--- title: "What an LLM Request Actually Costs in Production: OpenAI, Claude and Gemini Compared per Task | Vindler Blog" description: "Price per million tokens tells you very little about your bill. We priced three real task shapes, a RAG support answer, an eight-call agent run and a document extraction, on ten current models from OpenAI, Anthropic and Google, with and without prompt caching and batch. The model tier you pick moves the cost a hundredfold. Caching and batch move it by a quarter to a half." url: "https://vindler.solutions/blog/llm-api-cost-per-task-2026" locale: "en" updatedAt: "2026-11-06" --- [Back to Blog](https://vindler.solutions/blog) What an LLM Request Actually Costs in Production: OpenAI, Claude and Gemini Compared per Task [LLM Cost](https://vindler.solutions/blog?tag=LLM%20Cost)[AI Cost Optimization](https://vindler.solutions/blog?tag=AI%20Cost%20Optimization)[OpenAI](https://vindler.solutions/blog?tag=OpenAI)[Claude](https://vindler.solutions/blog?tag=Claude)[Gemini](https://vindler.solutions/blog?tag=Gemini)[Prompt Caching](https://vindler.solutions/blog?tag=Prompt%20Caching)[Batch API](https://vindler.solutions/blog?tag=Batch%20API)[Model Selection](https://vindler.solutions/blog?tag=Model%20Selection) # What an LLM Request Actually Costs in Production: OpenAI, Claude and Gemini Compared per Task Price per million tokens tells you very little about your bill. We priced three real task shapes, a RAG support answer, an eight-call agent run and a document extraction, on ten current models from OpenAI, Anthropic and Google, with and without prompt caching and batch. The model tier you pick moves the cost a hundredfold. Caching and batch move it by a quarter to a half. November 6, 2026 1 minute Share: # What an LLM Request Actually Costs in Production _OpenAI, Claude and Gemini compared per task, with list prices as of November 6, 2026._ On current list prices, the same agent run costs anywhere from about half a cent to $1.00 depending on the model, a spread of more than a hundred times. Prompt caching takes 27% to 47% off that, and a batch API takes 50% off work that can wait. The three big providers price their flagship, middle and small tiers within a few cents of each other, so the decision that moves your bill is the tier you run each step on, and the second is whether you cache. Pricing pages quote dollars per million input and output tokens. Nobody buys tokens. You pay for answers, agent runs and processed documents, and each of those has its own mix of cached prefix, fresh context and output. So we priced tasks. ## How we priced it We took list prices from each provider's own pricing page on November 6, 2026: OpenAI's API pricing page, Anthropic's Claude pricing page, and Google's Gemini API pricing page. For each, we picked a flagship, a middle tier and a small model. Anthropic gets four rows because it lists both Fable 5.1 and Opus 5.5 at the top. Then we defined three task shapes that match what we see in production systems: 1. **A support answer with retrieval.** A 3,000-token system prompt and tool definitions that repeat on every request, 5,000 tokens of retrieved documents and conversation that change each time, and a 400-token answer. 2. **An agent run.** Eight model calls to finish one user request. Each call repeats a 6,000-token prefix (instructions and tool schemas), carries an average of 5,000 tokens of the growing transcript, and writes 300 tokens. Per run that is 48,000 tokens of repeated prefix, 40,000 fresh, 2,400 out. 3. **A document extraction.** A 12,000-token document in, 1,000 tokens of structured output, nothing repeated, nothing urgent. The optimized column uses cache-read prices for the repeated prefix in the first two tasks, and the batch price for the third. Every input is listed above, so you can rerun the numbers with your own token counts. ## Cost per 1,000 tasks | Model | Support answer | With caching | Agent run | With caching | Extraction | With batch | | --- | --- | --- | --- | --- | --- | --- | | OpenAI gpt-6-astra | $100.00 | $73.00 | $1,000.00 | $568.00 | $170.00 | $85.00 | | OpenAI gpt-6.1-terra | $10.00 | $7.15 | $100.00 | $54.40 | $17.00 | $8.50 | | OpenAI gpt-6-luna | $1.00 | $0.73 | $10.00 | $5.68 | $1.70 | $0.85 | | Claude Fable 5.1 | $100.00 | $70.75 | $1,000.00 | $532.00 | $170.00 | $85.00 | | Claude Opus 5.5 | $40.00 | $28.60 | $400.00 | $217.60 | $68.00 | $34.00 | | Claude Sonnet 5.5 | $20.00 | $14.30 | $200.00 | $108.80 | $34.00 | $17.00 | | Claude Haiku 5.5 | $1.00 | $0.73 | $10.00 | $5.68 | $1.70 | $0.85 | | Gemini 4 Argon | $20.00 | $14.30 | $200.00 | $108.80 | $34.00 | $17.00 | | Gemini 3.8 Flash | $7.50 | $5.47 | $75.00 | $42.60 | $12.75 | $6.38 | | Gemini 3.1 Flash-Lite | $2.60 | $1.93 | $25.60 | $14.80 | $4.50 | $2.25 | Divide by a thousand for the cost of one task. A single agent run on a flagship model costs a dollar before caching. The same run on a small model costs a cent. ## What the table says **The tier matters, the provider much less.** OpenAI's gpt-6-astra and Anthropic's Fable 5.1 have the same list price, $10 in and $50 out. Sonnet 5.5 and Gemini 4 Argon match each other at $2 and $10. Haiku 5.5 and gpt-6-luna both sit at $0.10 and $0.50. Switching provider at the same tier changes little on price. Moving a step from the top tier to the bottom changes the cost of that step by a factor of one hundred. **Agents are where caching pays.** A support answer repeats a small share of its input, so caching saves about 27%. An agent run sends its instructions and tool schemas on every call, so the repeated share is large and caching saves 43% to 47%. Anthropic's cache reads are the cheapest relative to its base price (Fable 5.1 reads cached input at 2.5% of the normal rate), which is why its flagship saves the most. **Batch is a flat half.** OpenAI, Anthropic and Google all price batch at 50% of the standard rate. If a job can wait for results, such as a nightly enrichment, a backfill or an evaluation run, sending it through the batch API is usually the easiest saving available. Azure's page and AWS Bedrock's page state the same 50% for batch. **Output costs five times input.** Every model here charges five times more for output than for input, and Gemini Flash-Lite six times. A prompt that asks for a long explanation when the code only reads a JSON field pays for every word. ## What the table leaves out We did not measure quality, and on many tasks the small model is not good enough, while on some it is as good as the flagship. The only way to know is to score candidates on a set of your own requests. We wrote up how we do that in [Agent Evaluation at Scale](https://vindler.solutions/blog/agent-evaluation-at-scale). Cache writes cost extra. Both OpenAI and Anthropic now bill writing a prefix to the cache at 1.25 times the input price for the short cache window, more for the long one. With steady traffic the writes are a rounding error. With a request every hour they can eat the saving. Google's cached rates on the Flash models also carry a storage charge per hour. Long prompts change the price. OpenAI's flagship doubles its input price above 272,000 tokens. Claude Haiku 5.5 moves to higher rates above 100,000 tokens, and Anthropic's page notes that cache reads count toward that threshold. Gemini 3.1 Pro Preview has a separate tier above 200,000. Regional endpoints cost more. OpenAI adds 10% for data residency endpoints on recent models, and Anthropic's page lists a 10% premium on regional and multi-region endpoints on the partner clouds. ## Prices that change in the next two months This table has a short shelf life. Google's page already lists the next steps. Gemini 4 Argon goes from $2 to $4 per million input tokens and from $10 to $20 output on December 1, 2026, which doubles every Argon number above. Gemini 3.8 Flash doubles on January 1, 2027. On December 1, Google also routes Gemini 2.5 Pro to 3.8 Flash and Gemini 2.5 Flash to 3.5 Flash-Lite, so anyone still pinned to the 2.5 models gets a model change and a price change on the same day. OpenAI's promotional price on GPT-5.6 Sol ends on November 30. On the Anthropic side, Claude Opus 4, Opus 4.1 and Sonnet 4 are retired on Anthropic's own API (they remain on Bedrock and Google Cloud), and Sonnet 4.5 is marked deprecated. We will update this post when the December prices take effect. ## What to do with it Measure cost per task before you change anything. Most teams know their monthly invoice and very little below it, and the expensive feature is rarely the one people guess. Once every model call is traced with its tokens and the feature that caused it, the order of work is usually the same: batch what can wait, cache the repeated prefixes, trim context the model does not use, route the simple steps to a small model, and only then consider switching the main model or self hosting. Test each change against your own requests before it ships. If you would like that done for your system, our [AI Cost and Model Review](https://vindler.solutions/ai-cost-review) is one week at a fixed $4,500, and it ends with the cost per task, a ranked list of savings with the quality risk of each, and the model retirement dates that apply to you. More on how we approach it is on the [AI cost optimization](https://vindler.solutions/hire/ai-cost-optimization) page. _Prices checked against the providers' pricing pages on November 6, 2026. List prices only; enterprise and committed-use discounts are not included._ Share: Carlos Dutra, founder of Vindler Solutions ### Carlos Dutra Founder of Vindler Solutions, where I help organizations put AI into production and design the operating model around it. I write about AI adoption, agent architectures, and what actually changes inside a company once these systems ship. Completed Leading the AI-Driven Organization at MIT Sloan Executive Education. [Get in Touch](https://vindler.solutions/contact) ## Related Posts [How to Build a Defense Factory: OpenAI's Playbook, Read by Engineers Who Build Agents Defense FactoryOpenAIAI AgentsApplication SecurityVulnerability ManagementAgentic SecurityCodexClaude CodeMCPSecurity ### How to Build a Defense Factory: OpenAI's Playbook, Read by Engineers Who Build Agents In September OpenAI published the Defense Factory playbook: an agent-first operation that connects your security tools, skills and reproducible environments to fix vulnerabilities continuously. It explains what to build and gives a six-week plan, and it leaves the engineering to you. We build agent systems for a living, so here is what each part of the playbook means in practice, where the work actually is, and how to keep the agents themselves from becoming the next incident. November 6, 2026 1 minute](https://vindler.solutions/blog/how-to-build-a-defense-factory)[OpenAI's Decisions API: A Fast Classifier for Agent Routing, and the Calibration Test You Have to Run Yourself AIOpenAIAgentsArchitectureClassificationEngineeringLLMProduction ### OpenAI's Decisions API: A Fast Classifier for Agent Routing, and the Calibration Test You Have to Run Yourself OpenAI's Decisions API went into public beta on October 6. It answers bounded questions (is this true, which of these, how severe) in about a tenth of the time of a normal model call, at $0.10 per million input tokens. It is the right shape for the routing and gating steps inside an agent. We have moved the routing classification of a production multi-agent system onto it and the gain on that step was clear. Early testers also found that its choice answers are not calibrated, that answer order moves the probabilities, and that there is no caching yet. Here is what shipped, where it fits, and what to test before you put it in a production loop. October 15, 2026 1 minute](https://vindler.solutions/blog/openai-decisions-api-agent-routing)[One Setting That Makes Claude Code Dramatically Faster at Navigating Code ClaudeClaude CodeAIDeveloper ToolsProductivity ### One Setting That Makes Claude Code Dramatically Faster at Navigating Code A single environment flag unlocks LSP support in Claude Code, replacing slow text searches with the same semantic code intelligence your IDE uses. Two minutes of setup, permanent improvement on any non-trivial codebase. March 12, 2026 1 minute](https://vindler.solutions/blog/claude-code-lsp) ### Subscribe to our newsletter Get notified when we publish new posts on AI development, AWS, and software engineering. ## Structured data ```json {"@context":"https://schema.org","@type":"Organization","@id":"https://vindler.solutions/#organization","name":"Vindler","alternateName":"Vindler Solutions","legalName":"Vindler Desenvolvimento de Sistemas de IA Ltda - EPP","url":"https://vindler.solutions","logo":"https://vindler.solutions/logo.png","description":"Vindler is a global AI engineering consultancy specializing in custom AI solutions, LLM integration, and enterprise-grade AI systems.","foundingDate":"2025-06-02","foundingLocation":{"@type":"Place","name":"São Paulo, Brazil"},"taxID":"61.102.920/0001-05","email":"contact@vindler.solutions","telephone":"+1-646-846-5339","numberOfEmployees":{"@type":"QuantitativeValue","minValue":1,"maxValue":10},"areaServed":[{"@type":"Country","name":"United States"},{"@type":"Place","name":"Europe"},{"@type":"Country","name":"Brazil"}],"knowsAbout":["Multi-agent systems","Retrieval-augmented generation","Voice AI agents","LangChain","LangGraph","LiveKit","AWS Bedrock","Model Context Protocol","LLM observability and evaluation"],"founder":{"@type":"Person","@id":"https://vindler.solutions/about#carlos-dutra","name":"Carlos Dutra","jobTitle":"Founder and CEO","url":"https://vindler.solutions/about","sameAs":["https://www.linkedin.com/in/carlosdutra","https://github.com/cdutr"]},"contactPoint":{"@type":"ContactPoint","contactType":"Customer Support","email":"contact@vindler.solutions","telephone":"+1-646-846-5339","url":"https://vindler.solutions/contact","availableLanguage":["English","Portuguese","Spanish"]},"address":[{"@type":"PostalAddress","addressLocality":"São Paulo","addressRegion":"SP","addressCountry":"BR","streetAddress":"Av. Paulista, 1636"},{"@type":"PostalAddress","addressLocality":"New York","addressRegion":"NY","addressCountry":"US","streetAddress":"524 Broadway","postalCode":"10012"}],"sameAs":["https://www.linkedin.com/company/vindler-tech/","https://github.com/vindler-solutions","https://clutch.co/profile/vindler-solutions","https://www.wikidata.org/wiki/Q141018871","https://www.crunchbase.com/organization/vindler-solutions"],"memberOf":{"@type":"Organization","name":"Brazilian-American Chamber of Commerce, Inc.","url":"https://brazilcham.com/","sameAs":"https://www.linkedin.com/company/brazilian-american-chamber-of-commerce-inc./"},"hasOfferCatalog":{"@type":"OfferCatalog","name":"AI Engineering Services","itemListElement":[{"@type":"Offer","itemOffered":{"@type":"Service","name":"AI Engineering","description":"Custom AI systems development from prototyping to deployment."}},{"@type":"Offer","itemOffered":{"@type":"Service","name":"Custom ChatBots & Agents","description":"Intelligent, multilingual, and domain-specific conversational AI."}},{"@type":"Offer","itemOffered":{"@type":"Service","name":"AI Architecture Consultancy","description":"Designing scalable, secure, and ethical AI infrastructure for enterprises."}},{"@type":"Offer","itemOffered":{"@type":"Service","name":"Training & Team Building","description":"Upskill your team or recruit top-tier AI talent."}}]}} ``` ```json {"@context":"https://schema.org","@type":"BlogPosting","headline":"What an LLM Request Actually Costs in Production: OpenAI, Claude and Gemini Compared per Task","description":"Price per million tokens tells you very little about your bill. We priced three real task shapes, a RAG support answer, an eight-call agent run and a document extraction, on ten current models from OpenAI, Anthropic and Google, with and without prompt caching and batch. The model tier you pick moves the cost a hundredfold. Caching and batch move it by a quarter to a half.","image":"https://vindler.solutions/images/blog/llm-cost-per-task-nano-v3.jpg","datePublished":"2026-11-06","dateModified":"2026-11-06","author":{"@type":"Person","@id":"https://vindler.solutions/about#carlos-dutra","name":"Carlos Dutra","jobTitle":"Founder and CEO, Vindler","url":"https://vindler.solutions/about"},"publisher":{"@type":"Organization","@id":"https://vindler.solutions/#organization","name":"Vindler","logo":{"@type":"ImageObject","url":"https://vindler.solutions/logo.png"}},"mainEntityOfPage":{"@type":"WebPage","@id":"https://vindler.solutions/blog/llm-api-cost-per-task-2026"}} ```