Check the route before the rate
As of this October 4 research check, OpenAI's official documentation states that Sora 2 and the Videos API shut down on September 24, with no one-to-one replacement API. Replica's indexed official homepage carries a farewell notice; direct retrieval was unavailable, so current pricing cannot be confirmed. Neither belongs in a dependable new production plan on the strength of an old price page. This is a practical procurement lesson: verify that the intended service route exists before comparing what a generation used to cost.
Source: OpenAI deprecations: Sora 2 video generation models and Videos API · 2026-03-24
Source: Replica Studios Farewell
Separate plan price from rights
ElevenLabs lists Starter at a regular $6 per month with 30,000 credits. Its help page permits commercial use of qualifying paid-plan output subject to rights and applicable terms, while excluding free-plan and Beta Services output from commercial use. A production budget should therefore name the plan and feature, not just the brand. Keep the script, voice authorization where relevant, generation date, and applicable terms with the project. A subscription receipt does not resolve every right associated with a voice, likeness, script, music track, or supplied reference.
Source: ElevenLabs Pricing
Source: Can I publish the content I generate on the platform?
Compare the same deliverable
Replicate prices many models by runtime and others by output units; its documentation says model licenses vary. Google's current Veo 3.1 API table lists distinct preview variants, with Standard video and audio at $0.40 per second for 720p and 1080p. These are different billing structures, not directly comparable quotations for a finished advertisement. Specify resolution, length, audio, acceptable consistency, and intended use before comparing suppliers. A rate becomes meaningful only when the team understands how many attempts and how much editing that deliverable is likely to require.
Source: Replicate Pricing
Count the rejected attempts
For an illustrative eight-second shot, a posted rate of $0.40 per second implies $3.20 for one successfully billed generation at that rate. Five billed attempts imply $16 before editing, narration, music, or tax. Those figures are arithmetic examples, not a quote or a prediction of quality. Track the acceptance rate during a small approved pilot. Distinguish shots rejected for factual or continuity problems from shots rejected because the brief changed. This reveals whether the next saving should come from better prompting, stronger art direction, a simpler scene, or another production method.
Human review belongs in the unit cost
Google's API terms say it does not claim ownership of generated content and leave responsibility for use with the user. That supports a careful review process rather than a promise that every output is automatically cleared. Review product appearance, claims, captions, audio, and any real person's likeness before release. Keep the approved version and its source assets. An attractive frame can still show the wrong product detail or imply a result the business cannot support. The finished-minute budget should include the human work needed to catch those issues and produce a usable final asset.
Source: Gemini API Additional Terms of Service · 2026-03-23
Design for continuity and reuse
Keep storyboards, approved scripts, visual references, editable project files, and final masters outside a single supplier's interface. Build a short reusable format that can survive a provider change: one message, a few clear shots, a controlled narrator, and an explicit approval step. Budget the first edition as a learning exercise and later editions against observed acceptance and editing effort. AI media can make experimentation more accessible, but commercial value comes from a finished piece that is accurate, licensed for its intended use, and straightforward to update when the business changes.