AI-Generated Release Notes and Changelogs: Cost per Release Token Math

The Real Cost of a Release Note

Every product ships releases. Every release needs a set of user-facing notes and a technical changelog. Doing this manually takes a person 20–45 minutes. Doing it with AI takes seconds and costs pennies to a few dollars — but the “few dollars” ceiling matters when you release 4× a day, or run continuous deployment.

The cost split between two strategies is bigger than most people assume.

Strategy A: Conventional Commits Only

Feed the model just the commit list from the release, filtered to Conventional Commits format:

feat(auth): add SSO provider selection
fix(billing): correct currency conversion for JPY
chore(deps): bump next from 15.4.0 to 16.0.1
docs(api): document new /v2/subscriptions endpoint

Token profile per release: 500–3K input, 300–1K output. Cost at Claude Sonnet 5 promo pricing: $0.003–0.016 per release. Basically free.

Tradeoff: quality is bounded by commit message quality. If your team writes “fix stuff” commit messages, you get “fix stuff” release notes. If your team disciplines around descriptive Conventional Commits, output rivals a human writer.

Strategy B: Full Diff

Feed the model the raw diff of the release plus commit messages. The model figures out what changed independent of commit quality.

Token profile per release: 50K–500K input, 1K–3K output. Cost per release: $0.10–1.10. Substantially more expensive, but the output is dramatically better when commit messages are noisy.

This is the strategy to use when you inherit a codebase with poor commit hygiene, or when your team just doesn't enforce Conventional Commits.

Strategy C: Hybrid — Commits + Selected Diffs

Feed all commit messages, plus diffs only for the commits that don't follow Conventional Commits format. This lets the model rely on structured input where available and fall back to diff reading where necessary.

Token profile per release: 5K–20K input, 500–1.5K output. Cost per release: $0.015–0.055. Close to Strategy A on cost, close to Strategy B on quality — but requires more setup logic.

Monthly Cost Comparison

For a team shipping 20 releases per month at moderate size (30 commits per release, 3K LOC diff):

Every AI strategy pays back against manual work within a single release. The choice between AI strategies comes down to commit hygiene and how strongly you care about the note quality.

Semantic-Release Integration

If your team already uses semantic-release or release-please, both tools now integrate optional LLM enhancement steps. The typical setup:

1. Semantic-release generates the structural changelog from Conventional Commits.
2. An LLM post-processor rewrites bullet points into user-facing prose.
3. The polished output goes into GitHub release notes / product changelog / marketing email.

Post-processing adds roughly $0.02–0.05 per release on top of Strategy A costs. Total monthly cost for 20 releases: about $1. Worth every cent for teams that publish release notes as marketing surfaces.

Multi-Audience Notes

Some releases need three different note formats: technical changelog, user-facing highlights, marketing announcement. Generating three variants in one pass costs slightly more than one variant but much less than three sequential calls, because the input tokens are only paid once with prompt caching.

Structure the prompt to output all three variants as separate sections. On Claude Sonnet 5 promo pricing, this triples output tokens (roughly $0.03–0.08 extra per release) but keeps input cost flat. Total for 3-audience release notes: ~$0.05–0.15 per release using hybrid strategy.

Cost Traps to Avoid

• Passing the full diff on every release when Conventional Commits would suffice. Easy to slip into if you copy-paste an example prompt.
• Regenerating notes on every commit rather than on release cut. Only re-run at release boundaries.
• Using a frontier model where a mid-tier model is fine. GPT-4o or Sonnet 4.6 handles this task well; Opus 4.8 doubles cost for no measurable quality gain.
• Skipping prompt caching. Even simple caching of the system prompt saves 50–70% on repeated release-note generations.