Every AI Cost Tracker Has a Hand-Typed Price List. Monday Breaks Them.

Claude Sonnet 5 introductory pricing ends September 1, 2026, moving from $2 and $10 per million tokens to $3 and $15. Most developers have no idea what they currently spend, because Claude Code, Codex, Cursor plus Gemini CLI each write session logs to separate places, none of which talk to each other. A category of local tools reads those logs and adds up the cost. A new one called tokentab launched yesterday and picked up 213 stars in a day. But every tool in this category prices tokens from a hand-maintained table in its own source code, deliberately avoiding the network. Which means on Monday morning they will all under-report Sonnet 5 by roughly a third until somebody edits a file and ships a release. Best for finding out what you spend before the price moves. Not ideal for trusting the number afterward.

Your machine already knows what you have been spending.

Claude Code writes JSONL files to ~/.claude/projects/. Gemini CLI logs to ~/.gemini/tmp/. Codex uses ~/.codex/sessions/. Cursor keeps everything in a SQLite database.

Every one of those files records token counts per call. Nothing is estimated. The numbers are sitting on your disk right now.

None of those tools talk to each other, so if you use more than one, which most people do, you are left doing arithmetic across four dashboards that agree on nothing.

On Monday, Sonnet 5 gets 50% more expensive. Which makes this a reasonable week to find out what you are actually spending.

What Changes on September 1

The specifics, since a lot of coverage has been vague about it.

That is a 50% rise on both. There is also a reported tokenizer change adding up to 35% more tokens per equivalent piece of text, which compounds the increase for anyone processing the same workload.

Subscription users on Claude Code do not pay these rates directly. What they pay for is a quota, which is priced against these numbers. When the underlying rate goes up, quota policy tends to follow, which is what we saw with the Codex five-hour limit returning this week.

Anyone using the API directly just pays more from Monday.

tokentab, Verified

The tool that surfaced this week, checked before writing.

It reads the session logs those tools already leave on disk and adds up tokens plus cost, broken down by model, by project, by day, plus by the kind of work each session was doing. No account, no API key, nothing leaves your machine. If a tool is not installed it simply gets skipped.

Install is a clone plus pip install -e ., which puts a tokentab command on your path.

tokentab              # last 7 days across everything
tokentab today
tokentab month
tokentab -p all       # everything you have ever run
tokentab --provider claude

The caching handling is the part that shows care. Claude splits cache reads and writes into separate figures while Gemini reports input including the cached portion, so tokentab pulls cached tokens back out before pricing, avoiding a double charge for the same thing.

The Price Table Is a Python File

Here is the detail that makes this a story rather than a tool announcement.

tokentab’s prices live in tokentab/pricing/prices.py, in dollars per million tokens. It is a table on purpose. The author’s stated reasoning is that the tool never reaches out to the network to price anything, since a slightly stale number beats a crash when a vendor renames a model overnight.

That is a defensible engineering decision. It is also the reason your cost report breaks on Monday.

At 00:00 on September 1, Sonnet 5 costs $3 and $15. Every local tracker will keep pricing it at $2 and $10 until a human edits that file, tags a release, then you update. Your reported spend will sit roughly a third below what you are actually billed. The tool will not warn you, because from its perspective nothing has changed.

The fuzzy match makes it worse

Model matching is fuzzy, so claude-sonnet-5-20260514 still finds claude-sonnet-5. Helpful most days. This week it means the wrong price attaches confidently to the right model.

There is one saving grace worth knowing: if a model shows up at $0.00, that means the name did not match anything in the table. A zero is the tool telling you it does not know. A wrong number looks exactly like a right one.

This Category Is More Crowded Than It Looks

tokentab is one day old and it is not first. Anyone about to install something should know the alternatives, all verified.

toktrack, 186 stars, 214 days old, written in Rust. Covers eight CLIs including Claude Code, Copilot, Codex, Gemini, Qwen Code, OpenCode, Pi Agent plus Antigravity. Its distinguishing feature is a persistent cache. The reason for that cache is the single most useful fact in this article.

TokenBar, 302 stars, 79 days old, native Swift for the macOS menu bar. Tracks 25+ agents on-device, shows today’s tokens, cost, live tokens per minute, or remaining subscription quota. No dock icon, no telemetry, no account.

codeburn runs via npx codeburn and claims coverage across 37 tools and agents, with a menu bar plus web dashboard.

TokenTelemetry is a local dashboard covering coding agents plus autonomous ones, including a dedicated surface for Hermes Agent across 38 source platforms.

SessionWatcher is the paid option, a macOS menu bar app from $6.99 one-time, up to $59 for the tier covering everything.

Five free tools and one paid one, all reading the same logs off your disk. The differences are coverage, interface, plus whether they keep history.

Your Cost History Is Being Deleted Right Now

This is the fact worth acting on today, regardless of which tool you pick.

Claude Code deletes its session files after 30 days.

Most trackers, tokentab included, re-read those session files on every run. They hold no state of their own. So when Claude Code cleans up, your cost history goes with it. No tool reading live files can show you a month you no longer have.

toktrack’s entire pitch is built on this. It maintains a persistent cache specifically so history survives the CLI deleting its own logs.

Which means anybody who has been coding with Claude Code since spring has already lost most of their record. Whatever exists today is roughly the last 30 days and nothing before it.

If you want to understand your spending pattern across a longer window, that window closes a bit more every day. There is no recovering it later.

What You Actually Learn From Running One

Worth setting expectations, because the output is more interesting than the total.

The total is the least useful number. What matters is the breakdown by model plus by kind of work, because that is where the decisions live.

Output tokens cost between six and ten times more than input tokens across essentially every model. A workflow generating verbose output costs dramatically more than one reading a lot then replying briefly, even when the token counts look similar at a glance. Most people have never seen that split for their own usage.

Per-project breakdown tells you which repository is expensive, which is usually not the one you would guess.

And the model split tells you whether you are reaching for an expensive model out of habit. If most of your spend sits on a flagship doing work a cheaper model handles fine, that is the finding that pays for the afternoon.

None of these tools charge anything to show you that, since they are reading files you already have.

How to Do This in Ten Minutes

Concrete steps, since the window is short.

Pick one tool. On Linux or macOS wanting a number fast, tokentab is a clone plus a pip install. Wanting it permanently visible on macOS, TokenBar sits in the menu bar. Caring about keeping the history, toktrack is the one that survives log deletion.

Run it across everything you have:

tokentab -p all

That gives you your entire recorded history rather than the default seven days. Note whatever comes back, because on Monday the same command returns a differently-wrong number.

Then read three things off the output.

The monthly total, which tells you the scale of the decision. Then the split between input and output tokens, since output costs six to ten times more and that ratio is where savings hide. Finally the split by model, which tells you whether expensive models are doing work cheap ones could handle.

Write those three numbers down somewhere outside the tool. They are your baseline. After Monday no tracker will reproduce them correctly until it gets patched.

Ten minutes, no account, no signup, nothing leaves your machine.

What the Vendors Show You Instead

Worth understanding why this gap exists at all, because it is not an oversight.

Anthropic shows Anthropic usage. OpenAI shows OpenAI usage. Google shows Google usage. Each dashboard is competent within its own boundary, plus completely blind outside it.

Subscription users get a thinner picture still. What arrives is quota remaining, expressed as a proportion rather than a dollar figure. That is deliberate. A percentage bar tells you when to stop working. It does not tell you what the work was worth, which is the number that would let you compare your plan against paying per token, or against a competitor, or against buying hardware.

None of that is sinister. No business volunteers the comparison that might lose it a customer.

It does mean the only honest total comes from tools built by people with no stake in the answer, reading the logs the vendors were obliged to write anyway. Which is exactly the same shape as the open weights argument we keep returning to. Transparency arrives from outside or not at all.

The Honest Limits of the Number

Four caveats, because a confident dollar figure invites more trust than it has earned.

Prices are hand-maintained. Covered above. This week it matters more than usual.

Some tools do not record their own cost. Gemini, Qwen, Codex, Antigravity, GitHub Copilot plus modern Claude logs all report tokens without prices attached, so the tracker computes cost from a list-price estimate. toktrack marks these with a tilde to signal an estimate. Not every tool is that careful.

Subscription users are not being billed these numbers at all. If you are on a Claude Code or Codex plan, the figure a tracker shows is what the same work would cost at pay-per-token rates. It is a useful comparison, not your bill. Read it as an answer to whether the subscription is worth it.

A zero means unknown, not free. If a model name does not match the price table, it prices at $0.00. Anyone scanning a summary for a total will read that as cheap.

Does the Rise Actually Change Your Decision

Run a tracker, then do this arithmetic.

Take your monthly Sonnet 5 spend and add 50%. Add more if your workload is text-heavy, given the reported tokenizer change. That is your September number if nothing else changes.

If that figure stays under a subscription price, nothing changes for you. Keep doing what you are doing. Our Claude Pro review covers what that tier includes.

If it goes meaningfully above, you have three moves. Route cheaper work to a cheaper model, which the per-model breakdown makes obvious. Move to a subscription then live with quota limits, which now include the returning five-hour window on Codex. Or run models locally, which costs hardware and setup rather than tokens.

That third option changed this month. We covered an inference engine from Berkeley running a 35B model at 39.3 tokens per second on an 8GB laptop GPU, against a 33 token median in real Codex production traces. It speaks Anthropic-compatible APIs, so Claude Code can point at it. A year ago that was a research demo. Now it is a hardware purchase with a payback period. A tracker is how you work out whether that period is short enough to bother.

Why Everyone Is Building the Same Thing

Six tools solving one problem in three months is a signal about the problem rather than the tools.

The reason is that vendors have no incentive to build this. Anthropic will happily show you your Anthropic spending. It has no reason to build the view that puts its numbers next to OpenAI’s and Google’s on one screen, sorted by cost.

So the cross-vendor view only exists because independent developers keep building it. Every one of these tools is free, local, reading files the vendors already write.

That is the same dynamic behind a lot of what we have covered this month. DeepSeek gave away an entire agent runtime then raised model prices four days later. The community immediately built a plugin marketplace, four desktop clients plus a router around it. When a vendor leaves a gap, somebody fills it inside a week.

The gap here is that nobody selling you tokens wants to make the total easy to see.

Which is worth remembering when the next round of pricing changes arrives, because there will be one. Every major vendor repriced something in August. DeepSeek scrapped flat pricing then raised V4 Flash 93%. Grok added a clause doubling the rate on long requests. OpenAI cut Luna 80% while metering Codex more tightly. Anthropic is ending an introductory rate.

The pattern across all of it is cheap acquisition at the top of the funnel, rising prices where the usage actually sits. A tracker will not stop that. It will at least tell you when it happens to you.

The Part Worth Keeping

The numbers are already on your disk. Every tool you use writes down exactly what it consumed, then never adds it up for you.

Six separate people have now built the thing that adds it up. All of them price the result from a list somebody typed by hand into a source file.

On Monday one line in that list becomes wrong by 50%, in every one of those tools at once. None of them will say anything. The report will look exactly as confident as it did on Sunday.

Which is a decent argument for running one this weekend, while the number still means what it says.

It takes ten minutes. The alternative is finding out in October from a credit card statement.

Charts and Blocks

The tools, verified

What Monday does to your report