Remember when running a single GPT-4 query felt like burning cash? In 2023, you paid roughly $60 per million tokens. Today, that same quality of output costs just $0.75 per million tokens. That is a 98% drop in three years. If you are building or scaling an AI product right now, this shift isn't just a nice-to-have; it changes your entire unit economics. The market is splitting into two distinct worlds: cheap, fast commodity models for everyday tasks, and expensive, slow premium models for complex reasoning. Understanding where your use case fits in this split is the key to keeping your margins alive.
The Current State of LLM Pricing
The landscape as of August 2026 is no longer about one size fits all. Prices vary wildly depending on whether you need a model to summarize emails or solve novel mathematical proofs. We see a clear tiering system emerging. On the low end, Meta's Llama 4 Maverick offers input tokens at a median of $0.27 and output tokens at $0.85 per million. This model supports a massive context window of over 1 million tokens, making it ideal for processing long documents without breaking the bank. Slightly higher up, DeepSeek R1 sits around $0.60 for input and $2.34 for output. These are the workhorses of the industry. They handle summarization, basic question answering, and content generation efficiently because they don't require heavy internal computation for every token generated.
On the other side of the spectrum, we have the premium tier. OpenAI's GPT-5.2 Pro charges $21 per million input tokens and a staggering $168 per million output tokens. That is an 8x multiplier between input and output. Why so high? Because these models perform deep, internal chain-of-thought reasoning. They aren't just predicting the next word; they are simulating multiple paths of logic before giving you an answer. For tasks where accuracy is non-negotiable, like legal contract review or complex financial modeling, this price tag might be worth it. But for general chatbots? It’s overkill. The median output-to-input ratio across the market is about 4x, but premium models push that to 8x, which means verbose responses become prohibitively expensive if you aren't careful with your prompts.
Why Prices Are Dropping So Fast
You might wonder how prices fell so dramatically in such a short time. It’s not just marketing tricks. Three technical shifts are driving this down. First, Mixture-of-Experts (MoE) architectures have become standard. Instead of activating every parameter in a model for every request, MoE models only wake up a subset of experts relevant to the specific task. This drastically reduces compute requirements per token, allowing providers to lower their rates while maintaining high performance.
Second, Speculative Decoding is improving throughput. This technique pairs a smaller, faster draft model with a larger, more accurate one. The small model guesses the next few tokens, and the large model verifies them in parallel. If the guess is right, you save significant compute time. This is particularly effective for deterministic tasks where the path is predictable.
Third, quantization is changing the game. Running models at 4-bit or 8-bit precision instead of full 16-bit allows inference on cheaper hardware. You lose a tiny bit of precision, but for many applications, the difference is imperceptible. This efficiency gain permits deployment on consumer-grade hardware or lower-cost cloud instances, further squeezing infrastructure costs. When your server bill drops by half, you can pass some of that savings to the customer. This mirrors what happened in the early days of cloud computing, where basic compute capacity became a commodity as providers achieved scale.
The Hidden Costs You Might Be Ignoring
Token pricing is only part of the story. If you are using reasoning models, you need to watch out for "hidden" tokens. Providers like OpenAI and Anthropic introduced reasoning tokens that incur higher costs than standard output tokens. These are the internal steps the model takes to think through a problem. Even if the final answer is short, the model might have processed thousands of internal tokens to get there. In long analytical workflows, these can add up quickly. A simple data analysis task might seem cheap per token, but if the model spends 5,000 reasoning tokens to figure out the best approach, your bill jumps significantly.
Beyond the core model, your total cost of ownership includes vector database operations. Services like Pinecone or Weaviate charge per query and per gigabyte of storage. If you are storing millions of embeddings for retrieval-augmented generation (RAG), these costs can rival your inference costs. Additionally, reranking models used to filter results before sending them to the main LLM add another layer of expense. Always calculate your total monthly spend including storage, retrieval, and post-processing, not just the raw token count from the API call.
Commodity vs. Premium: The Market Split
The market is bifurcating. General-purpose language models are becoming commodities. Think of them like electricity. You don’t pay extra for the electrons that power your lights; you pay for the grid service. As open-source alternatives proliferate and competition intensifies, the price for basic capabilities-summarizing, translating, drafting-will continue to fall. This segment is characterized by high volume and low margin. Providers compete on speed, context length, and price.
Premium models, however, will maintain their pricing power. These are specialized tools for high-precision analytical tasks. They command a premium because they offer something hard to replicate: reliability in complex scenarios. If you are building a medical diagnosis assistant or a legal advisor, you can’t afford for the model to hallucinate. The computational intensity required to ensure accuracy justifies the higher price. This divide resembles the software industry, where basic utilities are free or cheap, but enterprise-grade solutions with guaranteed uptime and support carry hefty price tags.
| Model Tier | Example Models | Avg Input Cost ($/M Tokens) | Avg Output Cost ($/M Tokens) | Best Use Case |
|---|---|---|---|---|
| Commodity / Entry-Level | Llama 4 Maverick, GPT-4o Mini | $0.00 - $0.30 | $0.50 - $1.00 | Chatbots, Summarization, Basic QA |
| Mid-Tier / Balanced | DeepSeek R1, Claude Sonnet 4 | $0.60 - $3.00 | $2.30 - $15.00 | Complex Drafting, Data Analysis, Coding |
| Premium / Reasoning | GPT-5.2 Pro, Claude Opus 4 | $15.00 - $21.00 | $75.00 - $168.00 | Legal Review, Scientific Research, High-Stakes Logic |
Shifting Business Models: Beyond Per-Token Billing
Per-token billing is great for developers who understand APIs, but it confuses business users. Who wants to estimate how many words a summary will be? To address this, we are seeing a rise in per-action pricing. This model assigns a fixed cost to a defined task, such as "Review Contract" or "Extract Invoice Data." This simplifies budgeting for non-technical teams. If you know you will process 10,000 invoices a month, you can predict your cost exactly. It aligns vendor incentives with customer value perception because you are paying for the outcome, not the raw compute.
We are also seeing the emergence of Service Level Agreement (SLA) based tiers. Enterprise customers often care less about the exact token count and more about latency guarantees and uptime. SLA-based pricing allows providers to charge a premium for guaranteed performance windows. For example, a plan might guarantee sub-second response times for critical business processes. This shift indicates that the market is maturing. We are moving away from selling raw intelligence and toward selling reliable, integrated business capabilities.
What to Expect in 2027 and Beyond
Looking ahead, the trend of price compression for general models will continue. As 7-billion parameter models achieve scores that previously required 70+ billion parameters, the efficiency gap widens. You will see capable models running on local laptops, reducing dependence on cloud APIs for basic tasks. However, the premium segment will remain robust. Multimodal capabilities-combining text, image, and video understanding-will drive new pricing structures. Processing video tokens is far more expensive than text, so expect specialized pricing for media-heavy applications.
For businesses, the strategy should be hybrid. Use cheap, fast models for 80% of your routine tasks. Reserve premium models for the 20% of cases where error tolerance is zero. Monitor the release cycles of open-source models closely; every six months, a new open-weight model often matches the previous year’s flagship closed model. By staying agile and ready to switch providers, you can keep your AI costs aligned with the rapid pace of technological improvement. The goal isn't just to cut costs, but to optimize the balance between capability and expense for each specific workflow.
Are LLM prices going to stop dropping?
For general-purpose models, prices will likely continue to drop due to increased competition and improved efficiency. However, premium reasoning models may stabilize or even increase in price as they differentiate themselves through superior accuracy and specialized capabilities. The market is splitting, so the "average" price is misleading; you must look at specific tiers.
For general-purpose models, prices will likely continue to drop due to increased competition and improved efficiency. However, premium reasoning models may stabilize or even increase in price as they differentiate themselves through superior accuracy and specialized capabilities. The market is splitting, so the "average" price is misleading; you must look at specific tiers.
What is the difference between input and output token pricing?
Input tokens are the words you send to the model (prompts, context). Output tokens are the words the model generates. Output tokens usually cost more because generating text requires more computational effort than reading it. The ratio varies by model, ranging from 2x for efficient models to 8x for premium reasoning models.
Is per-action pricing better than per-token pricing?
It depends on your needs. Per-action pricing is better for predictable, discrete tasks like invoice processing or email classification because it simplifies budgeting. Per-token pricing is better for variable, creative, or open-ended tasks where the length of the response is unpredictable. Many companies use a mix of both.
How do hidden reasoning tokens affect my bill?
Reasoning tokens are internal steps the model takes to solve complex problems. They are billed separately and often at a higher rate than standard output tokens. If you use a reasoning model for a simple task, you might pay for extensive internal thinking that doesn't appear in the final answer. To avoid this, use simpler models for straightforward queries.
Should I switch to open-source models to save money?
Yes, for many use cases. Open-source models like Llama 4 Maverick offer competitive performance at a fraction of the cost of proprietary equivalents. However, you must factor in the cost of hosting and maintenance. If you have the engineering resources to manage self-hosted models, the savings can be significant. If not, managed open-source options via cloud providers are a good middle ground.
