# 2026 AI Stack *Insight · April 10, 2026 · 14 min read* # The AI Capability Stack for 2026 ## What’s Worth Building On and What Will Be Shut Down Next ----- ## Why This Doc Exists On March 24, 2026, OpenAI killed Sora. Not paused. Not sunset. Killed. The app shuts down April 26. The API follows September 24. Disney walked away from a billion-dollar deal. The whole thing lasted six months. Sora was burning roughly a million dollars a day in compute costs while user numbers cratered from a million to under 500,000. The product generated $2.1 million in total lifetime revenue. The math never worked. But the real reason Sora died wasn’t the money. It was compute allocation. Anthropic was winning the coding and enterprise market with Claude Code. OpenAI needed every GPU it had to compete. A consumer video app that nobody was using didn’t survive that calculus. This doc maps where every major AI capability stands right now: what’s stable, what’s fragile, and what’s likely next to get the Sora treatment. If you’re building on any of these tools, this is the risk assessment. ----- ## The Sora Post-Mortem: What Actually Happened **Timeline:** - Feb 2024: Sora previewed. Generated massive hype. - Dec 2024: First public launch for ChatGPT Plus/Pro users. - Sep 2025: Sora 2 launches as a standalone app with social features, deepfake “cameos,” and native audio. - Nov 2025: Downloads peak at 3.3 million across iOS and Android. - Dec 2025: Disney signs a $1B deal to license characters for Sora. - Feb 2026: Downloads collapse to 1.1 million. Daily compute costs remain ~$1M. - Mar 24, 2026: OpenAI announces shutdown. Disney exits the deal. The $1B investment dies. **What killed it:** 1. **Compute economics.** Each 10-second clip cost OpenAI roughly $1.30 to produce. Users were paying $4-8 per clip. At scale, that’s a net loss on every generation. Video inference is orders of magnitude more expensive than text inference. The per-token economics that make ChatGPT viable don’t transfer to video. 1. **Competition caught up.** By Q1 2026, Runway Gen-4, Kling 3.0, and Google Veo 3 had all matched or exceeded Sora’s quality. Sora’s only remaining advantage was the OpenAI brand. That’s not enough to justify $1M/day in burn. 1. **Strategic misfit.** OpenAI is pivoting to a “super app” model focused on coding agents, enterprise tools, and a next-gen model codenamed “Spud.” A consumer video social network doesn’t fit that roadmap. Leadership called Sora a “distracting side quest.” 1. **Liability exposure.** Sora was a deepfake factory. Users generated videos of public figures, copyrighted characters, and fake news footage despite moderation efforts. The legal and reputational risk was growing faster than the user base. **The lesson:** A technically impressive product that burns compute, doesn’t generate sustainable revenue, and doesn’t align with the company’s strategic direction will get cut. Brand equity and billion-dollar partnerships don’t override unit economics. ----- ## The Capability Stack: Category by Category ### 1. Text Generation and Reasoning **Status: Stable. This is the core business for every major AI lab.** |Tool |Best For |Risk Level| |------------------------------|-------------------------------------------------------------|----------| |ChatGPT (GPT-5.4) |General purpose. Broadest capability set. 800M+ weekly users.|Very Low | |Claude (Opus 4.6 / Sonnet 4.6)|Long-form reasoning, writing, code. 200K context window. |Very Low | |Gemini |Google ecosystem integration. Strong multimodal. |Very Low | |DeepSeek |Open-weight reasoning. Chain-of-thought on technical tasks. |Low | |Perplexity |Research with citations. Every response includes sources. |Low | **Why it’s safe:** Text generation is the revenue engine for every lab. ChatGPT, Claude, and Gemini are the highest-margin products these companies have. They’ll be the last things to get cut. **What to watch:** Model commoditization. As the quality gap between providers shrinks, differentiation moves to tooling, integrations, and ecosystem. The model itself becomes less important than what you build around it. **Which model for which task:** - Complex reasoning, analysis, long documents: Claude Opus 4.6 - General daily tasks, image generation, browsing: ChatGPT GPT-5.4 - Google Workspace native workflows: Gemini - Research with verifiable sources: Perplexity - Budget-sensitive technical tasks with open weights: DeepSeek ----- ### 2. Code Generation and Development Tools **Status: Stable and growing. This is where the money is moving.** |Tool |Best For |Risk Level| |--------------|---------------------------------------------------------------------|----------| |Claude Code |Terminal-native agent. 80.8% SWE-bench. Complex multi-file reasoning.|Very Low | |Cursor |AI-native IDE. Fastest autocomplete. Background agents. $2B ARR. |Very Low | |GitHub Copilot|Best value at $10/mo. Deepest IDE integration. 20M+ users. |Very Low | |Codex (OpenAI)|Autonomous coding agent. Async task execution. |Low | |Windsurf |Agentic coding with strong context management. |Low-Medium| **Why it’s safe:** Coding tools are the primary reason OpenAI killed Sora. They redirected compute here because this is where enterprise revenue grows. Anthropic’s Claude Code reaching #1 forced OpenAI’s hand. Both companies are pouring resources into this category. **The production stack most developers are converging on:** GitHub Copilot ($10/mo) for always-on completions, plus either Cursor or Claude Code ($20/mo) as the primary reasoning tool. $30/month covers the full workflow. **Which tool for which developer:** - Terminal-first, complex debugging, infrastructure work: Claude Code - Full IDE experience, multi-file editing, visual workflow: Cursor - Lightweight integration, budget-conscious, existing IDE: GitHub Copilot - Async background tasks, issue-to-PR automation: Codex ----- ### 3. Image Generation **Status: Stable but consolidating.** |Tool |Best For |Risk Level| |------------------------------|----------------------------------------------------------------|----------| |Midjourney |Highest aesthetic quality. Best for creative/artistic work. |Low | |DALL-E (via ChatGPT) |Convenience. Integrated into ChatGPT workflow. |Low | |Adobe Firefly |Commercial-safe. IP indemnification. Creative Cloud integration.|Low | |Stable Diffusion (open-source)|Full local control. No usage limits. Customizable. |Very Low | |Flux |Open-weight. Strong quality-to-compute ratio. |Low | **Why it’s mostly safe:** Image generation is computationally cheaper than video by orders of magnitude. The per-image cost is fractions of a cent. The economics work. And it’s deeply integrated into existing creative workflows. **What to watch:** Standalone image generation apps that don’t integrate into broader workflows. The trend is toward image gen as a feature inside larger platforms (ChatGPT, Canva, Adobe), not as standalone products. **Which tool for which use case:** - High-end creative, concept art, editorial: Midjourney - Quick generation inside a chat workflow: DALL-E via ChatGPT - Commercial work where IP liability matters: Adobe Firefly - Full control, no API dependency, fine-tuning: Stable Diffusion / Flux - Design production, social media, marketing: Canva Magic Studio ----- ### 4. Video Generation **Status: Volatile. This is the category Sora just blew up.** |Tool |Best For |Risk Level | |------------------------|-------------------------------------------------------------------|-----------| |Google Veo 3.1 |Highest quality. 4K, 60fps, native audio. Ecosystem play. |Low-Medium | |Runway Gen-4.5 |Professional creative control. Best editing suite. Studio adoption.|Medium | |Kling 3.0 |Best value. $0.07/sec. Up to 2-min clips. 60M+ users. |Medium | |Pika 2.5 |Social-first. Creative effects. Fast generation. |Medium-High| |Luma Dream Machine 3 |Atmospheric/environmental footage. Physics realism. |Medium-High| |Seedance 2.0 (ByteDance)|Multi-modal reference. Character consistency. |Medium | **Why it’s risky:** Video generation is the most compute-intensive AI capability by a wide margin. Sora proved that even OpenAI couldn’t make the economics work at scale. Every tool in this category faces the same fundamental cost problem: inference is expensive, and users expect low prices. **What Sora’s death tells us about this category:** The market has restructured into four tiers: - **Quality-first:** Runway. Highest Elo rating in blind evaluations. $0.14/sec. The filmmaker’s tool. - **Cost-efficient:** Kling. Production quality at $0.07/sec. Best for volume social content. The market leader by user count. - **Ecosystem-integrated:** Google Veo. Lives inside Google Workspace, YouTube Studio, Google Ads. 4K with native audio. - **Creative/experimental:** Pika. Viral effects, fast generation, social-first. **Shutdown risk factors for this category:** - Any tool burning VC money to subsidize generation credits without a path to profitability. - Any tool that can’t demonstrate enterprise revenue or sustainable unit economics. - Any standalone app without ecosystem integration (the Sora pattern). **Which tool for which use case:** - Cinematic quality, advertising, narrative: Runway Gen-4.5 - High-volume social content, product demos: Kling 3.0 - Native audio, highest resolution, Google ecosystem: Veo 3.1 - Short-form social hooks, viral effects: Pika 2.5 - Character-consistent multi-shot narratives: Seedance 2.0 **The survival rule for this category:** Don’t build your workflow around a single video generation tool. Use aggregators or multi-model APIs. The companies in this space are still defining their business models. Some of them won’t survive 2026. ----- ### 5. Audio and Voice **Status: Stable and maturing.** |Tool |Best For |Risk Level| |---------------------|------------------------------------------------|----------| |ElevenLabs |Voice cloning, TTS, dubbing. Market leader. |Low | |OpenAI TTS/Voice |Integrated with ChatGPT. Real-time conversation.|Very Low | |Suno / Udio |Music generation. Text-to-song. |Medium | |Whisper (open-source)|Speech-to-text transcription. Local deployment. |Very Low | **Why it’s mostly safe:** Audio inference is cheaper than video. ElevenLabs has built a real business with enterprise customers. OpenAI’s voice features are integrated into ChatGPT’s core product. **What to watch:** Music generation apps (Suno, Udio) face the same pattern as Sora: high compute costs, unclear monetization, and growing copyright litigation from the music industry. These are the most Sora-like products in the audio category. ----- ### 6. AI Agents and Automation **Status: Early but strategic priority for every major lab.** |Tool |Best For |Risk Level| |----------------------------|---------------------------------------------------------------|----------| |Claude Computer Use / Cowork|Desktop automation. Browser and file tasks. |Low | |OpenAI Operator / Agent SDK |Web-based task execution. ChatGPT ecosystem. |Low | |n8n |Open-source workflow automation. Self-hostable. AI agent nodes.|Very Low | |Zapier AI |No-code automation for non-technical users. |Low | |LangChain / LangGraph |Agent orchestration framework for developers. |Low | |CrewAI |Multi-agent frameworks. Team-of-agents architecture. |Medium | **Why it’s mostly safe:** Agents are the stated direction for both OpenAI and Anthropic. OpenAI’s “super app” strategy is explicitly about agents. Anthropic’s Claude Code is an agent. This is where both companies are betting their futures. **What to watch:** The fragmentation. There are dozens of agent frameworks, and most of them will consolidate or die. The ones that survive will be the ones that integrate directly with the major model providers or build defensible open-source communities. ----- ### 7. App Building (No-Code / Low-Code) **Status: Growing fast. The “vibe coding” category.** |Tool |Best For |Risk Level| |------------|---------------------------------------------------|----------| |v0 (Vercel) |Frontend components. React/Next.js. Design-to-code.|Low | |Bolt.new |Full-stack apps from prompts. Browser-based. |Medium | |Lovable |Rapid prototyping. Non-developer friendly. |Medium | |Replit Agent|Code generation with deployment built in. |Low-Medium| |NxCode |App building for founders. Full ownership. |Medium | **What to watch:** This category is crowded and most entrants are burning money. The products that survive will be the ones that integrate into existing developer workflows (v0 with Vercel) rather than trying to replace them entirely. ----- ### 8. Research and Knowledge **Status: Stable.** |Tool |Best For |Risk Level| |-----------------|------------------------------------------------|----------| |Perplexity |Cited research. Real-time web search. |Low | |Google NotebookLM|Document analysis. Audio overviews. |Low | |Elicit |Academic research. Paper discovery and analysis.|Low-Medium| ----- ### 9. Productivity and Enterprise **Status: Deeply embedded. Not going anywhere.** |Tool |Best For |Risk Level| |---------------------|--------------------------------------------|----------| |Microsoft 365 Copilot|Enterprise productivity. Office integration.|Very Low | |Notion AI |Workspace assistant. Notes, docs, databases.|Low | |Otter.ai |Meeting transcription and summaries. |Low | |Grammarly |Writing assistance. Tone detection. |Low | ----- ## The Shutdown Risk Framework Sora’s death provides a clear pattern for predicting which AI tools are at risk. Five signals: **1. Compute cost per interaction is unsustainably high.** Sora was spending $1.30 per 10-second clip. Any tool where the inference cost exceeds what users are willing to pay, or what advertising/subscription revenue can cover, is running on borrowed time. Video generation and music generation are the highest-risk categories here. **2. Usage is declining after the novelty period.** Sora downloads dropped 67% from November to February. If a tool’s growth curve peaks in the first three months and then declines, the parent company will notice. The initial spike was hype. The decline is the market’s verdict. **3. The tool doesn’t align with the parent company’s strategic direction.** OpenAI is pivoting to enterprise and coding. Sora was a consumer social app. When strategy shifts, products that don’t fit get killed regardless of their technical quality. Check whether the tool you’re using aligns with where its parent company is heading. **4. Competitors have caught up or surpassed it.** By March 2026, Runway, Kling, and Veo had all matched Sora’s quality at lower price points. When a product’s only advantage is the parent company’s brand, the justification for continued investment evaporates. **5. The tool is a standalone product without ecosystem integration.** Sora was a standalone app. Veo lives inside Google’s entire ecosystem. The standalone AI app model is fragile. Features integrated into platforms survive. Standalone products get cut. ----- ## What’s Most Likely to Get the Sora Treatment Next **High risk (12-18 months):** - Standalone music generation apps without clear paths to profitability - VC-funded video generation tools offering below-cost free tiers - Agent framework startups that haven’t found product-market fit - No-code app builders competing in a crowded market without differentiation **Medium risk:** - Specialized image generation apps that don’t integrate into creative workflows - Meeting/productivity tools in categories being absorbed by Microsoft and Google - Any tool whose primary value is a thin wrapper around a foundation model API **Low risk:** - Tools with sustainable unit economics and growing enterprise revenue - Open-source projects with strong communities (Stable Diffusion, Whisper, n8n) - Features integrated into major platforms (DALL-E in ChatGPT, Veo in Google, Firefly in Adobe) ----- ## The Builder’s Decision Framework When choosing an AI tool to build on or integrate into your workflow: **Ask these five questions:** 1. **Can this company afford to run this product at scale?** Check whether the compute cost per interaction is sustainable. If the product feels too cheap for what it does, someone is subsidizing it. Subsidies end. 1. **Does this tool fit the parent company’s stated direction?** Read their earnings calls. Read their CEO’s blog posts. If the tool you’re using isn’t mentioned in their forward-looking strategy, that’s a signal. 1. **Is this a standalone product or an integrated feature?** Integrated features survive reorganizations. Standalone products don’t. Prefer tools that live inside ecosystems you already use. 1. **What happens if this tool disappears tomorrow?** If the answer is “my entire workflow breaks,” you have a single point of failure. Either diversify or ensure you’re on a tool with very low shutdown risk. 1. **Are there open-source alternatives?** Open-source tools can’t be shut down by a corporate strategy shift. Whisper, Stable Diffusion, and n8n will exist as long as their communities exist. That’s a different kind of durability than any commercial product offers. ----- ## The Two-Tool Stack That Covers 80% of Use Cases For most professionals in 2026, the optimal setup is simpler than the landscape suggests: **Primary reasoning tool:** Claude or ChatGPT ($20/mo). Pick based on whether you prioritize writing quality and code reasoning (Claude) or breadth and multimodal features (ChatGPT). **Primary coding tool (if you code):** GitHub Copilot ($10/mo) for completions, plus Claude Code or Cursor ($20/mo) for complex tasks. **Everything else is situational.** Add Perplexity for research. Add Midjourney or Firefly for images. Add Kling or Runway for video if your workflow requires it. But don’t pay for tools you use once a month. Total monthly cost for a comprehensive stack: $30-60. Less than most SaaS subscriptions. More capable than what entire teams had access to two years ago. ----- ## The Bottom Line The Sora shutdown is a correction, not a crisis. The AI capability stack is stronger in April 2026 than it’s ever been. But the era of assuming every AI product will keep running indefinitely is over. Build on tools with sustainable economics. Prefer ecosystems over standalone apps. Keep open-source alternatives in your back pocket. And never build a critical workflow on a single provider without an exit plan. The compute is finite. The companies allocating it are making hard choices about what to keep running. Make sure the tools you depend on are on the right side of that calculation. [Read on the site →](https://www.agenticamit.com/insight/2026-ai-stack)