SED News: Restricted Models, IDE Wars, and the DeepMind Mafia

What It Covers

SED News examines AI model access restrictions by the US government affecting Anthropic's Fable and OpenAI's GPT-5.6, the IDE landscape shift following SpaceX's $60 billion Cursor acquisition, London's failure to retain DeepMind alumni value, and the emerging cost gap between Claude and open-weight models like DeepSeek.

Key Questions Answered

• •AI Model Vendor Lock-in: Choosing a coding tool now means buying into an entire model ecosystem. Claude Code ties developers to Anthropic's models, Codex to OpenAI, and Cursor now to SpaceX's Grok. Evaluate switching costs before committing — accumulated context, chat history, and workflow integration make migration expensive, similar to database migration costs.
• •Open-Weight Cost Arbitrage: DeepSeek V4 Pro costs approximately $0.44 per million tokens versus Claude Opus 4.7 at roughly $5.00 — an 8x price difference. Teams spending heavily on Claude Code should benchmark open-weight models against their specific workloads, as developers report increasingly competitive output quality that may justify the performance tradeoff.
• •Government AI Restrictions Create Sovereign Model Demand: US government restrictions on Fable and GPT-5.6 have no published compliance standards — no defined checklist exists, making it effectively arbitrary. Non-US organizations should factor sudden model unavailability into procurement risk assessments and consider open-weight or regionally developed alternatives like Mistral as hedges against foreign government intervention.
• •LLM Resume Screening Unreliability: HackerRank's open-sourced ATS, stress-tested by running identical resumes 100 times, produced scores ranging from 66 to 99 out of 100. LLMs perform reliably on binary checklist items like skill verification but produce coin-flip results on subjective criteria like architectural complexity — avoid using LLM-generated scores as dependable hiring metrics.
• •Token Efficiency Over Token Volume: The "compound correctness" principle argues that spending more tokens on capable models upfront reduces total token consumption versus cheaper models requiring multiple correction iterations. Teams optimizing purely for per-token cost may be measuring the wrong variable — measure total tokens consumed per completed task, not cost per individual generation.