The AI Cost Index is a blended price signal for comparing model families without doing spreadsheet gymnastics. It turns mixed input and output pricing into one number per 1M tokens, so you can see what frontier and budget usage really costs.
Why a blended index is useful
Raw model pricing is annoying to compare because input and output often cost different amounts, and real workloads are rarely 50/50. A blended index gives you a practical shorthand for planning, benchmarking, and arguing with your own estimates before you ship.
For developers, the point is not to memorize every rate card. The point is to answer a simpler question: if my agent burns through a lot of context and produces a decent amount of output, am I spending closer to budget-model money or frontier-model money?
That is exactly why a blended view is useful alongside the live AI Cost Index and the broader model catalog in live model prices. It helps you move from “this seems expensive” to “this class of usage is priced like this.”
The two baskets that matter most
The current index gives you two clean reference points:
| Basket | Blended price | Blend assumption |
|---|---|---|
| Frontier | $4.64 per 1M tokens | 3:1 input:output |
| Budget | $0.74 per 1M tokens | 3:1 input:output |
Those numbers are not random. They reflect blended cost, which is the right unit when you want to estimate real usage patterns instead of cherry-picking the cheapest side of a rate card. That matters because output-heavy tasks, long-context tasks, and agent loops can behave very differently from one another.
Here is the practical takeaway: if you are building with a model that lands near the frontier basket, your bill can rise quickly once output grows. If you are building with something near the budget basket, you still need to watch volume, but the same workload is much easier to absorb.
Worked example: what 12 million tokens costs
Let’s do the math with a real workload. Say your coding agent uses 12,000,000 tokens in a month and its usage pattern matches the frontier basket.
12,000,000 tokens / 1,000,000 = 12
12 × $4.64 = $55.68So that month would cost $55.68 at the frontier blended rate.
Now run the same token volume through the budget basket:
12,000,000 tokens / 1,000,000 = 12
12 × $0.74 = $8.88Same token volume, very different bill: $55.68 versus $8.88. That gap is why model selection matters even when the token count stays constant.
If you want a more concrete coding example, imagine a tool that does lots of retrieval, context refreshes, and iterative edits. The raw token count can look tame at first, but once you multiply by the blended rate, the monthly number starts to feel real. This is the same reason people get surprised by long-running agent workflows, even when each step seems small in isolation.
How the index compares to specific models
Blended baskets are useful for quick planning, but sometimes you need to see the actual model rates. Here is a small comparison using only current figures from the catalog.
| Model | Input | Output | Notes |
|---|---|---|---|
| gpt-4o-mini | $0.15 per 1M | $0.6 per 1M | Very low-cost baseline for many tasks |
| deepseek-chat | $0.28 per 1M | $0.42 per 1M | Cheap both ways, especially on output |
| gemini-2.5-flash | $0.3 per 1M | $2.5 per 1M | Input is cheap, output is notably higher |
| claude-sonnet-4-5 | $3 per 1M | $15 per 1M | Mid-to-frontier territory |
| gpt-4o | $2.5 per 1M | $10 per 1M | Frontier-grade generalist pricing |
| claude-opus-4-1 | $15 per 1M | $75 per 1M | Premium frontier pricing |
The table makes one thing obvious: output pricing can dominate fast. A model with cheap input but expensive output can still surprise you if your workflow asks it to explain, summarize, or generate long responses. That is why cost analysis should never stop at the prompt.
For a broader market view, the public State of AI page is useful when you want to zoom out from one model to the whole ecosystem. The point is not just which model is cheapest, but which one is cheapest for the way you actually work.
What the index tells you about real workflows
Most developer usage is not a single prompt. It is a chain: read files, ask a question, refine the answer, regenerate, then ask again after the first attempt misses. That means the same feature can land in very different cost bands depending on how often it loops.
- Short Q&A: usually low total tokens, but output can still be a meaningful share of spend.
- Refactor loops: often input-heavy because the model keeps rereading context.
- Agentic coding: can become output-heavy if the model narrates, plans, and retries a lot.
- Long-context debugging: often expensive because old context gets repeatedly re-sent.
The blended index helps you reason about those patterns without pretending every task is identical. It is a planning tool, not a replacement for actual usage data. If you want the real picture, you need telemetry, not vibes.
That is where open data helps too. The project publishes data under CC BY 4.0, which makes it easier to analyze model pricing and spend behavior without treating the numbers like a black box. If you want to dig deeper into the raw side of things, the open data is there for that.
How to track this in your own stack
If you want to see what this means for your own tools, start with npx mytokentracker. It reads the usage logs coding agents already keep locally, prints a 30-day cost table at API prices, and the usage data never leaves your machine.
If you want a synced dashboard, create a free account and run npx mytokentracker init. It asks for the API token, uploads history, then syncs every 30 minutes. It works with Claude Code, Codex, GitHub Copilot CLI, Antigravity, OpenCode, Amp, Qwen Code, and more. Only daily totals per agent and model are uploaded, never prompts, code, or file names.
For app code, the drop-in wrappers auto-capture OpenAI, Anthropic, Gemini, and Mistral. In Python, that means track_openai, track_anthropic, track_google, and track_mistral. In Node, it is trackOpenAI, trackAnthropic, and the rest of the same family. If you use another provider, a single POST to the events API is enough.
The useful part is not just cost. It also captures input, output, cache, and reasoning tokens, latency, and success, broken down by provider, model, platform, and use-case. That makes the blended index actionable, because you can compare your own workload against the public pricing signal instead of guessing.
FAQ
Is the AI Cost Index the same as a model price list?
No. A price list gives you raw input and output rates. The index compresses that into a blended number so you can compare usage patterns more quickly. It is a planning metric, not a replacement for actual rates.
Why does blended pricing matter if I already know input and output rates?
Because most real workloads mix both. If your tool sends a lot of context and gets a lot of generated text back, a blended number is closer to what you will actually pay than either side of the rate card alone.
Where should I look if I want live prices or a broader market view?
Use live model prices for the detailed catalog, and the AI Cost Index for the blended view. If you want the macro picture, the State of AI page is the right place to start.
If you want to see your own numbers instead of guessing, try tracking your AI usage locally.