--- title: "Old model usage" description: "antiburn flags sessions that used a deprecated model after its reviewed replacement became available." url: "https://antiburn.ai/docs/findings/old-model-usage/" --- ## What it detects antiburn maintains a curated list of deprecated models, each mapped to the reviewed replacement that succeeds it and the date that replacement became available - checked against the vendor's own announcement, not a guess. A finding fires when a session used a listed model after its replacement's availability date. Using an old model before its replacement shipped is not a finding. The list currently covers, for example, the Claude Opus 4.x line (replaced by Opus 5) and Sonnet 4.x line (replaced by Sonnet 5), and OpenAI's GPT-5.4, GPT-5.4 mini, and GPT-5.5 (replaced by GPT-5.6's Terra, Luna, and Sol tiers respectively). This is a maintained reference list, not something you configure, and it grows as vendors ship new replacements. ## Why it burns quota A model's replacement is added to this list because it gives better output at the same or lower cost, per the vendor's own positioning. Staying on the older model buys nothing and usually costs the same or more for weaker results. ## How to fix it Switch to the current replacement the next time you start a session. If your agent lets you set a default model in its configuration, update that default so future sessions don't repeat the finding - otherwise you'll need to select the newer model manually each time. ## When it shows "not assessed" antiburn needs to know which model a session used and when, so a session shows not assessed if model identity isn't recorded at all - true only for agents antiburn doesn't have a dedicated adapter for. For agents that do report model identity, a session using a deprecated model still comes back not assessed rather than a finding or clean if antiburn can't place the usage in time - no usable timestamp was recorded for that turn, so antiburn can't tell whether it happened before or after the replacement shipped. Partial model evidence also blocks a clean result even when no finding was observed, since a missing record could be hiding deprecated usage antiburn never saw. See [hygiene score and findings](/docs/concepts/hygiene-score-and-findings/) for how this fits into your overall score.
Old model usage
antiburn flags sessions that used a deprecated model after its reviewed replacement became available.
What it detects
antiburn maintains a curated list of deprecated models, each mapped to the reviewed replacement that succeeds it and the date that replacement became available - checked against the vendor’s own announcement, not a guess. A finding fires when a session used a listed model after its replacement’s availability date. Using an old model before its replacement shipped is not a finding.
The list currently covers, for example, the Claude Opus 4.x line (replaced by Opus 5) and Sonnet 4.x line (replaced by Sonnet 5), and OpenAI’s GPT-5.4, GPT-5.4 mini, and GPT-5.5 (replaced by GPT-5.6’s Terra, Luna, and Sol tiers respectively). This is a maintained reference list, not something you configure, and it grows as vendors ship new replacements.
Why it burns quota
A model’s replacement is added to this list because it gives better output at the same or lower cost, per the vendor’s own positioning. Staying on the older model buys nothing and usually costs the same or more for weaker results.
How to fix it
Switch to the current replacement the next time you start a session. If your agent lets you set a default model in its configuration, update that default so future sessions don’t repeat the finding - otherwise you’ll need to select the newer model manually each time.
When it shows “not assessed”
antiburn needs to know which model a session used and when, so a session shows not assessed if model identity isn’t recorded at all - true only for agents antiburn doesn’t have a dedicated adapter for.
For agents that do report model identity, a session using a deprecated model still comes back not assessed rather than a finding or clean if antiburn can’t place the usage in time - no usable timestamp was recorded for that turn, so antiburn can’t tell whether it happened before or after the replacement shipped. Partial model evidence also blocks a clean result even when no finding was observed, since a missing record could be hiding deprecated usage antiburn never saw.
See hygiene score and findings for how this fits into your overall score.