Sora Review: OpenAI’s AI Video Dream That Burned Too Bright

Hands-On Verdict

The honest way to judge Sora by OpenAI is no longer about whether it saves time on your weekly work. It can’t - Sora was shut down on April 26, 2026, and its API follows on September 24. What’s worth asking instead is: how did a tool with genuinely groundbreaking technology, a billion-dollar Disney deal, and the full weight of OpenAI’s brand behind it crash this badly, this fast?

As of my 2026-05-01 verification pass, this review serves as a postmortem. Sora 2 produced some of the best AI video I’ve personally seen - the physics, the lighting, the synchronized audio - but the product decisions around it were so fundamentally broken that 99% of users quit within 30 days. The tech was spectacular. The business side was a masterclass in how not to launch a product.

I’m not writing this to pile on. The lessons here matter because every AI tool that follows Sora will either repeat its mistakes or learn from them. And honestly, I’m still sad about it. I wanted this to work.

A Timeline of Rise and Collapse

Sora first appeared as a research preview in February 2024. OpenAI released a handful of demo clips - an SUV kicking up dust on a mountain road, a woman walking through Tokyo at night, woolly mammoths trudging through snow - that genuinely stunned the AI world. The wait for public access became a running joke in creative communities. Ten months passed.

On December 9, 2024, Sora finally launched to ChatGPT Plus and Pro subscribers in the US and Canada. The launch immediately buckled under demand - OpenAI paused new account creation the same day. The interface was barebones: text-to-video only, no audio, resolutions capped at 720p for free users and 1080p for paid. It was impressive, but a lot of us who’d been waiting since February felt the gap between the demos and the practical experience.

Then came Sora 2.

Released on September 30, 2025, Sora 2 was a complete overhaul. Not just a model upgrade - an entirely new product vision. OpenAI pivoted from “creative tool” to “social media app.” Sora 2 launched as a TikTok-like iPhone app with a vertical swipe feed, invite codes, and a “For You” algorithm. The model itself was meaningfully better: synchronized audio with dialogue and sound effects, longer videos (10-25 seconds depending on tier), a Storyboard feature for frame-by-frame direction, and the much-hyped Character Cameos for maintaining consistent characters across scenes.

But the rollout was chaotic. The app was iOS-only and US-only at launch. Android users didn’t get access until November 2025. Most of Europe never got the app without VPN workarounds. Paying ChatGPT Plus subscribers found themselves behind a waitlist while invite codes circulated on Discord and Reddit. The frustration was palpable, and the numbers tell the story.

By January 2026, third-party analytics showed Sora 2 had lost 99% of its active users within 30 days of signup. Installs fell 45% month-over-month. Total revenue across the product’s lifetime was reportedly just $2.1 million - against operating costs of roughly $1 million per day. On March 24, 2026, OpenAI announced the shutdown, and Disney - which had been negotiating a $1 billion investment and licensing deal - pulled out the same day. The app went dark on April 26. The API will follow in September.

What Sora Actually Got Right

None of this changes the fact that Sora 2’s output quality was genuinely impressive. I ran dozens of generations across both tiers, and here’s where it stood out:

Physics simulation was Sora’s defining superpower. Water behaved like water. Smoke billowed through doorways convincingly. Fabric draped under gravity instead of floating in that uncanny AI-purple blob way. In benchmark tests, Sora 2 achieved 9/10 realism scores. When I generated a crowded fish market with splashing water, loose ice, and vendor interactions, the scene held together in ways that made Runway Gen-4 look slightly synthetic by comparison.

Lighting was another clear strength. Sora 2 handled mixed-lighting scenes - practical lamps in moody interiors, golden hour exteriors with lens flare, neon reflections on wet pavement - with a cinematic quality that felt more like David Fincher than a neural network. The color grading in outputs often looked professionally post-processed.

Synchronized audio was the headline Sora 2 feature, and it actually delivered. Ambient soundscapes, footstep Foley, rustling leaves, even dialogue with lip-sync - it worked more often than not. Veo 3 matched this later, but Sora 2 shipped it first, and the initial quality was better than what Runway was offering at the time.

Creative interpretation was excellent. When I gave Sora 2 vague but evocative prompts - “a retired samurai running a Tokyo cat café at 4 AM, rain streaking the windows” - the model made genuinely interesting compositional choices rather than defaulting to generic mush. It interpreted mood, suggesting creative camera angles and atmospheric details I hadn’t specified.

Where the Wheels Fell Off

The technical capability was real. The product experience around it was not.

Region locking and invite hell. Sora 2 launched with an invite-code system on an iOS-only app available exclusively in the United States. I remember watching international creators on Twitter posting screenshots of “Not available in your region” while US-based influencers with invite codes showed off their generations. It created a two-tier resentment that never really healed, even as access expanded.

Pricing without value. ChatGPT Plus at $20/month gave you limited Sora 2 access - enough to dabble, not enough to produce. Pro at $200/month was the only path to Sora 2 Pro and 1080p 25-second clips. Meanwhile, Runway Pro was $28/month and Veo 3 offered free access with 500 monthly credits. For API users, a single 10-second Sora 2 Pro clip at 1080p cost roughly $5 in generation fees. At that rate, an afternoon of iteration could easily run $50-100.

Quality inconsistency. Users reported - and I experienced - noticeable quality degradation over the months following Sora 2’s launch. Early Sora 2 generations were stunning. Later ones, as OpenAI presumably throttled compute to manage costs, had more artifacts, blurrier details, and worse prompt adherence. A Reddit side-by-side comparison from October 2025 showed dramatic regression using identical prompts. This eroded trust fast.

Content moderation overreach. Sora 2’s safety filters were aggressive to the point of comedy. Innocuous prompts involving children or teenagers were frequently blocked. One user reported getting flagged for generating “a person walking a dog” because the model interpreted it as potentially involving minors. The NewsGuard study found Sora 2 generated false or misleading video content in 80% of tested prompts - so the safety system was simultaneously too strict and didn’t stop what it claimed to protect against.

The social-media identity crisis. This is the big one. OpenAI built a TikTok clone powered by an AI video model, apparently expecting people to scroll through AI-generated feeds the way they scroll through human content. They didn’t. The app was neither a serious creative tool (no editing suite, no robust export options, no collaboration features) nor a compelling social platform (the content was AI-generated and the community never materialized). It fell between two stools and landed on the floor.

The Financial Reality

Sora’s collapse wasn’t just a product failure - it was a financial catastrophe. Forbes estimated Sora’s annualized operating costs at over $5 billion, roughly $15 million per day. The Wall Street Journal reported the number at $1 million per day in March 2026 - likely reduced after compute optimizations. Against this, total lifetime revenue was reportedly just $2.1 million. OpenAI lost an estimated $8 billion on Sora in 2025 alone.

When OpenAI canceled Sora, Disney - which had been negotiating a $1 billion partnership to license its IP for AI-generated content - withdrew immediately. The Hollywood Reporter described Disney as “blindsided.” Ars Technica confirmed that no money had actually changed hands between the companies before the deal collapsed.

This wasn’t a profitable product that got sunset. This was a money furnace.

What Should You Use Instead?

The shutdown leaves a gap, but not an unfillable one. Here’s the practical reality in May 2026:

Google Veo 3.1 is the consensus best overall AI video model right now. It matches or exceeds Sora 2’s output quality, supports native audio sync, and outputs at 4K. The free tier gives you 500 credits monthly. The $19.99/month AI Pro plan covers roughly 100 Veo 2 videos. If you’re looking for the closest Sora replacement with better pricing and global availability, this is it.

Runway Gen-4 remains the best choice for professional video work. Motion Brush, camera controls, keyframing, and a genuine editing environment make it a production tool, not just a prompt-and-pray generator. Pricing starts at $12/month (Standard) and $28/month (Pro). It lacks Sora 2’s native audio, but the control precision is on another level.

Kling 3.0 continues to gain ground, particularly for longer clips and global accessibility. At roughly $0.50 per clip with 4K support, it’s aggressively priced and available in far more regions than Sora ever was.

For a full comparison of these options, see our AI Video Generation Guide.

What Sora Taught Us

I think about Sora as two separate things: the model and the product.

The model was remarkable. Sora 2’s physics simulation, cinematic lighting, and audio synchronization pushed the entire industry forward. OpenAI’s research team deserves recognition - they built technology that genuinely expanded what was possible in AI video generation.

The product was a case study in hubris. Launching region-locked, invite-gated, on a single platform, with pricing that made Runway look cheap and Veo look generous, while simultaneously trying to be a social network - it reads like a parody of Silicon Valley excess. The 99% user churn wasn’t about the technology. It was about execution.

For anyone building in the AI space, Sora’s arc carries a clear message: the best model in the world can’t save a product that ignores its users. Impressive demos and technical benchmarks mean nothing if people can’t access your tool, can’t afford to use it seriously, and don’t trust it to be here next quarter.

I started this review impressed by Sora. I’m ending it disappointed - not in the engineers who built it, but in the decisions that killed it.