AI Model API Pricing: Real Costs per 1,000 Prompts

👁 113 reads

How much does an AI API actually cost? Running 1,000 typical requests currently ranges from [pw_rc_fakt k=”najtanszy_koszt”] with [pw_rc_fakt k=”najtanszy”] up to [pw_rc_fakt k=”najdrozszy_koszt”] with [pw_rc_fakt k=”najdrozszy”] - a [pw_rc_fakt k=”roznica”] price gap for the exact same workload. I update this ranking weekly using public pricing endpoints, incorporating direct token overhead benchmarks across models. Since October 8, the comparison also includes Claude Haiku 5.5.

[pw_ranking_cen]

How the Cost per 1,000 Requests Is Calculated

List prices per million tokens mean very little until you know how many tokens your actual prompts consume. That is why I standardize every model against a single, uniform unit of work:

• A standard request assumes a baseline English prompt of 400 input tokens (an instruction paired with context, such as a customer email) and 600 output tokens (a multi-paragraph response).
• When working with non-English inputs, I adjust this baseline using a multiplier derived from my empirical tokenization tests - for instance, 1.46 for GPT-6 and 2.37 for Claude 5.5 when processing Polish text.
• Total cost = 1,000 × multiplier × (400 × input price + 600 × output price) / 1,000,000.
• Prices are fetched directly from the public OpenRouter /api/v1/models endpoint, reflecting base provider rates without markup, tracked in USD.
• Week-over-week changes compare current costs against snapshots taken at least 6 days prior.

Factors omitted from this baseline that may increase your actual bill:

• Hidden reasoning tokens: models that “think” before responding bill internal reasoning steps as output tokens, which can multiply costs significantly when reasoning mode is active.
• Long context overhead: for documents running tens of thousands of tokens, input rates dominate, and several providers apply higher tiers above specific thresholds (e.g., 200,000 tokens).
• Discounts: prompt caching and batch API modes can cut costs by 50% or more, as detailed in my guide to tokens and caching.

Token Efficiency Benchmark: How Tokenizers Affect Real Costs

I tested 35 models via OpenRouter using identical test text - a 120-word customer support inquiry regarding an espresso machine warranty claim - sent in both Polish and English. I extracted the exact input token count reported by each provider’s API. Chat template framing overhead was isolated and subtracted by sending a single-character baseline message. The entire measurement run cost roughly $0.15.

Key takeaways from these measurements:

• Claude 5.5 and GPT-6.1 Sol advertise identical rates ($2.00 per million input tokens, $10.00 output). However, the tested non-English passage consumed 341 tokens on Claude versus 210 on GPT. Consequently, 1,000 requests cost [pw_rc_fakt id=”anthropic/claude-sonnet-5.5″] on Sonnet 5.5 versus [pw_rc_fakt id=”openai/gpt-6.1-sol”] on GPT-6.1 Sol - roughly 62% more for the exact same nominal pricing.
• The Claude 5 tokenizer is noticeably dense even in English: 207 tokens where GPT used 144. Inefficient tokenization is not limited to non-Latin scripts or accented text.
• Gemini and Grok encode multilingual text most efficiently: a 1.16-1.37 multiplier means their effective cost stays closest to baseline list prices.
• Kimi K3 offers very low input pricing ($0.50), but higher output pricing ($15.00) combined with nearly double the token volume places it alongside top-tier flagships in overall expense.

An earlier test conducted with a shorter passage on the AI Tokens page yielded between 1.38x and 2.11x token inflation relative to English. Longer texts compress the variance slightly, though relative model rankings remain consistent. I rerun this benchmark script whenever a new tokenizer enters the index.

Choosing the Right Model Tier

The cheapest model is not necessarily the right fit for your pipeline. Here is how the ranking segments into practical tiers:

• High-volume automated tasks (email classification, summaries, data extraction): GPT-6 Luna ([pw_rc_fakt id=”openai/gpt-6-luna”]), Gemini 2.5 Flash-Lite ([pw_rc_fakt id=”google/gemini-2.5-flash-lite”]), Mistral Small 4 ([pw_rc_fakt id=”mistralai/mistral-small-2603″]). Luna passed my non-English benchmarks - see the breakdown in the review of GPT-6 Luna.
• Production chatbots and daily generation: Gemini 3.8 Flash ([pw_rc_fakt id=”google/gemini-3.8-flash”], introductory rate through late 2026), DeepSeek V4.1 Flash ([pw_rc_fakt id=”deepseek/deepseek-v4.1-flash”]), Mistral Large 4 ([pw_rc_fakt id=”mistralai/mistral-large-4-0″]).
• Complex agent workflows and coding: GPT-6.1 Sol ([pw_rc_fakt id=”openai/gpt-6.1-sol”]), Claude Sonnet 5.5 ([pw_rc_fakt id=”anthropic/claude-sonnet-5.5″]), Grok 4.7 ([pw_rc_fakt id=”x-ai/grok-4.7″]).
• High-end reasoning when required: Claude Opus 5.5 ([pw_rc_fakt id=”anthropic/claude-opus-5.5″]), GPT-6 Astra ([pw_rc_fakt id=”openai/gpt-6-astra”]), Claude Fable 5.1 ([pw_rc_fakt id=”anthropic/claude-fable-5.1″]).
• Open-weights models suitable for self-hosting: Gemma 4 31B, Qwen 3.8, GLM-5.3, DeepSeek V4 - explored further in the guide to running local AI models.

To evaluate relative quality independently, explore the multilingual model arena, and review the cost per benchmark intelligence point on the price of intelligence tracker.

AI Model Pricing by Provider

Estimating Monthly API Costs: Three Real Scenarios

Since the index displays cost per 1,000 calls, monthly bills scale linearly. Consider these three typical implementation profiles:

• Customer service bot, 10,000 interactions monthly (single-turn average): GPT-6 Luna [pw_rc_fakt id=”openai/gpt-6-luna” x=”10″], Gemini 3.8 Flash [pw_rc_fakt id=”google/gemini-3.8-flash” x=”10″], Claude Sonnet 5.5 [pw_rc_fakt id=”anthropic/claude-sonnet-5.5″ x=”10″].
• Automated inbound email triage, 3,000 messages monthly: DeepSeek V4.1 Flash [pw_rc_fakt id=”deepseek/deepseek-v4.1-flash” x=”3″], GPT-6.1 Sol [pw_rc_fakt id=”openai/gpt-6.1-sol” x=”3″].
• Dedicated power-user assistant, 50 queries daily across 22 workdays (1,100 calls): Claude Opus 5.5 [pw_rc_fakt id=”anthropic/claude-opus-5.5″ x=”1.1″], GPT-6.1 Sol [pw_rc_fakt id=”openai/gpt-6.1-sol” x=”1.1″].

If you primarily interact via conversational interfaces rather than direct API calls, compare these figures against standard subscriptions. For guidance on setting up credentials, review the tutorial on getting started with AI APIs, or access these models through a single key via OpenRouter.

Image Generation: API Costs per 1,000 Images

Vision generation models bill image generation output as tokens. Here, costs are normalized to 1,000 images at approximately 1024×1024 resolution (prompt baseline: 100 English text tokens). Image token counts follow official documentation: 1,120 tokens for Gemini 3 image endpoints and 1,056 for GPT Image at medium quality. Generating at 2K, 4K, or higher quality presets scales token usage accordingly.

• GPT Image mini: [pw_rc_fakt id=”openai/gpt-5-image-mini”] per 1,000 images,
• GPT Image 2: [pw_rc_fakt id=”openai/gpt-5.4-image-2″],
• Nano Banana 2.1: [pw_rc_fakt id=”google/gemini-nano-banana-2.1″] - performance tested in the Nano Banana review,
• Nano Banana Pro: [pw_rc_fakt id=”google/gemini-3-pro-image”].

Switch the ranking view to the “Images” filter to inspect all six models visually. Quality differences are analyzed in the GPT Image comparison.

Frequently Asked Questions

How much does an AI API cost?

Pricing ranges from fractions of a cent to several cents per call. A bundle of 1,000 standard requests spans from [pw_rc_fakt k=”najtanszy_koszt”] on the most economical model to [pw_rc_fakt k=”najdrozszy_koszt”] on top-end frontier models. You pay strictly for consumed tokens, with no recurring subscription fees.

What is the cheapest AI model right now?

This week, the lowest-cost model indexed is [pw_rc_fakt k=”najtanszy”]. For robust multilingual support backed by major infrastructure, leading low-cost contenders include GPT-6 Luna and Gemini 2.5 Flash-Lite.

Why does non-English text cost more in AI APIs?

Most tokenizers are trained predominantly on English corpora. Non-English vocabulary, inflected grammar, and diacritics split into smaller subword fragments, yielding higher token counts for identical semantic content. In my benchmark, non-English text required 17% to 98% more tokens than the equivalent English phrasing on the same model.

Is Claude more expensive than ChatGPT via API?

At identical listed token rates, yes, whenever non-English text or token-dense content is involved. Claude Sonnet 5.5 and GPT-6.1 Sol both list at $2.00 input and $10.00 output per million tokens, but our benchmark text expanded to 341 tokens in Claude versus 210 in GPT, making Sonnet approximately 62% more expensive in practice.

How often are AI model prices updated?

The entire pricing directory is refreshed weekly, calculating net price deltas against previous snapshots. Price adjustments are covered as they break, and the AI cost calculator recalculates everyday workflow projections continuously.

Are OpenRouter prices identical to direct provider rates?

OpenRouter lists upstream provider pricing directly without per-token markups; its margin is handled via account top-up fees. Open-weights models are hosted by multiple third-party infrastructure providers with varying price floors, so rates at specific providers may deviate from directory averages.

How much do 1,000 requests cost in GPT models?

Baseline benchmarks run at [pw_rc_fakt id=”openai/gpt-6-luna”] for GPT-6 Luna, [pw_rc_fakt id=”openai/gpt-6.1-sol”] for GPT-6.1 Sol, and [pw_rc_fakt id=”openai/gpt-6-astra”] for GPT-6 Astra, assuming standard request length without reasoning tokens enabled.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources and Methodology

• Pricing data: public API endpoint OpenRouter /api/v1/models, fetched [pw_rc_fakt k=”data”].
• Token benchmark: empirical tests conducted October 7, 2026 (measuring usage.prompt_tokens across 35 models on identical multi-language sample paragraphs).
• Vision token allocations: Gemini API documentation (Gemini 3 image endpoints) and OpenAI documentation (GPT Image medium quality).
• Token fundamentals and interactive calculator: AI Tokens Explained.

Kinetyra