ChatGPT Just Became a Work Agent

What It Covers

OpenAI releases GPT-5.6 model family and ChatGPT Work agentic harness, while Meta launches Muse Spark 1.1 at one-tenth the cost of rivals. The episode covers how cost efficiency has become the primary competitive battleground across all frontier AI labs, displacing raw benchmark performance as the key differentiator.

Key Questions Answered

• •GPT-5.6 Cost Advantage: GPT-5.6 Sol matches or exceeds Claude Opus 4.8 and Fable 5 on most benchmarks while costing 40% less than Opus 4.8 and one-third the price of Fable 5 on the Artificial Analysis coding index. Enterprises blocked by Anthropic's data retention policies on Fable are actively redirecting budgets toward Sol as a direct replacement.
• •ChatGPT Work Harness: OpenAI's new ChatGPT Work interface extends the Codex agentic approach to all knowledge work, connecting to Notion, Google Drive, and Microsoft 365. It runs on cloud instances so tasks continue with the laptop closed. Early enterprise results include compressing month-end financial close from days to hours and generating executive pipeline dashboards from CRM data.
• •Knowledge Work Loop Shift: Early power users report GPT-5.6 Sol enables full autonomous loops of knowledge work, not just task assistance. The practical shift is from executing individual tasks to supervising systems that execute them. Sol's speed advantage over Fable 5 makes it suited for iterative, collaborative workflows rather than long-running autonomous runs.
• •Muse Spark 1.1 Pricing Disruption: Meta's Muse Spark 1.1 costs one-tenth the price of both Fable 5 and GPT-5.5, benchmarks competitively with Opus 4.8 and GPT-5.5, and runs at one-quarter the latency of Opus 4.8. On agentic benchmarks like Jobbench and MCP Atlas, it leads both rival models, making it a viable enterprise option for cost-sensitive agentic workloads.
• •Benchmark Reliability Collapse: OpenAI audited SWE-bench Pro and formally retracted support after finding 30 broken tasks, hidden requirements, contradictory instructions, and contaminated training data. Cursor, Cognition, and Databricks have each launched proprietary benchmarks. Enterprises evaluating models should weight internal task-specific testing over published leaderboard scores, as no single public benchmark reliably measures frontier coding capability.