Model scale remains the strongest predictor of LLM performance gains
Source published ·Modelwire updated
Original coverage: Simon Willison ↗·How Modelwire adds context
The development
Ted Sanders clarifies a fundamental principle in LLM development: performance across tasks is multidimensional and unpredictable, yet model scale remains the single strongest predictor of capability gains. Larger models consistently deliver better reliability, factual recall, and general performance, though they incur higher per-token costs. Sanders outlines a clear hierarchy in his company's lineup (Luna through Astra), suggesting that version numbering now encodes multiple signals simultaneously. This insight matters because it challenges the assumption that specialized models will dominate; scale appears to be the dominant lever for practitioners choosing between options.
Modelwire’s AI-generated summary of coverage from Simon Willison.
Modelwire analysis
Skeptical readOur AI-generated reading of the wider context and the next developments to watch.
Sanders frames scale as the single strongest predictor, yet his own company's lineup (Luna through Astra) appears to encode multiple signals simultaneously. The tension between 'scale dominates' and 'version numbering encodes multiple signals' deserves scrutiny: is he describing what practitioners should do, or what OpenAI's portfolio actually does?
This directly contradicts OpenAI's framing from yesterday, where Sanders himself articulated a framework treating model size, reasoning compute, and multi-agent orchestration as independent design levers rather than a unified axis. Yesterday's story showed that smaller models like Terra can dominate specific domains while remaining cost-efficient. Today's claim that scale is the single strongest predictor flattens that nuance. Either the portfolio strategy has changed in 24 hours, or Sanders is conflating 'scale is important' with 'scale is sufficient,' which are different claims.
If OpenAI releases a new smaller model in the next two quarters that outperforms a larger competitor on a specific workload class (like the Terra example), that confirms yesterday's framing and undermines today's scale-dominance claim. If instead all new releases follow a strict size hierarchy, Sanders has walked back the orthogonal-levers position.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·The Decoder
OpenAI positions model size and reasoning as separate optimization axes
OpenAI's Ted Sanders articulates a framework for model selection that challenges the industry's tendency to rank systems linearly by scale. GPT-6.1 Terra demonstrates that smaller models can dominate specific domains like technical writing while remaining cost-efficient, yet lose ground in general reasoning. Sanders positions model size, reasoning compute, and multi-agent orchestration as orthogonal design levers…
MentionsTed Sanders · Luna · Terra · Sol · Astra
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. Simon Willison originally reported this story as “Quoting Ted Sanders”. The full content lives on simonwillison.net. If you’re a publisher and want a different summarization policy for your work, see our takedown page.