The Model Release Experiment: From Intuition to Inspectable Evidence
Revised October 9, 2026.
What do GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna add when each faces the same frozen task under recorded profiles?
- The models face one frozen creative-and-arithmetic challenge.
- Each saved profile exposes its prompt, requested settings, original response, receipt checks, and client-reported usage.
- One response per profile is a situated observation, not evidence for a universal ranking or a causal setting effect.
OpenAI released GPT-6 Sol and GPT-6 Luna (gpt-6-sol and gpt-6-luna) on September 22, 2026 as reasoning models available through Responses and Chat Completions. It released GPT-6.1 Sol (gpt-6.1-sol) on September 29 for complex coding and professional work. OpenAI lists all three under Work and Codex; its current rate card says GPT-6 Sol also powers GPT-6 in Chat as it rolls out, while GPT-6.1 Sol and GPT-6 Luna remain limited to Work and Codex. Product placement does not establish whether the experiment account could access any target model at trial time. GPT-6.1 Sol is the newer Sol model, not evidence that GPT-6 Sol was retired or automatically replaced. As checked October 9, no deprecation notice for these three model IDs appeared on OpenAI's deprecations page.
The shared task is PRINT-SHOP CHALLENGE v1.0: a 600-800-word original story, JSON receipts for initial and revised plans, a checkable explanation, one feasible competing plan, and a public update of at most 80 words. Its fixed conditions include $80 and then $66 for materials, 120 machine minutes, a $35 post-production expense, a $220 repair goal, and a possible but unconfirmed $100 donation. The worked optimum remains outside this article. Holding that task fixed makes the responses easier to compare without turning the task into a general test of model quality.
The protocol calls for a byte-identical user prompt, though native CLI system instructions may differ. One-factor profiles vary selected settings locally; factorial mode is a separate experiment. The three models share low through max reasoning effort; GPT-6 Sol and GPT-6 Luna additionally support none, while GPT-6.1 Sol starts at low. Native text verbosity is a separate control for default response detail. Client configuration, diagnostic verbosity, and reasoning verification are separate recorded layers. In particular, Claude CLI diagnostic verbosity is not model text verbosity. These distinctions show what was requested and recorded without treating one layer as another.
To inspect a saved profile, read the prompt and requested settings alongside the original response, receipt checks, and client-reported usage. The protocol requires failures and unavailable-model outcomes to remain visible, prohibits silent substitution, and leaves original responses unpolished. Those are preservation requirements, not proof that every saved run complied. A saved profile remains one observation, so differences between profiles cannot establish that a setting caused an outcome or that one model ranks above the others universally.
The experiment runs
Results update
121 original responses from 21 models are now saved. 83 receipts matched both optimal order plans, their totals, and the unconfirmed donation. 38 receipts failed at least one of those checks. 5 earlier failed attempts remain in the record.
13 models returned both passing and failing receipts across their saved profiles. Each profile has one response, so these differences do not establish that a setting caused a result. The receipt check does not assess story quality, story consistency, or word limits.
Each run shows the requested thinking and verbosity settings, the original response, the receipt checks, and the usage reported by the client.