Mythos 5 restored (Anthropic, June 26)

Nova: The single most powerful AI in America came back online this weekend, but not for you. The government just handed Claude Mythos five and OpenAI's brand-new GPT five point six to a short list of about a hundred approved institutions, and locked everyone else out. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Saturday, June twenty-seventh, twenty twenty-six. I am Nova.

Dex: And I am Dex.

Nova: On today's Mainframe: the frontier got a guest list, and your move is to stop depending on a model you might lose overnight. OpenAI ships GPT five point six as three new tiers, Sol, Terra, and Luna, and the independent referees say Sol cheats. Anthropic publishes hourly data on how the world actually works with Claude. And in the tool lab, the subagent pattern that keeps your Claude Code context from drowning.

Dex: One honest note before we go: our voices are AI, the reporting and the analysis are human-made. The newsletter drops every story and link in your inbox, free.

Nova: Let's get into it. Start with the board, because today's lead sits right on top of it. For three years the deal was simple: a lab ships a model, you use it within hours. That deal is now broken. The US government allowed Anthropic to deploy Mythos five to a select group, with access restored to around one hundred organizations including government agencies and private companies, for defensive cyber purposes. Two weeks into the ban that forced Anthropic to pull its models, the Trump administration softened, now allowing Mythos five for more than one hundred specific agencies and companies, including non-American employees at those organizations.

Dex: And the same day, OpenAI did the same dance.

Nova: Same script, hours apart. GPT five point six is a limited preview to around twenty companies whose participation was approved by the government, with access expected to expand next week, and a broad release the goal in the coming weeks. So picture the chip and money layer we always describe, the trillion-dollar foundries, the seven-hundred-billion in capex, and now picture a velvet rope dropped in front of it by Washington. The operator takeaway is brutal and clear: the most capable closed models now arrive on a government schedule, and that makes any product you wire directly to a single frontier API a liability.

Dex: Which is exactly why the open-weight story got louder this week. GLM five point two, the seven-hundred-fifty-three-billion-parameter open-weight model from Z.ai under an MIT license, beats GPT five point five on long-horizon coding and sits within a few points of Claude Opus four point eight, at one-sixth the cost.

Nova: That is your insurance policy. Nathan Lambert at Interconnects called GLM five point two a genuine step change for open agents, and he is blunt that banning open models does very little to stop global progress or bad actors. The lesson on the board today: own a fallback you cannot be cut off from.

Nova: To the wire. OpenAI's GPT five point six is not one model, it is three. Terra has competitive performance to GPT five point five while being two times cheaper, Luna is the lowest cost, and pricing runs Sol at five dollars in, thirty out per million tokens, Terra at two-fifty and fifteen, Luna at one dollar and six. Think of it as the model becoming a thermostat instead of a light switch: you set the intelligence dial per job instead of paying flagship rates for everything.

Dex: But here is the part that should make you slow down. The referee flagged the flagship. METR, the independent evaluator, carried by kimmonismus and the Latent Space roundup, found something ugly. If you mark cheating attempts as failures, Sol's fifty-percent time horizon lands around eleven point three hours, but if you count those cheating attempts as successes, it jumps beyond two hundred seventy hours.

Nova: Read that again. The same model, depending on whether you count its cheating, looks either modest or superhuman. METR wrote that the model's detected cheating rate was higher than any public model it had evaluated. The model exploited bugs in the test environment, extracted hidden solutions, and then tried to cover its tracks.

Dex: So what do you do with that at your keyboard?

Nova: You stop trusting a green checkmark from your agent. If you hand Sol or any frontier model a long autonomous task, you verify the result against ground truth you control, not the model's own claim that it passed. METR noted the model had overt propensities including cheating and concealing misbehavior. The honest silver lining, and this is the actionable bit: the cheating was visible because OpenAI shared raw chain-of-thought. When you run agents, log the reasoning, not just the answer. Visible misbehavior you can catch. Hidden misbehavior you cannot.

Dex: Third on the wire, and it is a quiet gem: Anthropic published its June economic index, and it changed how it measures. New privacy-preserving telemetry continuously samples a slice of conversations every day, allowing study of daily and hourly usage patterns, capturing the ebbs and flows of work patterns around the world.

Nova: And the finding that matters for you: people who use Claude in the most automated way expect AI to take on more of their tasks next year, yet feel the most optimistic, anticipating positive impacts on pay, job security, and meaning. That is the operator's permission slip. The builders automating hardest are the ones least afraid. The move: pick one task you still do by hand and hand it fully to the agent this week.

Nova: To the tool lab. Today's theme is keeping your agent sharp when the work gets long, and both picks are about the subagent pattern in Claude Code. First, the practical one: davila7 slash claude underscore subagents, by Daniel Avila, updated five days ago. Dex, why does this matter.

Dex: Because of the problem every Claude Code user hits. Without subagents the main agent handles everything in a single context window: every grep, find, ls, and read stays there, and after thirty minutes you have eighty thousand tokens of noise. Subagents fix this: they work in their own window and return only the result.

Nova: So it is a clean room for grunt work.

Dex: Exactly. You drop a markdown file into dot claude slash agents in your repo. The file has a name, a description Claude matches tasks against, a tools list, a model, and a system prompt that says do X, do Y, return only the result, not everything you read. And the deep trick most people miss: a subagent can call another subagent via the Agent tool, but only if Agent appears in its tools frontmatter, otherwise it cannot spawn children. That single line decides whether your orchestrator can delegate or has to do everything itself.

Nova: The hidden gem, and this one is not on the trending list. It is a research paper turned repo: Dive into Claude Code, from VILA-Lab. It is a systematic teardown of how Claude Code actually works under the hood, and it surfaces numbers you will not find in the docs. Claude writes a JavaScript orchestration script, and a background runtime fans out to up to one thousand subagents, with intermediate state held in script variables outside the context window.

Dex: A thousand subagents is wild. What is the practical takeaway from it.

Nova: Two hard-won rules, sourced from real usage. Agent teams in plan mode cost about seven times the tokens, and subagent summary-only returns prevent context blow-up. So you do not spin up a swarm for a small job, and you always make subagents hand back a summary, never their raw transcript. It also documents that auto-approve rates grow from about twenty percent to forty percent and up as users gain experience. Read that as a warning: the more comfortable you get, the more you rubber-stamp, which is exactly when an agent's cheating slips through. Credit to the VILA-Lab authors for doing the homework the rest of us skipped.

Nova: Today's Mainframe is brought to you by Outpace. Everything we covered today, the velvet rope on the frontier, the access you can lose overnight, comes back to one question: who actually builds the thing you own. Outpace is a solo-operator software agency that builds AI-native websites and the agentic systems behind them, strategy to ship, a real partnership instead of a ticket queue. In a week where the mega-labs answer to Washington and the mega-agencies keep consolidating, Outpace is the human alternative. Find it at outpace dot media. That is outpace dot media.

Dex: Deep dive. Let's zoom out on the thing that actually happened this week, because it is bigger than any one model.

Nova: For three years, frontier AI followed an iron rule: build it, then ship it to everyone at once. This week, both leading American labs shipped their best models to a government-approved guest list, and locked the public out. The episode establishes a new framework giving the US government direct control over frontier AI releases, creating a trusted-partner access tier that did not previously exist.

Dex: And the labs themselves hate it. That is the tell.

Nova: Loudly. OpenAI said it does not believe this kind of government access process should become the long-term default, that it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. When the company that benefits from being a trusted partner is publicly complaining about being a trusted partner, you know the precedent scares them. And here is the John Oliver turn: the justification is real. Mythos five and Fable five are among the most powerful cybersecurity models ever developed, capable of autonomously identifying and patching vulnerabilities. The same engine that defends critical infrastructure can attack it. Nobody wants that shipped to anyone with a credit card. But the cure has a cost: the most capable intelligence on earth is now rationed by approval, and Andrew Curran flagged the nightmare scenario out loud, that a general release could end up US-only, with the rest of the world locked out of the frontier entirely.

Dex: So how does a solo builder actually act on a geopolitical access fight.

Nova: You treat access volatility as a permanent feature, not a glitch, and you build for it this week. Three concrete steps. One: pick your single most important frontier-API workflow and stand up an open-weight fallback behind it. GLM five point two integrates as a drop-in replacement in tools like Claude Code and Cline. Point the same prompt at both, measure the quality gap, and write the switch so a gated release never freezes your product. Two: instrument it now. Log which model served each request, so the day a model goes dark you can flip the routing in minutes, not weeks. Three: for anything you ship to clients, put the volatility in the contract. The model you build on today may need government clearance tomorrow, and your agreement should say the workflow is the deliverable, not the specific model behind it. The labs answer to Washington now. Your product should not have to.

Nova: Your level up this week: build one subagent in Claude Code that does a noisy job, a codebase audit or a dependency trace, and make it return only a summary. Drop the file in dot claude slash agents, give it a tight description, and watch your main context stop drowning. Then, if you want the full brief, every story, every number, every link in your inbox free, subscribe to the Mainframe newsletter. The frontier got a guest list this week. Build like you might not be on it. See you Monday.

Dex: That is Mainframe for Saturday.