Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
LLM API prices are not comparable until you specify the exact model, input and output mix, context length, cache behavior, service tier, deployment geography, and date. Providers charge separately for generated output, cache writes or storage, and sometimes server-side tools, so a headline “price per million tokens” can be misleading.
The examples below are official USD rates checked on October 13, 2026. They are illustrations rather than a complete market survey; verify the linked pricing page immediately before procurement because models, prices, and scheduled changes can move.
What an LLM API bill actually contains
Input tokens
Input tokens include the user prompt and the context sent with it. Long system instructions, conversation history, retrieved documents, tool definitions, and tool results can all increase input usage. Some providers count reasoning or “thinking” tokens in a separate output category; use the model’s billing definition rather than assuming hidden reasoning is free.
Output tokens
Output tokens are generated text and, for some models, billed thinking tokens. Output rates are commonly several times higher than input rates, so limiting unnecessary verbosity can reduce cost more than shaving a small amount from the prompt.
Cache reads, writes, refreshes, and storage
“Cached tokens” is not one universal category. A provider may distinguish a discounted cache hit, the initial write, a refresh after expiry, and storage charged by time. A fair estimate must state how many tokens are written, how often they are reused, how long they remain valid, and whether storage is metered separately.
Tool work
A tool call can add tokens for the tool name, description, JSON schema, an automatically inserted system instruction, the model’s tool-use block, the tool result, and any follow-up reasoning. Server-side tools such as web search may add a request fee on top of those model tokens. There is no reliable universal “cost per tool call.”
Official pricing examples
All token figures in the tables are USD per 1 million tokens unless a different unit is shown. They are provider-published examples, not independent market statistics.
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1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsOpenAI API
The OpenAI API pricing page separates standard input, cached input, cache writes, and output, and it has different columns for short and long context. The following short-context standard rows were listed in the pricing material checked on October 13, 2026:
| Model | Input | Cached input | Cache write | Output | Long-context column |
|---|---|---|---|---|---|
| gpt-6-astra | $10 | $1 | $12.50 | $50 | Higher; exact amount depends on the model row |
| gpt-6.1-sol | $2 | $0.10 | $2.50 | $10 | Higher; exact amount depends on the model row |
| gpt-6-luna | $0.10 | $0.01 | $0.125 | $0.50 | Higher; exact amount depends on the model row |
The same pricing material also records a gpt-6.1-sol short-context example at $1 input, $0.05 cached input, $1.25 cache writes, and $5 output. Because multiple rows or revisions can appear under a model family, treat that set as a separate listed example, not as a reason to average the numbers. Confirm the exact model identifier and row you will purchase.
Eligible regional-processing endpoints for models released on or after March 5, 2026 have a 10% uplift. FedRAMP endpoints also carry a 10% uplift. Built-in tools use the selected model’s token rates; add a separate tool fee only when the official pricing table lists one.
Anthropic Claude API
Anthropic’s Claude pricing documentation lists input, output, 5-minute cache writes, 1-hour cache writes, and cache hits or refreshes as separate categories:
Free tools Windows power users keep installed
One-click scans. No signup required.
| Model | Input | Output | 5-minute cache write | 1-hour cache write | Cache hit or refresh |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 | $2.50 | $4 | $0.10 |
| Claude Opus 5.5 | $4 | $20 | $5 | $8 | $0.20 |
Anthropic states, “Tool use requests are priced based on:” the total input tokens sent to the model (including the tools parameter), generated output tokens, and additional usage-based pricing for server-side tools such as a charge per web search. Tool names, descriptions, schemas, tool-use blocks, and tool-result blocks all consume tokens. An automatically included system prompt also adds usage, and its size varies by model and tool-choice mode.
For Claude 4.6 and later models using US-only inference, Anthropic applies a 1.1× multiplier to token-pricing categories. Partner-operated cloud platforms have their own regional pricing, so do not copy first-party API rates to a partner deployment.
Google Gemini API
Google’s Gemini API pricing page has model-specific rows for text input, output (including thinking tokens where applicable), context caching, and cache-storage time. One visible example, Gemini 3 Flash Preview, is listed through December 31, 2026 at $0.50 input, $3 output, and $0.05 for context-cached tokens, plus $1 per million cached tokens per hour for storage.
Google also lists scheduled changes. For one model row, the rate through December 31, 2026 is $0.75 input, $3.75 output, and $0.075 context caching; from January 1, 2027 those become $1.50, $7.50, and $0.15. The exact model row matters, so do not apply those figures to every Gemini model.
Search grounding includes 5,000 free requests per month shared across Gemini 3 and newer models. After that allowance, supported listed models are charged $14 per 1,000 requests. Check the row for tool support and the current allowance before forecasting a production bill.
A fair way to compare providers
Use the same workload definition for every provider. A useful estimate is:
Total cost = (uncached input tokens × input rate) + (cache-hit tokens × cache-hit rate) + (cache-write tokens × write rate) + (output tokens × output rate) + (tool requests × tool-request fee) + cache-storage charges.
Keep each category separate; do not subtract cache hits from input twice or treat a write as a hit.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Worked example: Claude Sonnet 5.5
Suppose one month contains 4 million uncached input tokens, 6 million cache-hit tokens, 1 million tokens written to a five-minute cache, and 2 million output tokens. Using the published Sonnet 5.5 rates above:
- Uncached input: 4 × $2 = $8.00
- Cache hits: 6 × $0.10 = $0.60
- Five-minute writes: 1 × $2.50 = $2.50
- Output: 2 × $10 = $20.00
- Token subtotal: $31.10
Any server-side search or other metered tool requests would be added separately, as would storage if the selected service charges for it. This example is a calculation from the published rates, not a provider benchmark.
Seven steps for a procurement-grade estimate
- Freeze the model and date. Record the exact model ID, API or cloud platform, pricing page URL, and date checked. A family name alone is not enough.
- Measure input and output separately. Use production logs or a representative sample. Include conversation history, retrieval context, tool schemas, tool results, and billed thinking tokens where the provider counts them.
- Define the context tier. State the maximum context and whether requests cross a long-context threshold. OpenAI’s long-context prices are higher than its short-context column.
- Describe cache behavior. Estimate the percentage of requests that hit, the token volume written, the write duration, refresh frequency, and storage hours. A five-minute cache and a one-hour cache are different products.
- Count tool activity. Track calls by tool type, schema size, result size, retries, and intermediate model turns. Add web-search, grounding, code-execution, or other server-tool fees where the provider specifies them.
- Apply tier and geography modifiers. Check batch, flex, priority, regional-processing, FedRAMP, US-only inference, and partner-cloud terms. A regional or compliance endpoint can cost more than the base row.
- Run sensitivity cases. Price low, expected, and high output lengths; cache-hit rates; tool-call rates; and traffic growth. Recheck before signing because these rates are volatile.
How caching changes the economics
Cache writes are an upfront cost
A write can cost more than an ordinary input read. In the Anthropic example, a five-minute write is $2.50 versus $2 input, while a one-hour write is $4. The longer duration costs more but can avoid repeated writes when the same large prefix is reused.
Hits can be dramatically cheaper
Anthropic’s Sonnet 5.5 cache hit or refresh is $0.10, 95% below its $2 input rate. That discount only applies to eligible tokens actually served from cache; changing the prefix, allowing the cache to expire, or exceeding retention assumptions can turn the request back into ordinary input or a new write.
Storage may be metered independently
Gemini 3 Flash Preview lists $1 per million cached tokens per hour for storage in addition to its context-caching token charge. A workload with infrequent reads can therefore pay meaningful storage cost even when inference volume is low. Include retention hours in the estimate.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Tool calls: what to count and where estimates fail
- Definitions: long descriptions and JSON schemas are sent with requests and consume input tokens.
- Control instructions: providers may insert system text for tool choice; Anthropic says the amount varies by model and mode.
- Arguments and results: both the model’s tool-use block and the returned payload are billable context.
- Intermediate turns: an agent may call a tool, inspect the result, revise its plan, and call again; each turn can generate input and output tokens.
- Separate request fees: server-side web search or grounding can add a per-request charge even when model tokens are also billed.
For OpenAI built-in tools, start with the chosen model’s token rates and then add any separately listed tool fee. For Anthropic, use the documented total-input, output, and server-tool components. For Gemini Search grounding, apply the shared 5,000-request monthly allowance and then the listed $14-per-1,000 rate only to supported models.
Context, service tier, and geography can outweigh the headline rate
Two teams can select the same model and receive different effective prices. A request that crosses a long-context threshold may use a higher column. Priority or low-latency tiers can have different rates from standard service. Regional-processing and FedRAMP endpoint uplifts change OpenAI totals, while US-only inference changes Claude 4.6-and-later token categories by 1.1×. Partner cloud deployments may publish independent prices.
Always put these modifiers beside the number in an internal cost sheet. “$2 per million input tokens” without model, context, endpoint, and date is not a reproducible price.
Pricing changes to watch before launch
Google has explicitly scheduled January 1, 2027 changes for some Gemini rows, including the example that doubles input from $0.75 to $1.50 and output from $3.75 to $7.50. Other providers can also change rates, model names, cache rules, or tool fees without matching your original estimate. Re-open the official pricing pages immediately before publication, procurement, and any major traffic migration.
Quick Recap
Practical checklist
- Exact model/version and provider platform recorded
- Pricing-page date and URL recorded
- Input, output, and reasoning-token treatment documented
- Short- versus long-context tier identified
- Cache hits, writes, refreshes, retention, and storage hours estimated
- Tool schemas, intermediate turns, results, retries, and server-tool fees counted
- Regional, compliance, partner-cloud, and latency-tier modifiers applied
- Scheduled price changes and free allowances excluded or dated explicitly
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.
