The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →There is no permanent cheapest API. OpenAI, Anthropic Claude and Google Gemini publish model-specific rates, so your bill depends on the model tier, input/output mix, cache hits, processing mode, geography, tools and retries. The reliable way to compare them is to replay the same representative workload, apply each provider’s current rate-card definitions, and measure cost per accepted result—not just cost per million tokens.
Prices change frequently. The provider pages reviewed on 13 October 2026 should be treated as a dated snapshot; refresh the live rows immediately before budgeting or publishing a purchasing decision.
As an Amazon Associate I earn from qualifying purchases.
What an API price actually measures
Cloud AI APIs normally meter tokens rather than charging a single price per request. A request can contain input tokens (your prompt, conversation history and tool results), output tokens (the model’s response), and, depending on the service, cached reads, cache writes, thinking tokens, search queries or code-execution time. Two applications making the same number of calls can therefore have radically different bills.
Model names are not interchangeable units of capability. A fast, lower-cost model may need more retries or produce less usable output than a more capable model. Compare models that can perform the same task, then record quality and rework. Public rate tables are model-specific prices, not an independent market statistic.
Use the official live documentation for the three services: OpenAI API pricing, Claude Platform pricing and Gemini API pricing.
The cost categories to put in one comparison
| Category | What to record | Why it changes the result |
|---|---|---|
| Model and capability tier | Exact model ID, context limit and selected quality tier | A cheaper model may not complete the same work or may require more attempts. |
| Input tokens | Uncached input tokens for each request | Long system prompts and retrieved documents can dominate spend. |
| Output tokens | Billable answer tokens; check whether reasoning is included | Output rates are separate and often higher than input rates. |
| Prompt-cache reads | Tokens served from an existing cached prefix and the provider’s read rate | Repeated prefixes can lower input cost, but cache mechanics differ. |
| Cache writes | Tokens written, storage duration or time-to-live, and write rate | Writing a cache can cost more than reading it and may expire. |
| Processing mode | Standard, batch, priority or speed mode | The same model can have different rates by mode. |
| Geography and route | Region, residency option and direct-versus-marketplace endpoint | Regional processing or a cloud marketplace can add a premium. |
| Tools and media | Search, grounding, code execution, images, audio, storage and other meters | These charges sit outside ordinary token arithmetic. |
| Retries and tool loops | Failed calls, validation retries and follow-up tool requests | Every unsuccessful attempt still consumes capacity and may be billed. |
A reusable API cost formula
For a workload whose provider rate card separates cache categories, calculate each vendor independently:
(uncached input tokens × input rate + cached input tokens × cache-read rate + cache-write tokens × cache-write rate + billable output tokens × output rate) ÷ 1,000,000
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Then add request, tool, storage and processing-mode charges that the provider lists. Keep the units attached to every term. “Input rate” must be the rate for the exact model, region and mode; do not substitute another model’s number.
For a monthly estimate, multiply the per-request result by expected successful calls, then add failed attempts and retries as separate call groups. If 8% of calls retry once, for example, model those additional calls explicitly instead of hiding them in a rounded request count. Providers tokenize text differently, so use the token counts reported by each API rather than assuming one vendor’s tokenizer produces the same count as another’s.
OpenAI API pricing details to verify
The OpenAI rate card itemizes model rates for input, cached input, cache writes and output where applicable. Select the exact model row and copy each category into your worksheet; there is no single “OpenAI API price.”
Regional-processing uplift
OpenAI states that eligible regional-processing endpoints for models released on or after 5 March 2026 carry a 10% uplift. This applies only when both the model and endpoint are eligible, so do not add 10% to every OpenAI request. Record the endpoint’s geography and eligibility beside your rate.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsPrompt caching behavior
OpenAI says supported calls with prompts longer than 1,024 tokens automatically benefit from prompt caching. The cached prefix starts at 1,024 tokens and grows in 128-token increments. For a repeated system prompt or document prefix, split your estimate into cached and uncached input using the separate rates shown for the model. A changing prefix, short prompt or low hit rate may leave most tokens at the ordinary input rate. More implementation detail is in OpenAI’s prompt-caching documentation.
Time-limited promotion
The pricing page says promotional pricing for GPT-5.6 Sol is available at least through 21 November 2026. Treat that as a dated promotion for that model, not a permanent OpenAI price or a discount that applies to other models.
Anthropic Claude pricing details to verify
Anthropic’s pricing page lists model-specific input and output prices per million tokens and separate cache-read and cache-write prices. Copy the model row that matches your task, then state whether your estimate assumes a five-minute cache lifetime: the displayed prompt-cache pricing reflects a five-minute TTL.
US-only inference
Anthropic states that US-only inference is available at 1.1 times the normal input and output pricing. Apply that multiplier only when your deployment selects US-only inference, and check how cache categories are treated in the current table rather than assuming the same multiplier for every meter.
Recommended Free Tools
Platform features
The Claude platform also lists feature charges such as web search and code execution. Add them only if the workload invokes those features. A text-only benchmark should not include a search fee, while an agent that searches on most turns must record the number of searches and their billed units.
Google Gemini pricing details to verify
Google’s Gemini API page separates models and billing modes, including standard and batch rates where available. It also lists context-cache and Google Search grounding charges. Confirm the exact model table, mode and free-tier eligibility for your project; a generic “Gemini price” is not precise enough for a budget.
Thinking tokens
Google’s standard rates state that the output price includes thinking tokens. When comparing a reasoning-heavy task, count the provider’s billable output exactly as reported and do not remove thinking tokens from Gemini’s output total while leaving another vendor’s reasoning tokens in its total.
Search grounding
Grounding can result in one or more individual search queries, and Google bills per query performed. Estimate the expected number of queries, not merely the number of model responses. A request that triggers three searches has a different tool cost from one that triggers one search even if both return a single answer.
Worked comparison without inventing a rate
Suppose your evaluation represents one month of support traffic: 100,000 requests, 2,000 input tokens per request, 600 output tokens per request, and a 70% cache hit on a 1,400-token common prefix. Assume 5% of requests retry once, and 10% of successful requests use a search or grounding tool. These are declared workload assumptions, not provider statistics.
| Meter | Calculation for the month |
|---|---|
| Total input tokens | 100,000 × 2,000 = 200,000,000 |
| Cached-prefix tokens | 100,000 × 1,400 × 70% = 98,000,000 (subject to each provider’s cache rules) |
| Uncached input tokens | Total input minus tokens actually served from cache; use each API’s usage record |
| Output tokens | 100,000 × 600 = 60,000,000 before retries |
| Retry calls | 100,000 × 5% = 5,000 additional calls, with their own input and output usage |
| Tool calls | 10,000 requests initially use search or grounding; count actual queries or executions |
For each provider, replace the worksheet’s input, cache-read, cache-write and output-rate cells with the current model row. Add the provider’s search, grounding, code, storage, batch or regional charges. The resulting three totals are auditable because the token assumptions are identical while the billing definitions remain vendor-specific.
Step-by-step method for a fair price test
- Define the accepted outcome. Write the task, quality threshold, latency target, context size and required tools. “Generate text” is too vague; specify what counts as a usable answer.
- Choose comparable models. Select one model from each vendor that meets the capability and context requirements. Record model IDs and the date you captured rates.
- Build a representative sample. Include short and long prompts, normal and edge cases, tool calls, multilingual inputs if relevant, and the real system prompt. Use enough requests to expose retries and cache behavior.
- Measure tokens per provider. Export input, output, cached-read and cache-write usage from each API. Do not convert one provider’s token count into another’s.
- Declare cache assumptions. State hit percentage, prefix length, write frequency and TTL. For OpenAI, verify the 1,024-token threshold and 128-token growth increments; for Claude, account for the displayed five-minute TTL; for Gemini, use its context-cache rules.
- Run the intended processing mode. If production will use batch, priority or speed processing, test and price that mode rather than standard service.
- Apply geography and endpoint choices. Include regional or US-only inference premiums and distinguish a direct API from a marketplace route.
- Add optional features. Count search queries, grounding calls, code execution, image or audio units and storage only when the sample uses them.
- Replay failures. Record validation failures, timeouts, safety refusals that require a second call and agent loops. Include their actual usage in cost per accepted result.
- Compare quality-adjusted totals. Divide total spend by accepted outcomes, then report latency and failure rate beside the dollar figure.
Why cost per accepted result beats list price
Imagine Model A has a lower output rate but passes your evaluator on 88% of attempts, while Model B costs more per token and passes on 97%. If rejected answers trigger human review or a retry, Model A’s effective cost can exceed Model B’s. The correct comparison is:
cost per accepted result = total billed spend, including retries and tools ÷ number of outputs that meet your acceptance test
Use the same evaluator and sample for all vendors. Keep human-review labor separate from API spend, but show it as an operational consequence if it materially affects the decision. A provider with a slightly higher token rate may still be cheaper when it avoids rework.
Batch, priority, speed and deployment trade-offs
Batch processing can reduce a bill for asynchronous jobs, but it changes latency and sometimes eligibility. Priority or speed modes may cost more in exchange for service characteristics. Put the selected mode in the comparison table and do not mix a batch rate for one vendor with a standard rate for another.
Geography can change both price and architecture. OpenAI’s documented 10% regional-processing uplift is limited to eligible endpoints and models released on or after 5 March 2026. Anthropic’s US-only option is documented at 1.1× input and output pricing. Other routes, including cloud marketplaces, may publish their own rates or add platform fees; verify the bill from the route you will actually deploy.
Tools and multimodal features can dominate a small-token workload
Search and grounding are not interchangeable labels. Anthropic lists web search as a platform charge; Google bills grounding by individual search query; code execution has its own feature terms. For an agent, log every tool invocation, query count and returned context tokens. Image, audio, storage and other modalities can introduce separate meters, so a text-only token estimate is incomplete for multimodal applications.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Practical ways to reduce spend without damaging quality
- Keep stable instructions and schemas at the beginning of prompts to improve cache reuse where supported.
- Measure cache hit rate instead of assuming every repeated prompt is cached.
- Trim irrelevant retrieval context and cap output length after testing quality.
- Route simple requests to a lower-cost model and reserve a stronger model for exceptions, then include routing retries in the estimate.
- Use batch mode for work that does not require interactive latency.
- Cache application-level retrieval results separately from provider prompt caches so an expired provider cache does not force a full recomputation.
- Set budgets and alerts by model, project, region and tool, not only by total account spend.
- Recalculate when a model, rate row, cache policy, residency choice or tool contract changes.
Common comparison mistakes
Using a single “per million tokens” number
One number hides input/output asymmetry, cache categories and model choice. Always show the exact model and each meter.
Ignoring output and reasoning tokens
Output can be the expensive side of a workload. Gemini’s standard output price includes thinking tokens; preserve that definition when comparing totals.
Assuming cache mechanics are universal
OpenAI’s automatic caching threshold and 128-token increments, Claude’s displayed five-minute cache TTL and Gemini’s context-cache implementation are different systems. Apply each provider’s own usage fields.
Leaving out failed attempts
Retries, tool loops and validation calls are real usage. Measure them in an evaluation rather than multiplying only successful request counts.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchCarrying an old promotion into a long forecast
GPT-5.6 Sol promotional pricing is stated to run at least through 21 November 2026. Mark the end date and model scope, then run a scenario at the non-promotional rate if your contract extends beyond it.
How to choose a provider from the result
Choose the lowest cost per accepted result that satisfies your latency, reliability, privacy, residency and feature requirements. If two options are close, prefer the one with simpler caching, predictable tool billing and a route that matches your compliance needs. Keep a second model qualified for outages or price changes, but price that fallback with its own token mix and geography rather than assuming it is interchangeable.
Finally, save the rate-card URLs, model IDs, assumptions, usage exports and verification date with the forecast. Re-run the worksheet whenever a provider changes a model row or billing rule; API pricing is a moving input to your operating model, not a one-time fact.
Frequently Asked Questions
How do I calculate the cost of one million tokens?
Separate input, cached-read, cache-write and output tokens, multiply each by the exact model rate shown by the provider, divide token counts by 1,000,000, and add applicable tool, request, storage, mode and geography charges.
Is Claude API always cheaper than OpenAI?
No. The answer changes with model capability, input/output mix, cache hits, US-only inference, retries and optional features. Test identical work and compare cost per accepted result.
Does Gemini charge separately for thinking tokens?
Google’s standard Gemini pricing states that the output price includes thinking tokens. Use the model’s reported billable output and verify the current table for the selected mode.
When should I include search or grounding fees?
Include them whenever the workload invokes web search or grounding. Count actual queries or executions because one model request can trigger multiple searches.
The Bottom Line
Compare OpenAI, Claude and Gemini with the same workload, provider-specific token definitions, cache assumptions, processing mode, geography, tools and retry rate. Refresh the official price rows before committing: the cheapest list price is not necessarily the lowest cost per accepted result.
Free tools Windows power users keep installed
One-click scans. No signup required.
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.
