothing is prefilled from a model catalogue.

Formula, interpretation and limits

Cash cost per accepted result = (API & tool spend + other allocated cash spend) ÷ accepted final results. Optional time-inclusive cost adds review minutes ÷ 60 × your hourly time value to the numerator. All spending belongs to the full run, not just its successful attempts.

With zero accepted results, the unit cost is undefined, not zero. Acceptance share is accepted results ÷ original tasks tested; it describes your stated rule and sample, not a universal model-quality score. Calls per task can exceed one because of multi-step work as well as retries.

The worksheet does not fetch bills, verify your acceptance decisions, predict future failure rates or declare a winning model. Unequal task sets, missing cash allocations, quality differences, latency and severe errors can make a lower unit cost a poor choice. Use the same cost boundary and test representative tasks before acting.

Calculation happens in this browser. This worksheet does not send the entered numbers to HiseHub, save them in browser storage or add them to share links. It needs no prompts, keys or invoice uploads. Existing site analytics still applies; record your results privately before leaving or clearing the page.

Related reading: FinOps Foundation: Unit Economics. This is HiseHub’s independent worksheet, not a certified or endorsed benchmark.

Estimate a multi-step AI workflow, not just one request

Use Workflow budget when one job contains several calls or models. Add a stage for extraction, another for retries with their own prompt shape, and another for synthesis. Enter input and billed output per call, calls per whole task and tasks per month. The calculator applies each route’s token limits and long-input tier before multiplying; it never treats monthly totals as one prompt.

The document example assumes 100 extraction calls, 10 retries and one synthesis, all on Sonnet 5.5: $1.05 in tokens per task, $31.50 for 30 tasks. The mixed-model example changes only extraction to Haiku 5.5, giving $0.195 per task / $5.85 per month under the entered assumptions. Neither example establishes equivalent quality, an actual retry rate or API access. Read the worked workflow cost guide for the arithmetic, runtime fees and provider-specific budget limitations.

Optional USD extras remain unknown until entered. Count tools, runtime, cache writes/storage and fixed fees only once; known subtotal is not an all-in bill. After a real workflow finishes, use cost per accepted result to include failed work and review effort. Prospective stage budgeting and measured useful-result costing answer different questions.

Why API cost is harder than a single price per million tokens

Most providers publish separate rates for input and output tokens. A request that sends a long system prompt and receives a short answer has a different cost profile from a short question that produces a long explanation. In a chat application, the conversation history can also be sent again on every turn, so the amount of input grows even when the user’s latest message is small.

Some providers add more dimensions. Prompt caching can reduce the price of repeated context, while batch processing trades immediate responses for a discount. Reasoning tokens, image or video input, server-side search, and other tools may be billed separately or counted differently. A useful calculator therefore needs to show its assumptions instead of presenting one false “exact” number.

How the monthly estimate works

For a text-only request, the basic calculation is:

input cost  = input tokens × input price ÷ 1,000,000
output cost = output tokens × output price ÷ 1,000,000
monthly cost = (input cost + output cost) × requests per month

If you enter a cached-input share and the model has a cached-input rate in the current data, the tool applies that lower rate to the selected share of input tokens. It does not assume that all context is cacheable, and it does not hide a missing cache rate behind a guessed discount.

What the result does—and does not—tell you

Check the processing tier. The manually verified billing rows use Standard text-token rates; the calculator does not automatically apply Fast, Ultrafast, Batch or regional uplifts. For identical GPT-6.1 Sol usage—100,000 input tokens, 90% verified cache reads and 2,000 billable output tokens—the current API token subtotal is $0.049 Standard, $0.098 Fast or $0.294 Ultrafast. Earlier cache writes, tools, residency uplifts and tax are excluded. See processing-tier prices, eligibility and the worked example. This tier comparison was checked October 10, 2026; other provider review dates remain separate.

It does: turn a workload assumption into a comparable per-request and monthly estimate across representative GPT, Claude, Gemini, and Grok models.
It does: show a break-even request volume when a fixed subscription price is compared with the selected API workload.
It does not: promise that a consumer subscription includes API credits. Chat products and developer billing are normally separate.
It does not: account for every provider-specific fee, tax, quota, retry, tool call, rate-limit error, or application hosting cost.
It does not: measure answer quality. Use the AI Model Comparison for capability, speed, and benchmark context.

How to estimate your own workload without guessing wildly

Take a representative week of requests rather than using your busiest day.
Separate the reusable system prompt or reference context from the user’s changing message.
Measure typical output length; a long answer can cost several times more than a short answer on the same model.
Multiply by actual monthly requests, including automated retries and scheduled jobs.
Run a low, typical, and high scenario. The high scenario is more useful for setting a spending alert than a single optimistic average.

For a production service, compare the estimate with the usage fields returned by your provider’s SDK or API. That real usage should become your source of truth after launch; this page is a planning aid before and between deployments.

Upgrading a Haiku API integration? Our Haiku 5.5 migration troubleshooting guide separates invalid request parameters, missing response text and token-budget surprises. Recount the actual prompt with the new model before treating this calculator’s workload assumptions as your usage.

API or subscription: which route fits?

A subscription is usually easier for a person who wants a finished interface, file uploads, memory, voice, or a predictable monthly charge. An API is the correct route for an application, automation, or service that sends requests on behalf of multiple users. The cheapest option depends on workload, not only the headline price.

The calculator’s break-even result is only one input to that decision. Check whether the subscription’s usage caps fit your work, whether the product terms allow your intended use, and whether the API needs extra hosting, logging, retrieval, or moderation costs. When the result is close, a hybrid setup—a personal subscription plus API billing for a product—may be more practical than forcing one route to do both jobs.

For the decision behind the math, read AI Subscription vs API: Which Is Cheaper for Your Actual Usage? and compare its guidance with your own workload.

Privacy and data handling

The token counter estimates text in your browser. HiseHub does not ask you to paste an OpenAI, Anthropic, Google, or xAI API key, and the calculator does not need to upload your prompt to calculate the displayed estimate. Do not paste secrets, personal records, or confidential customer data into any online tool unless you have reviewed its data handling first.

Official references

OpenAI tiktoken — the official open-source tokenizer used with OpenAI models.
Anthropic Count tokens in a Message — counts message input including supported tools, images, and documents.
Google Gemini countTokens API — documents token counting for text and supported media inputs.
xAI API pricing — lists model rates and special pricing such as cached, batch, and priority processing.

HiseHub is an independent reference site, not an affiliate or representative of OpenAI, Anthropic, Google, xAI, or X. Provider prices, model names, limits, and terms can change. The calculator uses the latest server-side comparison data available to HiseHub and labels its output as an estimate.

Continue planning your AI setup

Compare model capability and value, check consumer subscription prices by country, or use the AI Plan Finder when you are choosing a paid chat plan. If a paid entitlement is missing, start with the AI Subscription Troubleshooter rather than buying a second plan.

If the API estimate is part of a larger system, compare local runtimes, chat interfaces, and workflow platforms in the Self-Hosted AI Tools directory before adding hosting and maintenance costs.

New billing choices: GPT-6.1 Sol and Sonnet 5.5 are consolidated API billing entries, not duplicate reasoning-effort rows. Compare GPT-6.1 Sol cache and long-context costs or Sonnet 5.5 / Opus 5.5 cost examples. OpenAI input limits are checked separately from total context capacity.

AI API Cost Calculator & Token Counter