Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors?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 reinstallLLM API prices are not directly comparable from a single input-token number. Your real bill depends on the exact model and date, input and output volume, cache reads and writes, context length, service tier, deployment geography, and any separately metered tools. The examples below use provider-published USD rates checked on October 13, 2026. Treat them as a framework for building a workload estimate, then recheck the provider page before procurement because prices and model names change.
What an LLM API bill actually measures
Most text APIs meter at least two token categories: tokens sent to the model (input) and tokens generated by it (output). A prompt containing a system message, conversation history, retrieved documents, tool definitions, and a new user request all contributes input tokens. The answer contributes output tokens. Models that expose reasoning or thinking tokens may count those tokens in the provider’s output or usage total, even when they are not shown as visible prose.
As an Amazon Associate I earn from qualifying purchases.
Caching adds more categories rather than replacing the normal input charge. A provider can charge for creating a cache entry, reading a cached prefix, refreshing an expired entry, and storing cached tokens over time. Tool use can add model tokens for schemas, tool calls, intermediate content and results, plus a separate fee for a hosted search or other server-side operation. A useful estimate therefore looks like this:
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →total = (input tokens × input rate) + (output tokens × output rate) + cache writes + cache reads/refreshes + cache storage + separately metered tool requests
#1 Best Overall
Every rate in that equation must carry its conditions: model version, context band, service tier, endpoint geography, cloud platform, currency and effective date. A “price per million tokens” without those assumptions is not a meaningful comparison.
Official example rates
The following figures are USD per 1 million tokens unless a separate unit is shown. They are illustrative rows from official pricing pages, not a complete catalogue of every model, modality or discount program.
| Provider and model | Input | Cached input or hit | Cache write | Output | Other published conditions |
|---|---|---|---|---|---|
| OpenAI gpt-6-astra, short-context standard | $10 | $1 | $12.50 | $50 | Long-context column is higher; exact amount depends on the model row. |
| OpenAI gpt-6.1-sol, short-context standard | $2 | $0.10 | $2.50 | $10 | Long-context pricing is listed separately. |
| OpenAI gpt-6-luna, short-context standard | $0.10 | $0.01 | $0.125 | $0.50 | Long-context pricing is listed separately. |
| Anthropic Claude Sonnet 5.5 | $2 | $0.10 cache hit or refresh | $2.50 for 5-minute; $4 for 1-hour | $10 | Cache duration changes the write price. |
| Anthropic Claude Opus 5.5 | $4 | $0.20 cache hit or refresh | $5 for 5-minute; $8 for 1-hour | $20 | Cache duration changes the write price. |
| Google Gemini 3 Flash Preview (through Dec. 31, 2026) | $0.50 | $0.05 context-caching charge | Not stated as a separate write category | $3 | Cache storage is $1 per million cached tokens per hour. |
Check the OpenAI API pricing, Claude API pricing and Gemini API pricing pages for the exact row you will deploy. OpenAI’s page also records a 10% uplift for eligible regional-processing endpoints on models released on or after March 5, 2026, and a 10% FedRAMP endpoint uplift. Those surcharges apply on top of the model’s listed categories where the endpoint qualifies.
How input, output and cached tokens differ
Input tokens
Input is everything transmitted for inference, not only the user’s latest sentence. Long chat history, retrieved text, images represented as tokens, JSON schemas and tool results can dominate a request. Separate input and output estimates because output rates are often several times higher than input rates.
Output and reasoning tokens
Output includes generated text and, for models whose pricing table says so, thinking or reasoning tokens. A short visible answer can still consume substantial billed output if the model performs extended internal reasoning. Use the provider’s usage fields to measure this rather than estimating from character count.
Cache hits, writes and storage
A cache hit means a previously stored prompt prefix was reused. It is usually cheaper than ordinary input but only applies to eligible tokens and an actual hit. A write creates or extends the cached prefix and can cost more than a hit. Some services separately charge for retaining cached tokens by the minute or hour. Expiration can trigger a refresh charge or a new write.
Rank #3
| Cache event | What to count | Question to answer in your estimate |
|---|---|---|
| Write | Tokens inserted into a cache and the selected duration | How many tokens are written, and is the lifetime 5 minutes, 1 hour or another period? |
| Hit/read | Eligible cached tokens read on each request | What percentage of requests reuse the prefix? |
| Refresh | Tokens revalidated after expiry or a miss | How often does the application exceed the cache lifetime? |
| Storage | Cached token volume multiplied by retention time | Does the provider bill token-hours separately from reads and writes? |
Anthropic explicitly separates 5-minute writes, 1-hour writes, and cache hits or refreshes. Google’s Gemini example separates a context-caching token charge from $1 per million cached tokens per hour of storage. OpenAI presents input, cached-input and cache-write rates. Do not map one provider’s “cached input” label to another provider’s storage or write category.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsTool calls are a bundle of costs, not one flat fee
Adding a tool to a request changes the token payload. The model receives each tool’s name, description and JSON schema; it may emit a tool-use block, receive a tool-result block, and then generate a final response. All of those blocks can count as input or output tokens. Large schemas and verbose results can cost more than the user prompt itself.
Built-in and client-executed tools
For OpenAI built-in tools, model tokens are billed at the chosen model’s rates. A tool that your application executes locally still creates model-token usage for the definition, call and result, even if there is no vendor tool fee. Count each round trip when a model calls several tools before answering.
Anthropic tool use
Anthropic states that tool-use pricing is based on total input tokens sent to the model, including the tools parameter, output tokens generated, and additional usage-based pricing for server-side tools. Its documentation notes that tool names, descriptions, schemas, tool-use blocks, tool-result blocks and an automatically included system prompt add tokens. The number of added system-prompt tokens varies by model and tool-choice mode.
Search and grounding fees
Google’s pricing page lists 5,000 free Search grounding requests per month shared across Gemini 3 and newer models. After that allowance, supported listed models are charged $14 per 1,000 requests. The search request fee is in addition to the model’s token charges, and support and price depend on the exact model row. A search-heavy workflow should therefore forecast both requests and the tokens used to formulate, ground and summarize each result.
A fair way to compare providers
- Freeze the model and date. Record the exact model/version, pricing page URL and date checked. “Claude,” “Gemini” or “GPT” alone is too broad for a procurement comparison.
- Define one workload. Specify requests per day or month, average input tokens, average generated tokens, maximum context, and the percentage of requests that include a tool.
- Separate token categories. Put ordinary input, cached reads, cache writes, output and reasoning tokens in different columns. Do not hide them in one blended rate.
- Model cache behavior. Estimate cacheable-prefix size, hit rate, write frequency, lifetime and refreshes. Include storage duration where the provider bills it.
- Inventory tool payloads. Count schema tokens, intermediate turns, result tokens and final output. Add web-search, grounding, code-execution or other server-tool fees using the provider’s stated unit.
- Apply context and service-tier rules. Long-context bands, batch or flex modes, priority service, promotional windows and regional processing can change the rate or latency. Use the row that matches production, not a cheaper default row.
- Match deployment geography. First-party APIs and partner-operated cloud platforms may have different regional prices. Anthropic says Claude 4.6 and later models using US-only inference apply a 1.1× multiplier to token categories; partner cloud pricing is independent.
- Run a sensitivity range. Calculate a low, expected and high case for output length, cache-hit percentage and tool frequency. A single average conceals the variables most likely to move the bill.
Worked estimate for a defined workload
Suppose one month contains 1 million input tokens, 200,000 output tokens, a 20% cached prefix (200,000 tokens), and a 1-million-token cache write. Using OpenAI gpt-6.1-sol short-context standard rates, the arithmetic is:
- 800,000 ordinary input tokens × $2 per million = $1.60.
- 200,000 cached-input tokens × $0.10 per million = $0.02.
- 200,000 output tokens × $10 per million = $2.00.
- 1 million cache-write tokens × $2.50 per million = $2.50.
The token subtotal is $6.12 before any tool fee, regional-processing uplift, long-context rate or other service-tier modifier. This is an illustration of the method, not a prediction of your bill; changing the hit rate or output length changes the result.
For the Gemini 3 Flash Preview example through December 31, 2026, 1 million input tokens cost $0.50 and 200,000 output tokens cost $0.60. If 500,000 tokens are retained in context cache for two hours, the stated storage rate contributes 0.5 × 2 × $1 = $1.00, while the context-caching token charge contributes 0.5 × $0.05 = $0.025 if those tokens are in the billed caching category. That illustrative subtotal is $2.125 before Search grounding or other usage. Confirm how your SDK reports eligible cached tokens before using this arithmetic in a budget.
Dates, scheduled changes and geography that can invalidate a quote
Google’s pricing page includes explicit rows with current rates through December 31, 2026 and new rates from January 1, 2027. One listed model changes from $0.75 input and $3.75 output per million tokens to $1.50 input and $7.50 output; its context-caching rate changes from $0.075 to $0.15. A forecast that spans the change must split usage by date instead of applying one annual average.
OpenAI’s short- and long-context columns make context length a pricing variable, while regional-processing and FedRAMP endpoints can add 10% for eligible models and endpoints. Anthropic’s US-only inference multiplier and independent partner-cloud rates create a similar geography issue. Record the endpoint and region alongside the model in every quote.
Common budgeting mistakes and fixes
- Comparing only input prices: Fix it by multiplying separate input and output volumes; output can dominate interactive workloads.
- Treating cache as free: Include writes, hits, refreshes and storage, and verify the cache lifetime.
- Counting a tool as one request: Add schema, call, result and intermediate model tokens, then add any server-tool unit fee.
- Using a headline model name: Pin the version, context band and endpoint; model families contain materially different rates.
- Ignoring thinking tokens: Inspect provider usage fields and include billed reasoning tokens where the pricing table requires it.
- Assuming partner clouds match first-party APIs: Check the platform’s regional price sheet.
- Using an old spreadsheet: Put an effective date in every rate cell and recheck official pages immediately before launch or renewal.
What to put in a procurement or architecture worksheet
Capture one row per model and deployment with these fields: model/version; provider and endpoint; region; pricing date; context band; service tier; monthly ordinary input tokens; cached-read tokens and hit rate; cache-write tokens and lifetime; cache-storage token-hours; output and reasoning tokens; tool calls by type; tool-result token volume; separate search or grounding requests; and any scheduled rate change. Keep the source URL beside each row so a reviewer can verify the value.
Then calculate monthly cost by category, not just a blended average. Report the expected total and the low/high sensitivity cases, explain which assumptions create the range, and set an alert for model, region or pricing-page changes. This makes a comparison reproducible when traffic, prompts or provider rates change.
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.
