Everyone shipped an AI coworker. Google unshipped one.

July 9 was a busy day. OpenAI launched ChatGPT Work alongside the GPT-5.6 family (Sol, Luna, Terra) — an agent that pulls context across apps and files, builds docs and spreadsheets, and runs scheduled tasks that repeat or watch for changes. Hours later, Meta Superintelligence Labs opened Muse Spark 1.1 to outside developers for the first time via the Meta Model API public preview: $1.25 per million input tokens, $4.25 output, $20 in free credits, 1M-token context, US-only for now.

The scoreboard is closer than the launch noise suggests. OpenAI claims 7M+ ChatGPT for Work seats, but Anthropic shipped Claude Cowork back in January and sits on a ~$47B annualized run-rate against OpenAI's estimated $25–33B. Ramp's 2026 AI Index has business adoption at 34.4% Anthropic / 32.3% OpenAI — a coin flip.

Then the other shoe. Google added Nano Banana image generation to Google Earth on July 30 and killed it on July 31, after researcher Henk van Ess generated convincing fake refugee camps on the US–Mexico border and an Iranian nuclear plant. Users produced fake bombings, riots, and a toppled Eiffel Tower from orbit. Google is rebuilding guardrails.

Ship-fast works for agents nobody trusts yet — and breaks instantly for imagery everyone does.