CostPerPrompt publishes calculators for estimating monthly spending on chatbots, agents, retrieval systems and other AI workloads. Its homepage says it tracks pricing for more than 232 models, and its flagship pricing table is marked as updated on August 2, 2026.
The product addresses a budgeting problem that has grown alongside the model market. Providers usually quote separate input and output prices per million tokens. Actual bills depend on how an application handles repeated context, prompt caching, batch processing, tool definitions, retries and conversation history. A token-price table leaves those decisions to the developer.
CostPerPrompt's modules cover general API traffic, chatbots, agents, retrieval-augmented generation, voice applications, image pricing, GPU rentals and token counting. The site presents the results as planning estimates for developers evaluating workloads before deployment.
From token rates to a monthly bill
The API cost calculator is designed to compare a workload across tracked models while incorporating caching and batch discounts, according to CostPerPrompt's homepage.
Model prices vary widely even inside CostPerPrompt's own table. The site lists DeepSeek V4 Flash 0731 at $0.09 per million input tokens and $0.18 per million output tokens, OpenAI's GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens, and Anthropic's Claude Fable 5 at $10 per million input tokens and $50 per million output tokens. Teams still need to check calculator results against provider documentation before setting a budget because provider pages can separate cache hits, cache misses, output rates, batch jobs and long-context rules in ways that a summary table may flatten.
Whether a lower-priced model can complete a job reliably remains a separate evaluation that a pricing calculator cannot answer. Price comparisons establish the cost of a given token volume, while model quality, latency and reliability determine whether that volume produces a usable result.
The chatbot calculator addresses a less obvious source of spending. Chat applications can resend system instructions and conversation history as context, increasing input volume as an exchange grows. CostPerPrompt says its chatbot estimates account for growing history, resent context and cache hits.
CostPerPrompt says naive chatbot estimates can miss the final cost by 2-3x because they ignore caching and history growth. Its homepage describes prompt caching as cutting repeated input prices by up to 90%, but realized savings depend on provider rules and how consistently an application preserves reusable prompt prefixes.
The hard part is choosing assumptions
CostPerPrompt's broader modules follow the same approach. The agent cost calculator covers multi-step loops, tool schemas and retries. The RAG calculator separates indexing, retrieval and generation. CostPerPrompt says its voice calculator combines speech-to-text, model inference and text-to-speech costs.
This packaging reflects how AI applications are built. An agent can invoke a model repeatedly before producing a user-facing result. A support bot sends old messages back to the provider. A voice product can pay several vendors or model endpoints during the same minute of conversation. Monthly request counts alone do not capture those patterns.
The assumptions still require scrutiny. Cache-hit rates vary by workload, and some providers apply different rules to cache reads, cache writes and long-context requests. A 90% cache discount can describe a provider's listed rate without guaranteeing that 90% of an application's input will qualify. Batch discounts also depend on whether the product can accept delayed results.
CostPerPrompt says in its methodology that pricing is pulled primarily from OpenRouter's public model index and cross-checked against official provider pages. It describes calculator outputs as planning estimates rather than quotes and directs users to verify final numbers with providers.
Public price pages provide useful reference points. OpenAI lists GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens, while Anthropic lists Claude Fable 5 at $10 per million input tokens and $50 per million output tokens. Anthropic also lists a 90% input-token discount for prompt caching on Claude Fable 5.
A crowded market, with a packaging bet
AI pricing references are already plentiful. Price Per Token advertises coverage of more than 300 models and adds benchmarks, price history and a Model Context Protocol server that can supply pricing data to agents. Provider dashboards show actual spending after deployment, while observability products can attribute model usage to applications and customers.
CostPerPrompt is concentrating on pre-deployment planning. Its utility will depend on whether the workload templates remain easier to use than a spreadsheet while staying current as providers add tiered pricing, regional premiums, long-context surcharges and changing cache policies.
The underlying pricing data is becoming valuable infrastructure in its own right. OpenRouter said in May 2026 that it raised a $113 million Series B. CostPerPrompt uses that type of public routing and provider data to produce planning estimates.
The available CostPerPrompt pages do not identify a founder or operator, disclose funding, or name a legal entity. That leaves the organization behind the calculators opaque and limits scrutiny of its resources and data-maintenance process.
Accuracy is the product
CostPerPrompt says it is independent of model providers and funded through advertising and some affiliate links. The disclosure matters because pricing comparison pages can direct valuable traffic toward vendors.
Its terms state that CostPerPrompt does not guarantee every figure is current or error-free. That caveat is understandable for a service tracking model names, discounts and access terms that can change quickly. It also defines the central operating challenge: teams will use the calculators only if they trust the refresh process enough to begin their budgeting there.
CostPerPrompt has built a practical interface around workload-based cost estimation. Developers can use it for a first pass, then verify model prices, discount rules and workload assumptions against provider documentation before committing production traffic.