OpenAI Agent Builder Shutdown | AI Field Notes #44

AnalysisAlibaba's Qwen 3.7 model family reached general availability in two stages last week, and the pricing is the headline. Qwen 3.7 Max ($2.50/$7.50 per million tokens, launched May 20) scores 56.6 on the Artificial Analysis Intelligence Index against Claude Opus 4.7's 57.3, a 0.7-point gap at roughly half the input cost. Qwen 3.7 Plus went GA June 1, adding vision and video understanding on a 1M-token context window at $0.40/$1.60 per million tokens, the lowest price point for a multimodal frontier agent currently available. Both are API-only with no open weights. The Max variant abstains on 52% of factual-retrieval questions, a real constraint for lookup-heavy workflows.

AnalysisOpenAI's biggest ChatGPT upgrade in over a year is a memory architecture overhaul rather than a model swap. Dreaming V3, which began rolling out June 4 to Plus and Pro subscribers in the United States, reduces the compute required to maintain conversation context by roughly 80%. The new system builds memory automatically across sessions, without users having to manually add or edit what the model retains, and doubles stored-memory capacity for paid subscribers. The efficiency gain is the point: it drops the compute cost low enough that OpenAI can now extend persistent memory to free users, expected in the coming weeks.

AnalysisA quiet product update on June 3 put a deadline on a lot of agent workflows. OpenAI notified developers that Agent Builder, its visual no-code canvas for composing agent logic, and its standalone Evals product will both go dark on November 30, 2026. The migration path is the Agents SDK for code-heavy workflows or ChatGPT Workspace Agents for natural-language ones. The shutdown follows the October 2025 launch of AgentKit as a unified platform and suggests the drag-and-drop builder never reached the critical mass needed to justify maintaining it alongside the SDK, which is where OpenAI has concentrated development resources since.

AnalysisColorado's AI Act, the first US state law with enforcement requirements on automated decision-making, takes effect June 30, roughly three weeks away. Companies deploying consequential AI in employment, housing, credit, or healthcare must disclose AI use, conduct bias impact assessments, and let affected individuals appeal. Running counter to it: the Great American AI Act, a 269-page federal bill moving in Congress, would preempt all state AI laws for three years and apply governance requirements only to companies above $500 million in annual revenue. If the federal bill passes before June 30, Colorado's compliance window could be superseded before it runs its first enforcement cycle. If it does not, Colorado becomes the reference point for how other states write their own rules.

AnalysisSpaceX's right to acquire Cursor for $60 billion is about to come due. The deal, struck in April, gives SpaceX the option to buy the AI coding tool (more than 1 million paying subscribers, 50,000 enterprise teams) roughly 30 days after SpaceX's own IPO. The roadshow launched June 4, with pricing set for June 11 and trading starting June 12 on Nasdaq under the ticker SPCX. That puts an acquisition close in July. SpaceX's plan pairs Cursor's IDE with xAI's Colossus cluster in Memphis, described as equivalent to 1 million H100 GPUs. The alternative is a $10 billion collaboration arrangement that stops short of a full purchase.

AnalysisStanford HAI's AI Index 2026 finds enterprise adoption of AI tools reached 65% of surveyed organizations, up 21 points from 44% in 2024. Public trust moved the other way, falling 11 percentage points over the same two years. The Index also found that AI system performance is advancing faster than the ability of evaluators to benchmark it, with frontier models outpacing existing test suites in at least five domains. The adoption-trust gap tracks closely to deployment pace: companies integrate AI faster than they document outcomes, and public trust has tended to follow evidence rather than announcements.