An LTX API migration is more than replacing one model ID. The official changelog shows a sequence: LTX-2 deprecation was announced on July 2, LTX 2.3 gained a 720p tier on August 2, LTX 2.5 arrived on August 11, and LTX-2 removal was recorded on August 16. A safe change therefore needs a dated dependency audit, matched visual tests, and a rollback path.

Retest LTX 2.3 with Your Own Assets

This plan turns the official LTX API changelog into a reproducible creator retest. It keeps vendor facts separate from your observations, preserves the old evidence bundle, and compares supported paths with the same prompts and owned inputs. It does not assume a newer model is automatically better for every production case.

About this guide: APOB AI prepared it from the first-party sources listed below and rechecked them on August 26, 2026. No performance, quality, latency, or cost result is claimed without a team running the suite and recording its own evidence.

Map the LTX API changes

Begin with a change map, not a code edit. For every entry, record the changelog date, affected model or endpoint, stated behavior, operational consequence, owner, and evidence link. The difference between an announcement date and the date a change is recorded matters during incident review.

July 2 deprecation

The July 2 notice said ltx-2-fast and ltx-2-pro would be deprecated on July 15, when requests would be served by LTX 2.3 without a price change, and removed on August 15. Treat those dates as the published plan. The later changelog entry records removal on August 16 and says requests specifying the retired IDs return an error.

Capture both entries in the migration record. Do not rewrite the planned date after the fact or use it as proof of the actual cutover time in your account.

August 2 resolution tier

On August 2, the changelog added 720p output for LTX 2.3 text-to-video and image-to-video: 1280x720 landscape and 720x1280 portrait. If an old preset assumed another output size, downstream crops, caption safe areas, upload validation, and cost estimates may change even when the prompt does not.

Add explicit resolution assertions to the retest rather than accepting any playable file.

August 15 removal

August 15 was the announced removal deadline; the current changelog records the removal on August 16. Its operational statement is unambiguous: the retired LTX-2 model IDs are unavailable and requests that specify them return an error.

Search configuration, environment files, job payloads, presets, saved templates, retry handlers, and documentation for both retired IDs. A wrapper may hide the string from application code while still sending it.

LTX-2.5 availability

The August 11 entry introduced ltx-2-5-fast and ltx-2-5-pro for text-to-video, image-to-video, and audio-to-video. This makes LTX 2.5 a current migration option; it does not prove it is the right replacement for a particular LTX 2.3 API workflow.

Keep the model families separate in tickets, filenames, and score sheets. “Current” is an access fact. “Preferred” must come from the matched test.

Inventory legacy dependencies

Freeze one inventory before changing production. Include direct calls, SDK wrappers, no-code automations, scheduled jobs, prompt templates, asset rules, storage, monitoring, and client-facing documentation. Assign an owner and stop/go decision to every row.

Endpoint inventory

Use redacted request and response examples. Preserve request IDs and timestamps, but remove keys, personal data, and private asset URLs from shared evidence.

Add a “how found” field beside every dependency: repository search, dashboard export, scheduler inspection, owner interview, or traffic log. That makes blind spots visible. Require the owner to confirm both positive findings and “not used” decisions; an empty cell is not evidence that a path is safe.

Preset and ratio audit

List duration, resolution, orientation, frame or camera controls, prompt enhancement, seed behavior where supported, and image constraints. Then trace assumptions into caption layouts, thumbnails, ad specifications, and delivery folders.

The current APOB AI Video Generator lists LTX 2.3 among its supported model options and can provide a creator-side comparison, but it is not evidence that a private API integration shares identical controls. Record the interface tested.

Billing and quality assumptions

Separate inputs from outcomes. Inputs include the selected model, resolution, duration, attempts, and any current price shown in the account. Outcomes include accepted clips, manual fixes, latency observed by your test, and final cost calculated from your own invoice or usage log.

Do not copy a temporary or regional number into a permanent forecast. Date-stamp every value and identify who can approve a budget change.

Build a matched regression suite

The suite should be small enough to run after every material API change and diverse enough to catch production breakage. Use only owned or licensed inputs, version every file, and preserve untouched exports.

Fixed prompts and inputs

Choose one product still, one fictional adult character, and one environment image. Lock prompt text, negative constraints, filenames, durations, aspect ratios, and any supported seed. Save the exact JSON payload for each path.

Do not “help” one candidate with a rewritten prompt unless the same revision becomes a new, separately labeled test case.

720p and 1080p cases

Run one 720p case because that tier was explicitly added for LTX 2.3. Run a matched 1080p case only where the current path supports it. Verify pixel dimensions with file metadata, not the export label, and inspect product geometry, text, face continuity, motion, flicker, and compression.

Portrait and landscape cases

Use the same narrative beat in 9:16 and 16:9. Keep the subject action and intended crop fixed. Check whether hands, product, face, caption-safe area, and important background elements survive both frames.

The broader APOB AI Video Generator can provide a second creator-facing path for the same owned inputs. Save its settings and output separately; do not present cross-interface results as an API benchmark.

Acceptance rubric

Score identity, prompt adherence, product fidelity, temporal stability, motion, framing, audio when applicable, technical file validity, retries, and manual repair. Use a 0–2 scale: 0 fails, 1 needs a documented fix, 2 passes as delivered. Define any hard stop—wrong product, unusable file, rights issue, or unsupported claim—before viewing the clips.

Permit two documented attempts per case and path. A third attempt begins a new labeled test with a written reason; it cannot replace the earlier failures in the score sheet.

Require screenshots, file metadata, request logs, and reviewer notes for every case, including failures.

Compare available migration paths

Compare only paths that were accessible on the test date. A supported ID in documentation, an APOB creator interface, and a production API integration are different surfaces and must be labeled accordingly.

Compatibility

Check payload schema, supported generation modes, input types, duration, aspect ratio, resolution, response parsing, and downstream delivery. Mark each item supported, changed, unsupported, or not tested. Compatibility is a prerequisite, not a quality score.

Quality and latency

Review outputs blind where practical. Measure submit-to-ready time from your logs, but report the sample count, network context, date, and failures. A single fast result is an observation, not a durable service claim.

Cost inputs

Record model, resolution, duration, successful jobs, failed jobs, retries, and the account’s dated usage result. Calculate cost per accepted clip as a local operational metric. Do not turn it into a universal vendor price.

Fallback path

Define one supported API choice, one creator-side fallback, and one manual delivery option. The fallback must preserve the same approved brief and ownership rules. Test it before a deadline, not during the first outage.

Roll out with a rollback plan

Production starts only after the inventory and regression evidence have named sign-offs. Keep the previous deployment artifact, configuration, and accepted output bundle until the new path meets the monitoring window.

Sandbox validation

Run every case with non-production credentials and destinations. Confirm errors are visible, logs are redacted, files reach the correct folder, and no test asset can publish to a live account.

Canary batch

Route a small, explicitly approved batch through the new path. Cap jobs and spend, label the outputs, and compare acceptance rate, retries, manual repair, and delivery errors with the frozen baseline.

Monitoring

Watch request errors, job completion, download failures, invalid files, retry rate, accepted-output rate, and cost per accepted clip. Each alert needs an owner, threshold, response time, and evidence link.

Rollback and owner sign-off

Rollback means switching to a tested supported path or pausing generation—not sending traffic back to removed model IDs. Record the decision owner, trigger, configuration change, client communication, and recovery verification.

Run a short rollback drill in the sandbox. Time how long it takes to stop new jobs, switch the supported fallback, verify one known-good file, and notify the owner. Attach the drill record to the release ticket so recovery is an exercised procedure rather than an optimistic paragraph.

One-page checklist: archive the LTX changelog evidence; find retired IDs; freeze presets and downstream assumptions; run matched 720p, 1080p where supported, portrait, and landscape cases; score technical and creative results; calculate from dated usage logs; test the fallback; canary with caps; monitor; obtain API, creative, operations, and budget sign-off. Evidence as of August 26, 2026; recheck by September 15.

Sources

• LTX API Changelog — Lightricks

• AI Video Generator — APOB AI