GPT-6 Astra, Sol, and Luna: Splitting Work Across One Lab's Stack

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

GPT-6 Astra, Sol, and Luna showed up in the notes as a lab split: computer-use strength near human levels, Excel tasks beating junior analysts on investment-banking drills, GDPval wins across dozens of jobs, plus paper hints about automated AI researchers later in the decade. The useful move was not worship. It was splitting work across named strengths — Astra for heavy computer use, Sol for cheaper near-peer cognition, Luna for decision-shaped gates — while I kept orchestration and accountability.

Stop asking if AI takes the job

The notes pushed a shift: stop doom-looping on replacement, start assigning daily tasks to Codex with Astra-class computer use. I agree with the operational half. I disagree with erasing labour politics. My version: assign tasks, measure quality, keep humans on accountability and taste, renegotiate roles with eyes open.

Unlimited AI coworkers without managers is how you get unlimited confident mistakes. Orchestration is the job.

Split by strength, not by brand loyalty

Computer-use scores near or above humans on some benchmarks, Excel agents beating juniors on IB tasks, broad GDPval wins — priors, not permission slips. I route UI-heavy chores to computer-use models, spreadsheet analysis to Excel-native surfaces, and decisions to a Luna-shaped gate even if I build the gate myself.

Steerable mid-thought behaviour matters for long tasks. I test steerability by interrupting mid-run. Models that ignore the interrupt lose production privileges.

Recursive research hype on a leash

Papers hinting at automated AI researchers and recursive self-improvement toward ASI are watchlist material. They are not architecture for a client CRM pilot in Karachi. I keep a horizon document separate from the sprint board so horizon awe cannot rewrite this quarter's scopes.

Clients who want ASI strategy get a workshop. Clients who want Excel agents get a scoped pilot with evals.

Daily tasks to Codex

Start small: inbox triage drafts, spreadsheet cleaning, UI regression clicks in staging. Each task has a definition of done and a rollback. Astra-level computer use stays in staging until the success rate is boring.

Boring success rates are the only rates I promote.

GDPval humility

Beating professionals on a benchmark suite across dozens of jobs is impressive and narrow. Real jobs include office politics, incomplete data, and moral blame. I use GDPval-like results to prioritise which chores to try, not to announce a department is obsolete.

Obsolete announcements without transition plans are cruelty dressed as foresight.

Sol as the cost twin

Near-Astra at a fraction of price is a routing gift when evals agree. I keep Sol on the candidate list for volume computer-use after Astra proves the pattern. Twin models need twin evals — never assume price cut equals quality twin without the table.

Luna as the habit of deciding

Even without a vendor Decisions API, I implement propose/decide/act. Luna is a name for the habit. Approvers see diffs and blast radius. Agents do not self-approve because the benchmark said they are peer-level.

Orchestration offer

Named model roles, daily task menu, staging-first computer use, decision gates, horizon doc quarantined from sprints. That is how I split one lab's stack across real work without cosplaying the singularity on a client's dime.

Interrupt test

Before trusting Astra-class computer use on a new chore, I interrupt mid-run and require a clean pause. Models that keep clicking fail the steerability bar. Sol does not get a free pass for being cheaper. Luna gates still see the proposal.

Orchestration without interrupts is just hope with a cloud bill.

Horizon doc quarantine

ASI timelines and automated researcher papers stay off the sprint board. Astra, Sol, and Luna get tasks with owners. Orchestration meetings that drift into eschatology get the horizon doc link and a redirect.

Staging computer-use first

Astra-class clicking stays in staging until success rates are boring across a week. Sol may follow on the same chores after a twin eval. Luna still gates production-shaped decisions. GDPval slides stay in the appendix.

Daily task menu

A literal menu of chores Codex-with-Astra may take: triage drafts, sheet cleans, staging UI clicks. Items not on the menu require a human add. Unlimited coworkers still need a menu or they invent work.