LLM Pricing Guide: Every Major Model Compared

There’s a pattern every developer stumbles into. You estimate your monthly AI spend during prototyping, watch it come in at $80, think “manageable,” ship, and then see a $4,200 invoice three weeks later. The token pricing page looked simple. What went wrong?

Usually several things at once. The context window was fuller than expected. Output tokens — which cost three to five times more per token than input on every major platform — ballooned. Reasoning models generated thousands of internal “thinking” tokens that billed as output. And nobody remembered to turn on prompt caching.

This guide tries to give you the complete picture. Not just headline rates, but the multipliers, batch discounts, caching math, and routing decisions that determine what you actually pay. We cover OpenAI, Anthropic, Google, DeepSeek, Mistral, and xAI’s Grok — every model family worth knowing about as of May 2026.

Token-Based Billing: The Mechanics

Every major LLM API bills by the token, not by the request. A token is roughly 0.75 words in English — so “1 million tokens” is about 750,000 words, or roughly six standard novels. That sounds enormous, but in practice a single well-structured RAG pipeline with a 4,000-token system prompt, five retrieved chunks, and a 500-token user query can consume 8,000+ input tokens before the model writes a single word.

The billing split matters enormously: input tokens and output tokens are priced separately, and output is almost always more expensive. The ratio varies by provider, but typically runs 3:1 to 5:1 output vs. input. That asymmetry is there for a real reason — generating tokens (autoregressive decoding) is computationally more intensive than processing the prompt (parallel encoding).

The Key Pricing Variables

The Price Collapse: 2022 to 2026

GPT-4 launched in early 2023 at $30 per million input tokens. Two years later, GPT-4-equivalent capability was available for around $2.50. By mid-2026, models matching GPT-4’s 2023 benchmark scores run for under $0.50 per million tokens on several platforms. Epoch AI’s analysis documents this trend in detail, finding that the price to achieve GPT-4’s performance on PhD-level science questions (GPQA Diamond) fell by roughly 40× per year between early 2023 and early 2025.

The acceleration is real but uneven. Standard text generation models fell faster than reasoning models. And at the absolute frontier — Claude Opus 4.7, GPT-5.5 Pro — pricing has actually been relatively stable over the past 18 months even as capability improved, because demand is high and the compute is genuinely expensive.

The forces driving these declines are well-understood: hardware improvements (H100s, next-gen Trainium, TPU v6), competition from DeepSeek’s open-weight models, and Mixture-of-Experts architectures that activate only a fraction of parameters per token. Introl’s infrastructure analysis documents GPT-4 equivalent performance dropping from $20/M tokens in late 2022 to roughly $0.40/M by end of 2025 — a 50× reduction in under three years.

This is Jevons’ paradox in action. Cheaper tokens lead to more ambitious architectures, longer contexts, more agent steps, and more reasoning — all of which consume tokens at rates that outpace the price drop. The engineers who benefit most from cheap tokens are also the ones most likely to build systems that need a lot of them.

Claude: The 4.x Family

Anthropic’s current lineup — Opus 4.7, Sonnet 4.6, and Haiku 4.5 — is structured around three clearly distinct tiers. The price points have been surprisingly stable since the 4.5 generation launched in late 2025. Anthropic hasn’t raised prices on a new model release in over a year, which is unusual in this market.

The Sonnet 4.6 vs. Opus 4.7 Decision

On SWE-bench Verified — the most widely cited benchmark for real-world coding tasks — the gap between Opus 4.6 and Sonnet 4.6 is just 1.2 percentage points: 80.8% vs. 79.6%. That’s a surprisingly small quality delta for a 40% price premium. For most production coding pipelines, Sonnet 4.6 is the right default unless you specifically need Opus-level autonomous agent capability or computer-use tasks.

Haiku 4.5 at 73.3% SWE-bench is within 7.5 points of Opus — while costing one-fifth as much per output token. For classification, moderation, extraction, and any task where you can tolerate an accuracy trade-off, Haiku has an overwhelming cost argument.

Prompt Caching: Anthropic’s Most Powerful Cost Feature

Cache reads on Anthropic cost just 10% of the standard input price. If your application reuses a large system prompt, a document corpus, or tool definitions across many requests, caching can cut your input costs by 90% on those tokens. The math compounds aggressively. A content moderation pipeline pushing one million posts per day through Haiku 4.5, with a 2,000-token shared system prompt, saves roughly $1,800/day from caching that prompt alone — before any other optimizations.

OpenAI: GPT-5.x, GPT-4.1, and the O-Series

OpenAI’s model lineup is now the broadest of any provider, which makes it genuinely confusing. There are four distinct families — GPT-5.x, GPT-4.1, GPT-4o, and the o-series reasoning models — each with their own pricing logic and use-case fit. The ranges are extreme: GPT-4.1 Nano at $0.10/M input vs. GPT-5.5 Pro at $30.00/M input, a 300× spread within a single provider’s catalog.

The GPT-4.1 Family: The Underrated Value Play

GPT-4.1 at $2.00/$8.00 with a 1M token context window is one of the more interesting pricing positions in the current market. For long-document processing — legal contract analysis, codebase reviews, lengthy RAG contexts — it’s cheaper than GPT-5.4 despite the lower per-token rate on short prompts, because you can batch more into a single request. GPT-4.1 mini at $0.40/$1.60 extends that advantage further down the cost curve.

Google Gemini: The Context Window Kings

Google’s Gemini family has become genuinely competitive in a way that wasn’t obvious 18 months ago. The free tier is the most generous in the industry, and the Gemini 2.5 Pro’s pricing — especially with prompt caching — is hard to beat for long-context work. The tiered pricing structure (below vs. above 200K tokens) is worth understanding before you build anything with long documents.

Gemini 2.5 Flash at $0.30/$2.50 is a remarkable data point. It supports a 1M token context window, handles multimodal input, and benchmarks competitively with models costing three to four times as much. For many mid-complexity use cases — document Q&A, structured extraction, summarization — it’s the most defensible budget choice from a major provider.

The Long-Context Calculation for Gemini Pro

Gemini 2.5 Pro’s tiered pricing deserves a specific example. If you’re processing a 500,000-token document (about 375,000 words), the first 200,000 tokens cost $1.25/M and the remaining 300,000 cost $2.50/M. Total input cost: $0.25 + $0.75 = $1.00. Output at $15/M for the same tier. For very long-context work, the math is reasonable, but plan carefully — the jump from $10/M to $15/M output above 200K is significant if your responses are long.

DeepSeek: The Pricing Disruptor

DeepSeek’s January 2025 release reshaped the conversation around LLM economics. V3 at $0.14/$0.28 per million tokens — later refined to V3.2 — offered performance comparable to the frontier Western models at a fraction of the cost. The reaction from established providers was swift: a wave of price cuts across the industry in early 2025 can be traced directly to DeepSeek’s market entry.

The cache pricing is worth dwelling on. DeepSeek charges $0.0028 per million cached input tokens — that’s 0.28 cents, or roughly two percent of what GPT-4.1 charges for the same tokens. For agent workflows with large, repeated system contexts, DeepSeek’s effective cost approaches zero on the input side.

What DeepSeek Doesn’t Tell You

The raw price is extraordinary, but several practical factors complicate the picture for production Western deployments. Data residency is primarily in China, which creates compliance concerns for healthcare, finance, and any business handling European user data under GDPR. Latency can be higher than US-hosted alternatives. And reliability SLAs are less mature than what you get from OpenAI or Anthropic. For non-sensitive, internal, or batch workloads? DeepSeek is hard to beat. For customer-facing production at enterprise scale with data residency requirements? Factor in the tradeoffs carefully.

Mistral, Grok, and Llama

The Complete Comparison: Every Major Model

Cost Optimization Strategies That Actually Move the Needle

Most cost optimization guides list ten things and nothing specific. Let’s be direct: three levers account for most of the savings available to a typical production team.

1. Model Selection (Biggest Impact)

The choice of model is the single largest cost lever. Switching from GPT-5.4 ($2.50/$15) to GPT-4.1 Nano ($0.10/$0.40) for classification queries can reduce costs by over 90%. The catch, of course, is that you can’t use Nano for everything — you need to know which tasks actually require frontier capability and which are genuinely routine.

A useful heuristic: run a sample of 500 real queries through both your current model and a cheaper one. If accuracy on the cheaper model is within 3-5 percentage points and errors are recoverable, it’s probably worth the switch. If errors are consequential and non-recoverable (medical decisions, financial transactions, legal documents), stay with the higher-capability model regardless of cost.

2. Prompt Caching (High Impact, Often Overlooked)

3. Batch API (High Impact for Non-Realtime Workloads)

Both OpenAI and Anthropic offer 50% discounts for batch processing — requests submitted asynchronously with up to 24-hour completion time. For analytics, nightly data processing, bulk classification, content generation pipelines that don’t need immediate responses, or any task running outside of a live user session, the batch API should be the default. There’s no meaningful downside other than latency, and half-price is half-price.

Additional Levers Worth Considering

The Model Routing Playbook

The most sophisticated production AI teams don’t use a single model — they route queries to different models based on complexity and cost tolerance. This isn’t complicated to implement, and the savings are real.

A Practical Routing Framework

Real-World Cost Scenarios

Abstract pricing comparisons have limits. Here are four concrete scenarios, built from common production architectures, with actual numbers.

Frequently Asked Questions

What’s the cheapest capable model available right now?

For pure token price, GPT-4.1 Nano and Gemini 2.5 Flash-Lite both hit $0.10/M input and $0.40/M output. DeepSeek V3.2 at $0.14/$0.28 is slightly more expensive but stronger on reasoning tasks. For production-quality generation with no data residency concerns, Gemini 2.5 Flash at $0.30/$2.50 offers the best balance of capability and cost at the budget tier.

Should I use Claude Opus 4.7 or 4.6?

For most workloads, 4.6. Opus 4.7 uses a new tokenizer that can generate up to 35% more tokens for the same input text — same sticker price, higher effective cost. If your prompts are short and unstructured (mainly code, prose), the tokenizer difference is smaller. If they’re data-heavy (JSON, structured output, tabular data), test before migrating. The quality uplift is real, but it’s concentrated in autonomous coding agents and computer-use scenarios. For general reasoning and analysis, the improvement over 4.6 is modest.

Is DeepSeek worth using in production?

For internal tooling, analytics pipelines, and non-sensitive workloads: genuinely hard to beat on economics. For anything touching user data subject to GDPR, healthcare records, financial data with specific residency requirements, or classified/sensitive information: the compliance picture is complicated, and most enterprise legal teams will push back. Evaluate your specific constraints before deploying.

How much does context window size affect cost?

Larger context windows don’t directly cost more per token — you pay the same rate regardless of context size. The problem is that larger windows encourage you to put more in them. A 1M token context window is only useful if you actually fill it, and filling it means spending on every token you send. RAG with a 200K window and precise retrieval is almost always cheaper than naive full-document loading with a 1M window.

What’s the real cost of reasoning models like o3?

Higher than the rate card suggests, for complex queries. The listed $2.00/$8.00 per million tokens for o3 doesn’t account for internal reasoning tokens that are billed as output but invisible in the response. On complex multi-step problems, visible output tokens may be 20–30% of total output consumption. For simple queries, the overhead is low. Monitor your actual token counts, not just visible output, if you’re using reasoning models at scale.

What are subscription plans and how do they differ from API pricing?

Subscription plans (Claude Pro at $20/month, ChatGPT Plus at $20/month, etc.) give individual users access to the consumer interfaces with usage limits. They don’t cover API access. The API is metered separately by token. Enterprise and team plans (Claude Team, ChatGPT Enterprise, Google Workspace AI) are per-seat, not per-token, and bundle product features with usage quotas. They’re structured differently and generally priced for business procurement rather than developer billing.