Claude and ChatGPT’s latest models are nitpicky and burning tokens. Here’s the fix
THE SO WHAT
More capable models that over-index on nitpicking and verbosity are quietly inflating token bills and latency. Operators should be tuning system prompts and response constraints as aggressively as they tune model choice — verbosity control is now a cost lever.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIMartha Stewart is the face of Hint, an AI app for home maintenance
Celebrity-led AI apps are a distribution play for narrow, habit-forming assistants. For consumer builders, the bar is shifting from generic chat to branded, domain-specific workflows that feel like a trusted persona, not a tool.
Applied AINimble claims its new, domain-specialized Web Search Agents cut token costs in half while boosting retrieval accuracy
Domain-specialized search agents that halve token costs point to a future where generic RAG looks wasteful. If you're spending heavily on LLM search, start benchmarking agentic, task-specific retrieval against your current stack and track both accuracy and unit economics.
Applied AIOpenAI launches ChatGPT for Academic Researchers, giving 100K scientists, mathematicians, and engineers free access to its frontier models through 2027
Putting frontier models in the hands of 100,000 researchers for free is a bet that breakthrough use cases will be discovered at the edge, not in-house. If you sell into R&D-heavy sectors, assume your customers' technical teams will rapidly normalize LLM-native workflows and raise the bar on your own tooling.
Disney is ditching Microsoft's GitHub Copilot and adding OpenAI's Codex
A major studio swapping out Copilot for Codex, Claude, and Cursor shows enterprises are willing to mix and match AI vendors at the tool level. If you're standardizing on a single assistant, expect developer pressure for a portfolio approach tuned to specific workflows and licensing terms.