fix(ai): read long-context thresholds from models.dev tiers by PavleProd · Pull Request #30141 · twentyhq/twenty

PavleProd · 2026-11-10T22:49:38Z

Problem

The catalog sync prices long context from models.dev's context_over_200k and gives every model a 200K threshold (LONG_CONTEXT_THRESHOLD_TOKENS). models.dev deprecated that field on 2026-08-23 (anomalyco/models.dev#5335). The real threshold now comes from cost.tiers[].tier.size, and it isn't always 200K:

13 OpenAI models (gpt-6-luna, gpt-6-sol, gpt-6.1-sol, gpt-5.6, gpt-5.5, gpt-5.4 and others) switch to long-context rates above 272K. computeCostBreakdown switches them at 200K, so a request of 200K–272K is debited at double the input rate.
Claude Haiku 5.5 switches above 100K, to 5× its base rates. models.dev publishes that tier only in tiers, so the catalog has no longContextCost for it, and requests over 100K are debited at base rates.

Debited by computeCostBreakdown today, against the provider's price (90% of input cache-read, 1K output):

Model	Input	Provider price	Debited today
gpt-6-luna	240K	$0.0051	$0.0099
gpt-6.1-sol	240K	$0.0796	$0.1542
claude-haiku-5-5	150K	$0.0168	$0.0034

Fix

extractCost takes the lowest context tier from cost.tiers and uses its size as thresholdTokens.
It falls back to context_over_200k at 200K when a model has no tiers.
ModelsDevModel gains the tiers field.

Built against today's models.dev payload, the only catalog changes are:

the 13 OpenAI thresholds (200K → 272K);
Haiku 5.5 gaining its 100K tier;
gpt-6.1-terra getting its 272K tier.

Rates and the Google and xAI entries are unchanged. ai-models.json and ai-providers.json aren't regenerated here; the next daily sync rewrites them. The regenerated catalog can be added to this PR if preferred.

Test plan

build-catalog.spec.ts: two new cases, a 272K tier quoted in both fields and a 100K tier quoted only in tiers. Both fail on main (200000 instead of 272000; longContextCost undefined) and pass with the fix.
All suites under scripts/ai-catalog-sync and metadata-modules/ai pass, except build-front-component-shared-dependencies, which needs a local twenty-ui build.
oxlint and oxfmt are clean on the changed files. tsgo reports no errors in them.

Written with Claude Code (Claude Opus 5.5).

twenty-ci-bot-public · 2026-11-10T22:50:00Z

👋 Thanks for contributing to Twenty! We're excited to have you on board.

Your PR has been set to draft while you work on it. Once you're done, mark it as Ready for review and our automated checks will run.

By submitting your Pull Request, you acknowledge that you agree with the terms of our Contributor License Agreement.

greptile-apps · 2026-11-10T22:52:11Z

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

Reads long-context prices and thresholds from models.dev context tiers instead of assuming 200K tokens.

AI model prices switch at each model’s published long-context threshold.

_{Reviews (3) · Last reviewed commit: "fix(ai): read long-context thresholds fr..." · Reviewed by Greptile}

The catalog sync gave every long-context model a 200K threshold taken from models.dev's deprecated context_over_200k field. OpenAI models switch at 272K and Claude Haiku 5.5 at 100K, which models.dev publishes only in cost.tiers. Take the threshold and rates from the lowest context tier, falling back to context_over_200k at 200K.

twenty-ci-bot-public · 2026-11-10T22:56:55Z

🤖 PR Review

Check	Result
🔍 Build safety	✅ passed
🧭 Triage	✅ passed
📐 Quality	⏭️ skipped — trusted-only
🛡️ Security	⏭️ skipped — trusted-only
🚦 Auto-approve	⏭️ skipped — trusted-only
🔐 Permissions	⏭️ skipped: no endpoint permission change

🔍 Automated Pre-Review

✅ No issues detected - This PR is ready for human review.

🧭 External PR Triage Review

✅ Looks good — a maintainer can pick this up.

Checks

⏳ CI: pending
🔗 Linked issue: Fixes #30144

Detailed findings (duplicate candidates, standards notes, summary) are in the workflow run logs.

View details

Automated pre-review — human approval still required.

etiennejouan

LGTM. Hi @PavleProd congrats for your first contribution to twenty repo !

PavleProd · 2026-11-13T09:38:51Z

Thanks @etiennejouan, glad it was useful!!!!! :D

PavleProd · 2026-11-13T11:18:37Z

Hey @etiennejouan, the server-integration-test failure is in application manifest sync (DEFERRED_WORKSPACE_MIGRATION_ACTIONS_IN_PROGRESS), which this PR doesn’t touch. Could you re-run the failed job when you get a chance?

…0371) ## Summary The AI catalog sync has failed every day since 2026-10-23 at "Create pull request" with `GitHub Actions is not permitted to create or approve pull requests`. That repo setting is off on purpose (twentyhq#29028), and i18n moved to the Twenty PR app, but this workflow kept `GITHUB_TOKEN`. The last catalog sync merged was twentyhq#28593, so `ai-models.json` and `ai-providers.json` are three weeks stale and catalog builder fixes like twentyhq#30141 can't reach billing. ```diff sync-catalog # GITHUB_TOKEN, contents: write run catalog sync compose PR body - peter-evans/create-pull-request # pushes, then fails to open the PR - mint dispatcher token, trigger twenty-infra automerge + force-push chore/ai-catalog-sync +open-pull-request # permissions: {} + open-pull-request.yaml # Twenty PR app opens the PR, triggers automerge ``` The reusable `i18n-open-pull-request.yaml` becomes `open-pull-request.yaml`, since it now serves more than i18n. It gets an optional `body` input, defaulting to its current text, and refreshes the body of an already open PR, so the catalog PR keeps a current coverage report. Its CI-trigger commit is now `chore: run CI on the latest push`. The i18n callers only change the path. ## Evidence - **Before:** run 43375772095 (2026-11-13) pushes `chore/ai-catalog-sync`, then fails with `GitHub Actions is not permitted to create or approve pull requests.` **After:** oxfmt and YAML parsing pass. The push step, replayed against a bare repo holding a stale `chore/ai-catalog-sync`, force-updates it with a single commit touching only the catalog folder. The app step only runs on GitHub, so the real check is the first run after merge: a PR from `twenty-pr` with CI running. ## Merge Danger **Door:** two-way Reverting brings back the current, failing workflow. **Blast Radius:** small - Catalog PRs now come from the Twenty PR app, so they get CI and twenty-infra's automerge turns on auto-merge. Main still requires an approval and the review bot only auto-approves translations, so a human still approves each catalog PR. - The first PR carries three weeks of catalog drift. - The five i18n workflows call the renamed file and behave the same, apart from the CI-trigger commit message. Nothing outside this repo references the old name or message. - The `ai, automated` labels are dropped. Nothing reads them.

twenty-eng-sync Bot added the -PR: draft label Nov 10, 2026

PavleProd force-pushed the fix/ai-catalog-context-tier-thresholds branch from 32d032b to 826787d Compare November 10, 2026 22:55

twenty-ci-bot-public Bot added the external-needs-triage label Nov 10, 2026

PavleProd mentioned this pull request Nov 10, 2026

AI credits use a 200K long-context threshold for every model (OpenAI is 272K, Claude Haiku 5.5 is 100K) #30144

Closed

PavleProd marked this pull request as draft November 10, 2026 23:19

PavleProd marked this pull request as ready for review November 10, 2026 23:19

twenty-ci-bot-public Bot added external-ready-to-be-reviewed and removed external-needs-triage labels Nov 10, 2026

twenty-ci-bot-public Bot mentioned this pull request Nov 11, 2026

fix(ai): dynamically read long-context threshold from models.dev tiers (#30144) #30151

Closed

etiennejouan self-assigned this Nov 13, 2026

chore(ai): drop redundant legacy threshold comment

e3d3342

etiennejouan approved these changes Nov 13, 2026

View reviewed changes

chore(ai): drop context tier comment

cbd3c05

etiennejouan enabled auto-merge November 13, 2026 09:32

etiennejouan mentioned this pull request Nov 13, 2026

Open AI catalog sync pull requests with the Twenty PR app #30371

Merged

Merge branch 'main' into fix/ai-catalog-context-tier-thresholds

1451bcb

etiennejouan added this pull request to the merge queue Nov 13, 2026

Merged via the queue into twentyhq:main with commit 6c95bf9 Nov 13, 2026
148 checks passed

Navigation Menu

Repository navigation

fix(ai): read long-context thresholds from models.dev tiers - #30141

fix(ai): read long-context thresholds from models.dev tiers#30141
etiennejouan merged 4 commits into
twentyhq:mainfrom
PavleProd:fix/ai-catalog-context-tier-thresholds

PavleProd commented Nov 10, 2026 •

edited

Loading

Uh oh!

twenty-ci-bot-public Bot commented Nov 10, 2026

Uh oh!

greptile-apps Bot commented Nov 10, 2026 •

edited

Loading

Uh oh!

twenty-ci-bot-public Bot commented Nov 10, 2026 •

edited

Loading

Uh oh!

etiennejouan left a comment •

edited

Loading

Uh oh!

PavleProd commented Nov 13, 2026

Uh oh!

PavleProd commented Nov 13, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

PavleProd commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Test plan

Uh oh!

twenty-ci-bot-public Bot commented Nov 10, 2026

Uh oh!

greptile-apps Bot commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 5/5

Summary

Uh oh!

twenty-ci-bot-public Bot commented Nov 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🤖 PR Review

🔍 Automated Pre-Review

🧭 External PR Triage Review

Checks

Uh oh!

etiennejouan left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

PavleProd commented Nov 13, 2026

Uh oh!

PavleProd commented Nov 13, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

PavleProd commented Nov 10, 2026 •

edited

Loading

greptile-apps Bot commented Nov 10, 2026 •

edited

Loading

twenty-ci-bot-public Bot commented Nov 10, 2026 •

edited

Loading

etiennejouan left a comment •

edited

Loading