06-reference

every agent model access gap

2026-06-28·reference·source: Every·by Mike Taylor

"Everyone Gets an Agent. Almost No One Gets the Model." — @miketaylor

Why this is in the vault

Mike Taylor's "metered intelligence" thesis takes its sharpest form yet: agent access is becoming universal while frontier model access is being actively rationed — first by U.S. government directive, then by capital allocation — and this bifurcation will define which organizations can actually compound on AI.

Issue contents

Lead: Token access as capital allocation (Mike Taylor / Context Window)

The trigger is concrete: OpenAI released GPT-5.6 Sol and, by U.S. government directive, access was restricted to roughly 20 pre-approved companies while Washington works out frontier-model release policy. Dan Shipper called this directly: a world where advanced models are locked to AI giants and select companies "is one where ambitious students, independent builders, and working professionals are denied the tools they need to learn, create, and compete."

Taylor's extension: that rationing is coming for everyone via economics, not just regulation. Token compute will be allocated like capital — the biggest budgets go to whoever can prove the biggest returns. As models get more capable, the gap between those with frontier access and those on yesterday's tier widens faster.

Counterpoint running alongside: agent access is genuinely expanding. Codex hit 5 million weekly active users. Anthropic's Claude Tag landed in Slack. Every's own Compound Engineering plugin can now run a coding agent unattended for hours — "long enough to build a feature, write its tests, and open a pull request on its own."

Alignment: Midjourney body scanner (Ashwin Sharma)

Midjourney is building a full-body ultrasound scanner — 500K sensors, 60-second scan, no radiation or magnets, targeting medical spas first. Stated goal: 1 billion scans/month. Sharma's verdict: the body-composition use case is solid, but routing anxious healthy people into a healthcare system already under strain risks manufacturing demand for follow-up biopsies and referrals rather than catching disease early. He'd do the scan for body composition; would not use it for diagnostic hunting.

Source Code: Claude Code as the only agent-builder you need (Nityesh Agarwal)

Makes the case that Claude Code subsumes specialized agent-builder tools (OpenClaw cited). Argues Claude Code's combination of file-system access, shell integration, and model quality makes it a general-purpose coding agent that most teams don't need to replace with a dedicated tool.

Working Overtime: AI career audit (Katie Parrott)

Parrott handed her career review to Codex. Framed as a practical test of whether AI can surface blind spots in self-assessment. Secondary piece — minimal RDCO relevance.

AI & I Podcast: What it means to be human (Dan Shipper × Edwin Chen, Surge AI)

Edwin Chen (Surge AI, former Twitter ML lead) on what remains distinctly human once machines can do everything. Philosophical register, not operational.

Studio updates

Mapping against Ray Data Co

Directly continuous with the metered-intelligence arc. Taylor's prior piece (2026-06-05-every-microsoft-metered-intelligence.md) framed the end of subsidized LLM subscriptions; this issue names the next-order consequence — not just that tokens cost more, but that frontier access itself gets gatekept. For RDCO, which runs an always-on Claude Code loop billed by the token, both dynamics are live operating costs, not abstractions.

The bifurcation thesis is a positioning opportunity. If frontier access concentrates at the top, the clients RDCO serves (mid-market data teams, phData engagements) will increasingly operate on yesterday's model tier while competitors with negotiated enterprise access move faster. The practical implication: understanding the model-tier gap becomes a consulting leverage point — helping clients maximize ROI on the tier they have access to, rather than assuming frontier parity.

Claude Code as the canonical agent-builder (Agarwal's piece) reinforces RDCO's current stack choice. No tooling switch warranted, but confirms the bet is mainstream rather than niche.

Government-directed access rationing is a new variable. The GPT-5.6 Sol restriction is the first concrete example of frontier access being regulated rather than just priced. This could affect phData client procurement timelines if enterprise agreements with OpenAI or Anthropic get caught in regulatory review windows.

Related