Claude Max plans meter usage by tokens processed, not by the number of messages you send or hours you're logged in. A quick question costs almost nothing; a long coding session where you paste several screenshots per review costs vastly more, because those images and the growing conversation around them get re-sent, in full, with every single turn.
Updated August 2026: none of this is a glitch, and it isn't something support can fix for you. It's simply how large language models process a conversation, and it applies to every plan, every provider. Once the mechanic is visible, the fixes are mostly free and take minutes, not money. This is part of the wider picture of Claude Code token costs, which breaks down where a session's tokens go in more detail.
- 2,691 tokens — one 1080p screenshot (Claude 4.7+)
- 10,764 tokens — 4 screenshots covering one page
- 80,730 tokens — 30 seconds of screen recording frame-dumped
Here's what's actually happening under the hood, and what you can change starting with your next session.
What actually counts against your limit?
Every token processed counts — the text and code in your message, any images or files attached, Claude's reply, and, in Claude Code, the output of every tool call and file read. Message count isn't the meter; a one-line question and a message carrying a screenshot are billed differently, even though the interface shows both as one message.
Anthropic doesn't publish the exact formula that converts tokens into "usage" on a Claude.ai or Claude Code plan. What is documented is how the underlying API bills: per input token and per output token, with images tokenized by size (more below). The reasonable, and widely stated, position is that plan usage scales with that same token volume rather than with a flat per-message count. Treat anything more specific than that as inference, not confirmed fact.
This article doesn't state a specific reset window, quota, or dollar cap for Claude Max, because none of those numbers are published in a way we could verify and stand behind. If a figure isn't sourced, it isn't here.
Why does context re-send on every turn?
Claude's API is stateless — it has no memory between requests. To keep a thread coherent, every message you send carries the entire conversation so far back to the model: your first message, every reply, every screenshot, every file. A screenshot pasted at turn two is still part of the payload at turn twenty, unless something removes it.
This is why a long visual-review session compounds. If turn two costs 2,691 tokens (Claude 4.7+) for one screenshot and nothing gets trimmed, that same 2,691 tokens is reprocessed at turn three, turn four, and every turn after, stacked on top of whatever new content you're adding. Anthropic's API supports prompt caching, which can make reprocessing repeated context cheaper on the API side, but whether that lowers usage-limit consumption on a Claude.ai or Claude Code plan the same way it lowers raw API cost isn't publicly documented. The safest assumption is the plain one: what you send stays in the conversation, and gets sent again.
What does a long session do to consumption?
A long session compounds two ways at once: each new turn adds fresh tokens, and every turn re-sends everything that came before. Twenty separate 30-second visual reviews compressed into one sprawling thread can push a single conversation's payload into the hundreds of thousands of tokens — well before you've reached anything close to a fair day's work.
Here's what that looks like with real numbers:
| Same 30-second visual review, sent as… | Tokens | Type |
|---|---|---|
| 1 screenshot | 2,691 (Claude 4.7+) | calculated |
| 4 screenshots covering one page, no narration | 10,764 | calculated |
| Full recording, frame-dumped at 1 fps | 80,730 | calculated |
| Voice + pointer bundle | ~3,500 | measured (typical) |
Twenty reviews frame-dumped in a single day is 20 × 80,730 = 1,614,600 tokens (calculated), mostly pixels the model is re-deriving from a frame it already saw a second ago. Compress the same 20 reviews into voice-plus-pointer bundles instead and the day's total drops to 20 × 3,500 = 70,000 tokens (calculated): a 23× reduction (calculated: 80,730 ÷ 3,500). Priced like API access at $10 per million input tokens, that's $0.81 for one frame-dumped review versus $0.04 for one bundled review (both calculated). Claude Max is a flat subscription, not metered per token, but the ratio holds regardless of how the bill is structured.
Even without going anywhere near full video, one 30-second frame-dump alone eats roughly 40% of a 200k-token context window before you've made a single fix (calculated), and that's the same context that then gets resent, in full, on every turn after it.
Are images the hidden cost?
Yes. Anthropic's published image-token formula charges roughly one token per 28×28-pixel patch, capped per image — a full 1080p screenshot lands at exactly 2,691 tokens (Claude 4.7+). That number is invisible; the interface shows "image attached," not a cost. Four screenshots to document one page run 10,764 tokens before a single word of feedback is written.
The formula is public for Claude 4.7 and later's high-resolution tier: tokens = min(4,784, ceil(width ÷ 28) × ceil(height ÷ 28)) (Source: Anthropic, 2026). For a 1920×1080 screenshot that's 69 × 39 = 2,691 tokens, exactly, since it falls under that tier's cap. Older, standard-tier models cap out lower and downscale first, landing at 1,560 tokens for the same image. A smaller, real test screenshot at 1,288×811 measured in at 1,334 tokens: still real cost, still invisible. See how many tokens a screenshot actually costs for the breakdown across common resolutions.
Most people don't paste one screenshot per review. They paste several, because one angle rarely tells the whole story. That's the honest problem: nobody pasting four screenshots is doing anything wrong. Nobody is shown the running total either.
What changes if you compress feedback?
Pixels are expensive and low in meaning per token; words are cheap and can be precise. A spoken description of a bug runs roughly 100–200 tokens (measured) and can name the exact problem — location, expectation, and what's wrong — while a screenshot alone, at over ten times the token cost, still can't tell Claude what you actually care about.
This is the actual lever, and it costs nothing: describe more, screenshot less. When a picture is genuinely necessary, for a layout bug, a color mismatch, something words fumble, crop it to the relevant region instead of the full window; a smaller image is fewer tokens under the same formula. A precise sentence ("the save button sits 8px too far right of the card edge, and should align with the title above it") often replaces a screenshot entirely, at a fraction of the cost.
Some bugs really are easier to point at than describe, and that's a real tradeoff. It's the reason tools exist that turn a screen-and-voice walkthrough into a compressed, text-first bundle before it ever reaches Claude. Walkie is one option built for exactly that; it's a purchase decision, not a requirement, and everything above works with nothing but habits you already have.
What can you control today?
Several changes cost nothing and take effect immediately: crop screenshots to the relevant region, describe instead of screenshotting when words suffice, start a fresh conversation for a new task instead of extending an old one, and avoid re-pasting context Claude already has. None require buying a tool or changing how you code — only how you hand Claude your feedback.
- Crop before you paste. A smaller image is fewer tokens under the same patch formula: cut the screenshot down to the region that matters instead of the full window.
- Describe first, screenshot second. Name the location and the expected-versus-actual behavior in a sentence before reaching for the capture tool; many bugs don't need a picture at all.
- Start new threads for new tasks. A fresh conversation doesn't carry forward every prior screenshot and file read. An old thread does, on every single turn, whether you're using that history or not.
- Don't re-paste what Claude already has. If a file or screenshot is already in the conversation, referencing it by name costs far less than pasting it again.
- Batch your visual feedback. One well-described review covering several issues costs less than several separate screenshot-heavy messages saying the same things piecemeal.
None of this requires spending anything, and none of it changes what you're building. It changes only how much of your limit gets spent showing Claude what you mean, instead of building it. If you reach the point where recording and narrating reviews is a daily habit and manual cropping starts to feel like a chore, that's the point where a compression tool starts paying for itself rather than being a nice-to-have. See 11 ways to cut Claude Code token costs for the fuller, ranked list. This article covered the honest mechanics; that one covers what to do about every part of them, not just the visual half.