Take-Two Cuts Its AI Team, Nexon's Trojan Horse, and Cursor Abandons VS Code
Studio AI disclosure debates are now in their second consecutive week of heavy coverage, and the stakes escalated. Take-Two cut its entire AI team (including the head of AI who built exactly what the CEO asked for), while Nexon pitched Arc Raiders as a “Trojan Horse” for company-wide AI adoption. On the tools side, Anthropic locked third-party harnesses out of Claude Code subscriptions and Cursor 3.0 abandoned its VS Code roots for an agent-orchestration platform.
Industry
Luke Dicken, Take-Two’s Head of AI, confirmed on LinkedIn that he and his team had been laid off. Fifteen outlets covered it.
About half reported it straight. The other half framed it as a contradiction: the CEO said he was “actively embracing” AI, then cut the AI team. One outlet called it “abandoning neural networks.” That framing oversimplifies. Zelnick has said the same thing publicly since at least 2023: AI for production efficiency, not creative work. He called the idea of AI-generated hits “laughable.” Dicken’s team, which described seven years of work building production tools, joined through the Zynga acquisition. A senior director cited “shifting priorities from upper management” as the reason, according to Game Developer. Take-Two declined to comment.
At a Nexon capital markets briefing, Embark CEO Patrick Söderlund said his team shipped The Finals and Arc Raiders with “significantly fewer people, at a fraction of the cost.” Nexon CEO Junghun Lee called Arc Raiders a “Trojan Horse” for Mono Lake, an internal AI platform the company intends to apply across all studios. Nine outlets covered the briefing. None verified the cost or headcount claims. These are investor-presentation assertions from a company pitching its AI story to the market.
Three quieter stories round out the week. Digital Extremes ruled out generative AI for Warframe. Sony acquired Cinemersive Labs for its visual computing team; the startup converts 2D photos and videos into 3D volumes. Pearl Abyss drew player backlash after undisclosed AI-generated assets surfaced in Crimson Desert promotional material.
Code
Anthropic changed Claude Code’s pricing to block third-party harnesses like OpenClaw from using subscription limits. The open-source tool now requires a separate pay-as-you-go plan. Anthropic’s head of Claude Code cited “engineering constraints,” but the timing matters: OpenClaw creator Peter Steinberger moved to OpenAI and called the move anti-competitive, claiming Anthropic copied features into its closed harness before locking out open-source alternatives. Anthropic is offering refunds. For studios evaluating coding agents, this is a vendor lock-in data point. techcrunch.com
Cursor shipped 3.0, a complete rebuild that abandons its VS Code fork for an agent-orchestration architecture with cloud and local agents, agent-first git management, and a proprietary model called Composer 2. Early reviews were mixed: fast and lightweight, but “not yet mature enough to merit a switch” from established agent workflows. One tester reported roughly $2,000 in costs during evaluation, raising questions about predictability at studio scale. every.to
OpenAI moved Codex to pay-as-you-go pricing for ChatGPT Business and Enterprise teams. Separately, according to Coder, the company raised a $90M Series C for its secure cloud development environments; KKR reported that after deploying Coder to 500+ engineers, over half of commits moved into Coder environments. openai.com
Audio & Speech
ElevenLabs launched ElevenMusic, an iOS app for generating songs up to three minutes from text prompts, entering the AI music market against Suno and Udio. Free users get seven songs per day; a Pro tier at $9.99/month offers 500 tracks. The company also opened a Music Marketplace inside ElevenCreative, where creators can license AI-generated tracks under three commercial tiers, one explicitly designed for games and apps. ElevenLabs reports its Voice Marketplace has paid out $11 million to creators. According to reports, the music push follows a $500M Series C at an $11B valuation in February, positioning it as a full-spectrum audio company rather than a pure voice vendor. tech.yahoo.com
IBM integrated ElevenLabs TTS and STT into watsonx Orchestrate for enterprise voice agents, supporting 70+ languages with HIPAA-aligned zero-retention mode. Mistral released its first TTS model. The annual Enterprise Tech 30 VC survey placed ElevenLabs at the top of its late-stage category, ahead of other AI infrastructure companies. sifted.eu
Microsoft entered voice AI directly. Its MAI Superintelligence team shipped MAI-Voice-1, generating 60 seconds of audio in one second with custom voice support, starting at $22 per million characters. MAI-Transcribe-1 handles speech-to-text in 25 languages at 2.5x the speed of Azure Fast. Both are available on Microsoft Foundry. techcrunch.com
Tool Updates
Autodesk Flow Studio: Autodesk launched Wonder 3D, a generative AI model integrated into Flow Studio that converts text and images into editable 3D assets with geometry and textures. Exports are available in USD, STL, and OBJ formats for downstream use in game engines and physical prototyping. Nikola Todorovic, co-founder of the acquired Wonder Dynamics, positioned the tool for rapid prototyping of characters, creatures, and props across game development and film production. (Autodesk)
Runway: Runway launched a $10M venture fund and a Builders program aimed at startups building on its foundation models. The fund writes checks up to $500K for pre-seed and seed companies working on new AI architectures or media formats. The Builders program offers seed through Series C startups 500,000 free API credits and early access to Characters, Runway’s real-time video agent API. Co-founder Alejandro Matamala-Ortiz said the goal is to explore use cases for its “general world models” beyond what Runway’s 150-person team can pursue, explicitly citing gaming applications. (TechCrunch)
Gemini: Google added Flex and Priority service tiers to the Gemini API, letting developers route different job types through a single synchronous interface instead of managing a separate Batch API. The Flex tier cuts prices by 50% for latency-tolerant background work like data enrichment and agentic reasoning. The Priority tier targets interactive, user-facing tasks requiring higher reliability. Both tiers are live, controlled by a single service_tier parameter in the API request. (Google AI Blog)
GitHub Copilot: GitHub Copilot injected promotional text for itself and Raycast into over 11,000 pull requests, marked “START COPILOT CODING AGENT TIPS.” After developer Zach Manson flagged the behavior, GitHub VP of Developer Relations Martin Woodward initially acknowledged the “product tips” were disabled due to negative feedback, calling them “icky.” A formal statement the following day reframed the insertions as a “programming logic issue” rather than advertising, and GitHub confirmed the tips have been removed from PR comments. (Windows Central)