Skip to content
xplainio
All films

Developer tools

Why your AI coding agent burns tokens, and how to stop it

Every message you send to a coding agent carries the whole conversation with it. This film follows one session from the first message, then looks at antiburn, an open-source tool that checks every session for the common causes of token burn.

Runtime
3:37
Claims checked
6
Sources
8
  1. 0:00The burn
  2. 0:11Your first message
  3. 0:33Sent to the model
  4. 0:51Cache hit
  5. 1:13Taking a break: cache miss
  6. 1:29The context grows
  7. 1:37/clear
  8. 1:49/compact
  9. 2:10The checks
  10. 3:03antiburn
Runtime
3:37
Tokens sent with the first message
44k
Session checks explained
6
Chapters on YouTube
10

The story in brief

The cost is not the message you type. It is everything sent with it.

One session
  1. 1First message: about 44k tokens
  2. 2The provider keeps a copy
  3. 3Cache hit: a fraction of the cost
  4. 4A break: cache miss
  5. 5The context nearly fills
  6. 6/clear or /compact
  7. 7antiburn checks the session

Context through the example session

  • First message43.7k
  • After the break59.0k
  • After an hour160.3k
  • After /compact47.7k
Source: Thousands of tokens, from the film's modelled session. Not a real invoice.

What we verified

Every claim the film makes, where it comes from, and how far the source lets it go.

Claim in the filmSourceHow far it goes
Reusing a cached prefix costs less than sending and writing it againAnthropic prompt caching docsAPI pricing, not a subscription-quota formula
/compact summarises the conversation and the summary replaces itClaude Code docsCompaction frees context; it does not refill quota
Estimated API cost and provider-reported plan usage are separate measuresantiburn docs: tokens, cost and quotaNo plan percentage is invented from token traffic
Repeated cache rebuilding is a cause of burn that antiburn checks forantiburn docs: excess cache rehydrationNot every cache write is a rebuild or avoidable
antiburn is free, open source and analyses sessions locallyantiburn FAQ and READMEPrivacy is quoted from the README as of 14 Oct 2026; zero network activity is not claimed
The first request is 7 tokensOpenAI o200k tokenizerClaude's own count may be 8 to 10

What we left out, and why

  • Prices and minutes on screen

    Cache lifetimes differ by provider, so any single figure would be wrong for most viewers.

  • Savings promises

    The film shows what to inspect before paying for more usage. It does not promise savings, more quota or automatic optimisation.

  • A $3,603 total from a supplied screenshot

    A screenshot is not an invoice, and that session never hit its limit.

  • antiburn's cost rows and efficiency section

    Cut to keep one story: one session, its context and its checks.

How it was made

Research
The app, its docs and its sourceantiburn's app was captured from public screenshots on 8 October 2026; its definitions come from its docs and pinned source.
Story
One session, one storyAfter five earlier cuts, the sixth was rebuilt scene by scene around one story: cache hit, cache miss, nearly full, compact, the checks. Then it was shortened by about half.
Voice
One calm narratorGoogle Gemini text-to-speech, voice Charon. Sound effects from ElevenLabs.
Build
The real interface, rebuilt in codeantiburn's views are rebuilt layer by layer, and every figure on screen comes from one modelled session, so a chip and the app always agree.
Review
Scene-by-scene reviewsEach scene was reviewed on its own. Scenes 1 and 2 follow the LGTM music video, scene 4 techniques from another explainer.

Sources

More films

See all 4 films
Work with us

Have something hard to explain? We make the film.

Every film here started as a dense source: a 91-page report, an app’s documentation, a funding dataset. We turn sources like yours into short films that make them clear.

Talk about a film