Answer in brief
An AI media comparison now starts with availability. OpenAI lists Sora’s API removal on 24 September; current image and audio-video preference tables answer different production questions.
September retirement changes the shortlist
For production teams, an AI media comparison now begins with whether the endpoint exists. OpenAI’s official calendar lists removal of its Videos API and Sora 2 variants on 24 September 2026. At the 1 October cutoff, a guide that treats Sora as a normal first-party production choice needs that availability fact checked before any quality discussion.
The change is practical for an ongoing campaign. A memorable old demonstration cannot promise a working endpoint, supported workflow or repeatable revision. Build a shortlist from current provider documentation, then attach quality evidence to each exact variant. We did not generate evaluation clips or run an image benchmark for this article.
A current video test must specify its sound
The video table uses Artificial Analysis’s audio-enabled AA-Video-T2V v2.0 board observed on 1 October. It records human preference within that task and retains confidence intervals. The producer’s methodology specifies generation defaults and a distinct silent board. Exact clip-generation and voting dates were not disclosed in the viewed snapshot, so they remain unknown.
Sound changes a brief: dialogue, ambience and synchronization can determine whether a shot works. A silent-video score cannot answer that question. Kling’s value of 1000 is the reference anchor for this scale, rather than a measured percentage. None of these numbers measures factual accuracy or guarantees that an entire sequence maintains continuity.
| Video model / test variant | Elo | 95% interval | Samples | Provider availability |
|---|---|---|---|---|
| Gemini Omni Flash 1.1 | 1117 | 1107–1127 | 6,974 | Gemini API, paid tier |
| Kling 3.0 1080p (Pro) | 1000 | Fixed reference anchor | 4,453 | Native audio mode documented |
| Veo 3.1 | 966 | 955–977 | 2,851 | Gemini API; generation with audio |
| Sora 2 / Sora 2 Pro | Unknown on this current board | Unknown | Unknown | Provider removal date: 2026-09-24 |
Provider features can decide a close practical choice
Google’s current video documentation recommends Gemini Omni Flash as the default and identifies Veo 3.1 for specific controls such as scene extension and frame direction. Kling’s official guide documents native audio and multiple-shot production. Those are provider-described capabilities; the preference table supplies independent evidence about a narrower generation task.
A studio might need a particular editing operation more than the model most frequently preferred in anonymous comparisons. Google’s Veo card was updated on 13 January 2026, and its product-page comparison panels carry October 2025 dates. Historical provider comparisons should therefore keep their original dates rather than becoming claims about today’s entire market.
Still images require a separate comparison
OpenAI’s 8 September Images 2.5 card documents Sunburst and Flare, while Google’s image API describes Nano Banana variants. The second table uses Artificial Analysis’s text-to-image board with named quality settings. Sunburst and Flare confidence intervals overlap, so their visible order should not be presented as a certain distinction.
An image generated from text and an edit to an approved reference are different deliverables. A higher text-to-image preference score does not establish whether a model preserves a product’s geometry during a local revision. Image Elo and video Elo also have separate anchors. Combining them into an invented all-purpose winner would discard the task each measurement represents.
| Image model / quality variant | Elo | 95% interval | Samples |
|---|---|---|---|
| GPT Image 2.5 Sunburst (max) | 1197 | 1188–1206 | 14,023 |
| GPT Image 2.5 Flare (max) | 1190 | 1181–1199 | 13,393 |
| Nano Banana 2 / Gemini 3.1 Flash Image | 1125 | 1117–1133 | 18,454 |
Choose a workflow, then budget accepted outputs
A useful creative trial would specify an approved reference, required language, duration, camera direction and revision constraints. Evaluate continuity, editability and sound on those requirements. Track rejected generations as well as accepted material. This is a proposed decision process; it is not a claim that our newsroom tested the listed models.
The September availability change makes that discipline timely. Keep the provider notice, model variant, modality and observation date together. The result is a shortlist a producer can actually act on, followed by a budget for accepted images or shots, rather than a promise inferred from an old launch demonstration.
Questions and answers
Is Sora still an available first-party API option?
OpenAI’s official deprecation calendar lists removal of the Videos API, Sora 2 and Sora 2 Pro on 24 September 2026. An older comparison does not override that provider notice.
Does a higher Elo guarantee a usable production shot?
No. Elo measures relative human preference under a defined test. It does not establish continuity, approved likenesses, precise spoken language or the cost of retries for your brief.
Can image and video Elo be combined into one score?
No. The boards use separate tasks, samples and reference anchors. Keep still-image generation, image editing and audio-video results attached to their own evaluation methods.