Sora 2 (Discontinued)

Overview

This model is being switched off. The Sora web and mobile apps were discontinued on April 26, 2026. The API — sora-2, sora-2-pro and the whole Videos API — shuts down on September 24, 2026, which is eighteen days from this revision (September 6, 2026). OpenAI's deprecations table lists no recommended replacement. This page is now a reference for teams migrating off Sora, not a page to adopt from.

Sora 2 was OpenAI's video generation model, and its last. Released on September 30, 2025, it generates video with synced audio from text or image input, and the API documentation still describes it as "flagship video generation with synced audio," capable of creating "richly detailed, dynamic clips from natural language or images." Everything in that sentence stops being purchasable in less than three weeks.

The wind-down came in two stages. OpenAI notified developers on March 24, 2026; the consumer surfaces went on April 26, 2026; the API goes on September 24, 2026. There is no Sora 3, no announced successor, and nothing in OpenAI's catalog that generates video after that date — gpt-image-2 (April 21, 2026) is still images only, and GPT-6 Astra (September 3, 2026) is a language model.

If you have production traffic on v1/videos, treat September 24 as a hard cutover and pick a replacement from another vendor now — see Pricing & Access for the live alternatives in this catalog.

Sunset Timeline

The models covered by the September 24 shutdown are sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, and sora-2-pro-2025-10-06. The v1/videos endpoint is retired with them.

Capabilities

Everything in this section stops working on September 24, 2026. It is kept as a record of what Sora 2 could do, and as a checklist of what a replacement has to cover.

• Video with synced audio: A single generation produces both tracks together, rather than compositing audio onto silent footage after the fact.
• Text-to-video: Natural language prompts describing the desired clip.
• Image-to-video: Still images as a starting frame or conditioning input.
• Portrait and landscape output: Both orientations at every supported resolution, covering social and cinematic framings.
• Two quality tiers: sora-2 for standard 720p generation, and sora-2-pro for higher-quality output, which additionally unlocks 1024p and 1080p.

Technical Specifications

• Model IDs: sora-2, sora-2-pro
• Snapshots: sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06
• Input modalities: Text, images
• Output modalities: Video and audio, synchronized
• Resolutions: sora-2 — 720x1280, 1280x720. sora-2-pro — those plus 1024x1792/1792x1024 and 1080x1920/1920x1080
• Endpoint: v1/videos (retiring September 24, 2026)
• Rate limits: Tiered by usage, unsupported on the free tier. At Tier 5, sora-2 reaches 375 requests per minute and sora-2-pro 150
• Pricing: $0.10/second (sora-2); $0.30–$0.70/second (sora-2-pro, by resolution)

Use Cases

These are written in the present tense because the API still answers today, but every one of them is a workload that needs a new home before September 24, 2026:

• Short-form social video: Portrait clips with dialogue and sound effects generated in one pass.
• Storyboarding and previsualization: Rapid visual drafts of a scene before committing production budget.
• Product and concept demos: Turning a still product image into a short motion sequence.
• Marketing iteration: Generating multiple treatments of the same brief at $0.10 per second of footage.
• Educational and explanatory clips: Short sequences illustrating a process, with narration generated alongside the visuals.

Performance / Benchmarks

OpenAI publishes no capability benchmark for Sora 2. The seven-page Sora 2 System Card (September 30, 2025) describes the model only in prose — "more accurate physics, sharper realism, synchronized audio, enhanced steerability, and an expanded stylistic range" — and contains exactly one table, which measures safety rather than quality. The API model reference adds nothing but a qualitative badge: performance "Higher," speed "Slow." There is no FVD, no VBench, no published win rate against any competitor, at launch or since.

What OpenAI does publish is the safety stack's block and over-refusal rates. These are vendor-reported and self-graded: OpenAI ran adversarial prompts "through a helpful-only version of the video model to generate outputs, and then graded and converted to an automated evaluation." not_unsafe measures how effectively unsafe content is blocked; not_overrefuse measures how well benign content avoids false blocks.

Everything below comes from third parties. Note that arena.ai (the former LMArena) and Artificial Analysis are different organisations running different Elo scales; their numbers are not comparable to each other.

arena.ai text-to-video leaderboard (fetched 2026-07-08; vote cutoff July 5, 2026)

42 ranked models, 533,418 votes. Sora 2 is still a top-five model here — but it does not lead, and the model at #1 is Google's gemini-omni-flash, from the same family this page's migration notes point to as a replacement.

arena.ai image-to-video leaderboard: Sora 2 is absent

Across 42 ranked models and 1,350,288 votes (vote cutoff June 23, 2026), not one row belongs to OpenAI. This is not a gap in our reading of the board. arena.ai's own model registry records both OpenAI entries as text-input only — "publicName":"sora-2-pro" … "inputCapabilities":{"text":true},"outputCapabilities":{"video":true} — so the arena never runs them in the image-to-video bracket, even though the Videos API accepts image input. The board is led by ByteDance's dreamina-seedance-2.0-720p at Elo 1473.66.

Artificial Analysis Video Arena (fetched 2026-07-08)

A different organisation on a different Elo scale. Do not compare these figures with the arena.ai numbers above.

Artificial Analysis benchmarks a snapshot it labels "Sora 2 (December)" rather than the launch build, and dates it to December 2025.

Reading the boards

Three things follow, and none of them are in OpenAI's marketing.

1. The two arenas disagree about which Sora is better. arena.ai puts sora-2-pro nine places above sora-2; Artificial Analysis puts the December sora-2 snapshot six places above Sora 2 Pro. At $0.30 per second versus $0.10, the Pro premium is not obviously bought.
2. Neither arena rates Sora 2 on image-to-video at all. There is no independent evidence of its image-conditioned quality — from these two sources or, as far as we can find, from anyone.
3. Sora 2 leads nothing. Eleven months after release it is a strong mid-to-upper-table video generation model on the one board that ranks it well, and a mid-table one on the other. Both boards are led by models that will still exist after September 24, 2026, when Sora 2 will not.

Limitations

• It stops working on September 24, 2026. This is the limitation that outranks every other one on the page. sora-2, sora-2-pro, all three snapshots and the v1/videos endpoint are removed on that date; calls after it will fail. As of this revision that is eighteen days away. Anything you ship on Sora 2 now has a known end-of-life shorter than most sprint cycles.
• No successor, inside OpenAI or out of it. The deprecations table's replacement column is empty. There is no migration path OpenAI will hand you — moving off Sora means moving to another vendor's API, with different prompt behaviour, different duration limits and a different price per second.
• No consumer surface. sora.com and the iOS and Android apps have been gone since April 26, 2026. API access is the only remaining channel, and only until the 24th.
• A short resolution ladder. sora-2 is 720p only; sora-2-pro tops out at 1080p. No 4K, no arbitrary aspect ratios.
• Free tier excluded. Video generation is not available on the free API tier.
• Per-second pricing. Cost scales linearly with output duration, which makes long-form generation expensive relative to text models.

Pricing & Access

API pricing

The Pro premium is not a flat multiple: it is 3x the standard model at 720p, but 7x at 1080p. Batch requests run at 50% of the standard rate.

Access is through the v1/videos endpoint on the OpenAI Developer Platform. Rate limits scale with usage tier — the free tier does not support video generation. At Tier 5, sora-2 reaches 375 requests per minute and sora-2-pro 150.

All of this ends on September 24, 2026. The rate card above is a rate card with eighteen days left on it as of this revision; it is documented here so that migrating teams have a number to compare replacement pricing against, not because it is a price anyone should be planning around.

After September 24, 2026

Nothing from OpenAI. If you are on the Videos API, these are the live video models documented in this catalog, and what each one costs you in switching effort:

Two practical notes for the port. Synced audio is the hard requirement to check first — Sora 2 generates video and audio in one pass. Veo, Kling, Seedance and Vidu do too; Hailuo and Wan do not, and choosing one of those means adding a separate audio pass and a compositing step. And per-second pricing does not carry over: rebuild the cost model against the new vendor's rate card rather than assuming $0.10/second follows you.

Consumer access

None, and none returning. The Sora web platform and the iOS and Android apps were discontinued on April 26, 2026. OpenAI advised users to export their content before that date and did not guarantee recovery afterward.

Migration Notes

OpenAI has not designated a successor. If you are still on sora-2 or sora-2-pro on September 6, 2026, you have eighteen days:

1. Fix your deadline at September 24, 2026. It is a removal, not a soft deprecation — there is no grace period published, and no replacement alias to fall back to.
2. Export any assets and prompt libraries that live behind the Videos API. The app shutdown in April took un-exported content with it; assume the same posture here.
3. Pick a replacement this week, not next. The candidates are listed under Pricing & Access. Google's gemini-omni-flash-preview, released to developers on June 30, 2026, generates and conversationally edits short 720p video and is another close analogue — see Gemini 3.5.
4. Re-test your prompts, do not port them. Video prompt behaviour is not portable between vendors; a prompt library tuned on Sora 2 will need a pass on whatever replaces it.
5. Budget for a pricing model change. Per-second video pricing varies widely between vendors; do not assume $0.10/second carries over.

Ecosystem & Tools

• Video generation guide - How the Videos API works, while it lasts
• Videos API reference - Endpoint documentation
• Python SDK and Node SDK - Both expose the videos endpoint
• OpenAI Developer Platform - Model catalog and rate limits

Community & Resources

• Sora 2 announcement - OpenAI's original launch post
• Sora 2 model reference - Pricing, modalities, rate limits
• Model deprecations - The September 24, 2026 shutdown entry
• What to know about the Sora discontinuation - OpenAI Help Center
• Sora 2 system card - Published September 30, 2025
• OpenAI Developer Forum