The activity page is easy to read: 10,000 tokens on a quiet day, 100,000 on a normal one, a million when a job runs long. Finance wants the month. The catch is that input and output are priced apart inside the same request, and GPT, Claude, Gemini, and DeepSeek each sit on several model IDs whose rates can differ by an order of magnitude. Before you multiply by 30, you have to decide which ID, and which mix, that daily total belongs to.
The figures below use catalog rates read from OpenRouter on 13 October 2026, a 30-day month, and a 3:1 input-to-output mix, with one everyday ID per family. Change the ID or the mix and the same day lands in a different cell. For how to pair models with tasks, see How to Choose an AI API After OpenAI DevDay 2026? OpenAI, Claude, and Gemini Model Pricing, Speed, and Agent Capability Comparison. Subscriptions, the agent shell, and the always-on machine are a separate bill: How Much Does an AI Coding Agent Cost Per Month? 2026 Total Cost of Subscriptions, Tokens, API, and Servers for Claude Code, Codex, Gemini, and Cursor.
What the activity total is actually counting
OpenRouter lists input (prompt) and output (completion) separately, in dollars per million tokens. A line that says “tokens today” is usually the sum, so it cannot be multiplied by just one of those columns. Rates come from pricing.prompt and pricing.completion on the models catalog as returned on 13 October 2026.
A 3:1 mix means three quarters of the day is input and one quarter is output, close to a system prompt plus one reply. A coding agent that keeps emitting long diffs spends more of the day on output, and the bill moves toward the output rate.
monthly tokens = daily tokens × 30 blended rate = (3 × input rate + output rate) ÷ 4 monthly cost = monthly tokens ÷ 1,000,000 × blended rate
Cache reads are cheaper. The tables assume no cache hits and no retries.
Which model ID each family uses here
A family name is not a price. Each row is one current, non-batch, non-image ID people actually route: GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 3.8 Flash, and DeepSeek V4.1 Flash. They are a common tier, not the cheapest or dearest SKU in the family.
| Model | Input | Output | Cache read | Blended |
|---|---|---|---|---|
GPT-6.1 Solopenai/gpt-6.1-sol |
$2.00 | $10.00 | $0.10 | $4.00 |
Claude Sonnet 5.5anthropic/claude-sonnet-5.5 |
$2.00 | $10.00 | $0.10 | $4.00 |
Gemini 3.8 Flashgoogle/gemini-3.8-flash |
$0.75 | $3.75 | $0.075 | $1.50 |
DeepSeek V4.1 Flashdeepseek/deepseek-v4.1-flash |
$0.30 | $1.20 | $0.006 | $0.525 |
How three daily volumes become a month
Ten thousand, one hundred thousand, and one million tokens a day become 300,000, 3 million, and 30 million tokens across 30 days. The table multiplies those volumes by the blended rate.
| Per day | Per month | GPT-6.1 Sol | Sonnet 5.5 | Gemini 3.8 Flash | V4.1 Flash |
|---|---|---|---|---|---|
| 10,000 | 300,000 | $1.20 | $1.20 | $0.45 | $0.16 |
| 100,000 | 3 million | $12.00 | $12.00 | $4.50 | $1.58 |
| 1 million | 30 million | $120.00 | $120.00 | $45.00 | $15.75 |
Check for Sol at 100,000 tokens a day: of 3 million tokens, 2.25 million are input and 0.75 million are output, so 2.25 × 2 + 0.75 × 10 = 12, or $12.00. DeepSeek’s $0.16 and $1.58 are $0.1575 and $1.575 rounded; 30 million tokens land on $15.75 exactly.
Where the cheap and expensive IDs sit
The main table is one everyday ID per family. Each family also has cheaper and dearer endpoints. Still at 3:1, and still at one million tokens a day (30 million a month):
| Model ID | Input / output | Blended | Month at 1M / day |
|---|---|---|---|
openai/gpt-6-luna |
$0.10 / $0.50 | $0.20 | $6.00 |
openai/gpt-6-astra |
$10.00 / $50.00 | $20.00 | $600.00 |
anthropic/claude-haiku-5.5 |
$0.10 / $0.50 | $0.20 | $6.00 |
anthropic/claude-opus-5.5 |
$4.00 / $20.00 | $8.00 | $240.00 |
anthropic/claude-fable-5.1 |
$10.00 / $50.00 | $20.00 | $600.00 |
google/gemini-3.1-pro-preview |
$2.00 / $12.00 | $4.50 | $135.00 |
deepseek/deepseek-v4-pro-0813 |
$1.32 / $3.96 | $1.98 | $59.40 |
At this volume, “GPT for a month” can be Luna at $6 or Astra at $600. A monthly quote needs the model ID, not only the family name.
What the catalog rate leaves out
- The fee on buying credit. The OpenRouter pricing page lists a 5.5% platform fee on the Standard plan, charged when you buy credit, not inside the token rates above. Covering a $120 Sol month entirely with that credit adds about $6.60, or about $126.60 in total.
- Cache and endpoint. A repeated system prompt that hits cache is billed at the cache-read rate, below the table. A different provider endpoint for the same ID can also disagree with the catalog. Check the model page before you commit.
- The token total itself moves. Retries and tool results inflate “tokens today.” The tables price a given token volume, not a given number of tasks. A 28-day February is about 6.7% under the 30-day figures.
Tokens move; the machine that stays on does not
Once OpenRouter is part of the day’s work, the bill above rises and falls with volume. The machine that runs the agent, holds the keys, and keeps the logs is a fixed monthly line. macOS already has a Unix shell, so Homebrew, SSH, and Docker are there without a compatibility layer for keys and the repo. Unified memory on Apple Silicon keeps local search, logs, and light scripts on one box, and an M4 Mac mini idles at about 4W, which is what you want from a machine left on.
Against a similarly priced Windows PC that also has to stay up, macOS crashes less often and is easier to leave unattended. Gatekeeper, SIP, and FileVault keep keys on the machine with a smaller exposed surface. The box is small and quiet, so power and after-hours attention stay low.
If you want that always-on machine on a line of its own, apart from the API curve—see the plans and pull the fixed cost out of the token chart.