Home / Research / AI Inference Cost Statistics

AI Inference Cost Statistics: September 2026

By Darin ten Bruggencate, Founder, AI Industry Reviews · September 16, 2026

Every time an AI model answers a question or writes a reply, its provider charges for the work, a step called inference. Headline prices for the best model on the market have barely moved in three years. But the price of reaching a fixed level of quality has fallen sharply, so we looked at what running AI models costs now.

We priced 164 commercially available models on published input, output, cached, and batch rates. Where a provider splits rates by context length, we used the rate that applies below 200,000 tokens. The blended rate in the tables weights input at 75 percent and output at 25 percent.

AI Inference Cost by Model Tier

Tiers follow capability and positioning. Several providers now serve frontier-class quality from models they price as mid-tier.

The AI Inference Cost by Model Tier, 2026

Insights

• The blended rate of the frontier reasoning tier is 128x above the small-model tier, wider than the 94x spread measured in the same exercise a year earlier.
• Output tokens price at a median of 5.0 times input tokens, a ratio that has held within half a point across every tier and every quarter tracked.
• 48% of models price below $1 per million blended tokens, so a team that matches each task to the cheapest capable model can keep most of its work inexpensive.

Blended Inference Cost per Million Tokens, Q1 2023 to Q3 2026

Headline prices for the best available model have barely moved in three years, which is why the cost of inference is so often described as flat. The more useful series holds capability constant and asks what the cheapest model meeting a fixed quality bar charges. In the table below, we show all three series, followed by the same data plotted on a logarithmic scale.

The Blended Inference Cost per Million Tokens, 2026

The capability series fell by a factor of 268 over the period, an average of roughly 5x per year, while the frontier series fell by 71 percent through the third quarter of 2025 and then reversed. Providers price the frontier at whatever the newest capability will bear, so that line follows what buyers will pay for new capability. Once a capability is two generations old, distilled and open-weight models compete on price alone, and the floor collapses. The 2026 uptick in the frontier line is the first sustained rise recorded in this series, and it coincides with reasoning models that bill for far more output tokens per answer than their predecessors did.

Inference Cost by Workload Type

A rate per million tokens tells a buyer very little without a token count attached to it. In the table below, we report the median token consumption measured for six common production workloads, along with what each one costs at mid-tier and budget-tier rates.

The Inference Cost by Workload Type, 2026

The agentic session costs roughly 370 times what a single chat reply costs, and 96 percent of its tokens, carrying 82 percent of its cost, are input. The cause is structural: an agent resends its accumulated context on every turn, so a fifty-turn session pays for the same files and the same conversation history dozens of times over. In our sample, doubling the number of turns in a session raised its cost by a factor of 3.4, well above a linear doubling, and two runs of an identical task on an identical agent differed in cost by as much as 26 times, driven by how many times the agent chose to re-read its working set. Forecasting inference spend from an average cost per request badly underestimates agentic workloads.

Discounts by Cost Reduction Mechanism

Every major provider now sells the same tokens at several prices depending on how the request is submitted. In the table below, we quantify the median discount each mechanism delivered across the providers offering it.

The Discounts by Cost Reduction Mechanism, 2026

Insights

• Cached input is the single largest lever available, and the one most often left unused: only 38 percent of the production deployments surveyed had structured their prompts so that the stable portion is at the front where it can be cached.
• Batch submission and cached reads are additive at most providers, producing an effective discount of 94 percent on repeated input tokens submitted asynchronously.

Self-Hosting Break-Even by API Tier

The alternative to paying per token is renting hardware and serving a model directly. In the table below, we solve for the monthly token volume at which an eight-GPU H100 node, rented at the median specialist cloud rate of $3.28 per GPU-hour and costing $18,900 per month, matches the cost of buying the same tokens from an API.

The Self-Hosting Break-Even by API Tier, 2026

The break-even volumes above count only the rented hardware. They exclude the engineering time to deploy, quantize, monitor, and update a served model, which in our sample added between 30 and 100 percent to the true cost for teams under twenty engineers. They also assume a node kept near full utilization, which almost never describes real traffic: a workload with a weekday peak and a quiet weekend runs at 30 to 40 percent average utilization unless it is deliberately batched. Self-hosting pays off against frontier pricing at volumes a mid-sized company can reach. It rarely pays off against open-weight hosted pricing, because the providers of those endpoints already run the same hardware at utilization rates a single tenant cannot match.

What AI Inference Cost Trends Mean for 2027 and Beyond

Running everyday tasks on AI keeps getting cheaper, because small models now cost very little to use. The most capable models still cost far more, so the price a company pays depends mainly on which model it sends each task to and whether it uses discounts like caching and batching.

Sorting work by difficulty and sending only the hard tasks to expensive models cuts cost. Many companies still leave easy savings unused, so costs could fall a lot without any change in quality.

So if your business pays for AI by usage, check which tasks really need the most capable model and turn on caching where your prompts repeat.

Requesting a Copy of This Report

If you would like a PDF copy of this report, or the full pricing dataset behind it, you can reach out here.