Console, Keys, and What It Costs

Keys, and the domain question

API keys come from https://console.ltx.io. Every request carries them the same way:

Authorization: Bearer YOUR_API_KEY

The base URL is `https://api.ltx.io`.

Key hygiene is unglamorous and worth thirty seconds. Keys go in environment variables or a secret manager, never in a repo, never in a ComfyUI workflow JSON you share, never in a screenshot in a Slack thread. Billing is prepaid against a credit balance, so a leaked key is not an abstract security problem — it is somebody else spending your balance until it hits zero and every generation in your pipeline starts failing.

The price list

Pricing is per second of output.

Notice the structure before the numbers. Each resolution step doubles the rate. Pro carries a consistent premium over Fast at the same resolution. And the post-production endpoints are priced above generation — reframe at 1080p output costs $0.20/sec against $0.08/sec to have generated the thing in the first place.

Worked example: a 96-second piece

Twelve shots, eight seconds each, 1080p, ltx-2-3-pro, three takes per shot because first takes are drafts.

Generation. 12 shots × 8s = 96 seconds of finished runtime. Three takes means 288 seconds actually generated.

288 s × $0.08/s = $23.04

The same work on Fast at 1080p would be 288 × 0.06 = $17.28 — a $5.76 saving that costs you access to retake, extend, and reframe on those shots. That is rarely a good trade for finals.

Delivery. Now reframe the 96-second master into three additional aspect ratios at 1080p output:

96 s × $0.20/s        = $19.20 per ratio
$19.20 × 3 ratios     = $57.60

Total: $80.64. Look at the split. Generation with three takes each: $23.04. Delivery: $57.60. Delivery costs two and a half times what generation cost.

That is the most useful number in this lesson. The instinct is to budget for generation and treat finishing as rounding error. The price list says the opposite. If your brief calls for four aspect ratios, decide that at the start — because the honest alternative, framing your original shots so a single master reads acceptably in more than one crop, is a creative decision made before you generate, not a cost optimisation applied afterwards.

Two more rates worth pre-computing. HDR on 1080p-or-under input is $0.20/sec, and the input is capped at 181 frames — roughly 7.5 seconds at 24 fps — so about $1.50 per shot, and only for shots that genuinely need an EXR pass. Audio-to-video and retake both sit at $0.10/sec at 1080p, which makes a targeted retake cheap in absolute terms and easy to do fifty times without noticing.

Auto Top-Up, and how it fails

Because the balance is prepaid, hitting zero is an outage: 402 `insufficient_funds_error`, which is not retryable. Auto Top-Up exists to prevent that.

The real ranges: threshold $10 to $10,000, top-up amount $5 to $2,500. It requires the owner role and a payment method on file. And the part to actually remember: it auto-disables after 3 consecutive payment failures.

That last line is the failure mode. An expired company card produces three quiet failures, Auto Top-Up switches itself off, and the next thing anyone learns is that the balance is zero mid-render. Set your threshold high enough that the gap between "top-up failed" and "generations fail" is measured in days, not minutes. For the $80 project above, a threshold of $50 with a $100 top-up gives you real headroom; a threshold of $10 gives you about two minutes of warning.

Concurrency is a planning constraint

The default concurrency limit is 2 simultaneous generations.

Return to the worked example: 12 shots × 3 takes = 36 generations. At two at a time, that is 18 sequential waves no matter how much money you have or how many threads you spawn. You cannot buy your way out of wall-clock time by fanning out. Treat 2 as a scheduling input when you plan a delivery date, and build your client around a semaphore rather than optimism.