October 14: the trade secret wars begin

A third credible lab is now shipping genuinely frontier-level models, which should mean more competition on price and faster feature cycles across the tools you already use.

Engineers: Wonder is priced at $4 per million input tokens and $20 per million output tokens via Meta's API, undercutting comparable models from OpenAI and Anthropic — worth benchmarking against your current provider for coding and agent workloads.

Managers: Meta built in "active memory" to retain context across long tasks, a feature aimed squarely at agentic workflows — relevant if you're evaluating AI assistants for team processes.

Do this: If you use AI coding or agent tools, try Muse Wonder via the Meta Model API and compare cost and output quality against your current model before switching anything.

Watch what happens when "AutoGov" meets messy reality — unclear consent rules around AI use of medical records, and inter-department turf wars — because those are the same snags any large organization hits when automating workflows.

Managers: Abu Dhabi's "crawl, walk, run" approach — align goals across departments before automating — is a more useful blueprint for rolling out AI internally than most vendor pitches.

Do this: Nothing to do yet — just note this as the reference case if your organization debates AI-driven process automation next year.

The era of near-free frontier-grade AI tokens is ending even for the budget leader, so build cost monitoring into any agent workflow now, regardless of which model you use.

Engineers: Harness treats every part of the agent runtime as a plugin, so it's worth evaluating as a model-agnostic alternative to Claude Code or Codex, especially since it already supports DeepSeek, Anthropic and OpenAI endpoints — but expect breaking changes, since the project explicitly warns it's a developer preview.

Finance: Budget for AI infrastructure like a variable cost, not a fixed one — peak-hour surcharges and steep cache-pricing hikes mean the same workload can now cost four times more depending on when it runs.

Do this: If you run agent workloads against DeepSeek's API, audit your cache-hit ratio and shift non-urgent jobs to off-peak hours before the new pricing lands.

Who can borrow cheaply is starting to shape who can afford the compute race, not just who has the best models.

Finance: Watch dim sum bond spreads and yuan funding costs as a new lens on competitive positioning between US and Chinese AI spenders, alongside their capex disclosures.

Do this: Nothing to do yet — just be aware that currency and credit markets, not just chip supply, are becoming a factor in the AI spending race.