Key Points
- OpenAI's GPT-5.6 Soul scores 7.78% on ARC AGI v3, a five-fold jump from Opus 4.8's 1.5%, showing progress on general reasoning tasks beyond pattern-matching.
- Soul autonomously post-trained the smaller Luna model, decoupling capability gains from raw pretraining compute and opening a new scaling dimension.
- The frontier remains fragmented: users report reaching for Soul 95% of the time for speed and cost while preferring Claude for the hardest problems.
Summary
OpenAI's GPT-5.6 Soul Shows Progress on General Reasoning, But Frontier Remains Fragmented
OpenAI released GPT-5.6 Soul with expanded coding and agentic capabilities, including a multi-agent Ultra mode. The model scored 7.78% on ARC AGI v3, up from Opus 4.8's 1.5%—a meaningful jump on a benchmark designed to test reasoning tasks any human should solve.
The context matters. ARC AGI v3 deliberately avoids the "spiky" problems where AI now dominates: mathematical optimization, programming, hacking. Instead it tests spatial reasoning, puzzle-solving, and generalization. At 7.78%, Soul is nowhere near human parity, but the progression from 1.5% suggests the model is developing broader reasoning chops beyond pattern-matching on internet-scale training data.
The frontier is still spiky and fragmented. Observers are comparing Soul to Anthropic's Claude (positioned as a "collaborative coworker") rather than declaring one model a clear winner. Dylan Field, Figma's CEO, argues the comparison between Anthropic and OpenAI's models is "apples and oranges," noting that model training still has unexplored territory ahead. Some users report preferring Claude for the hardest problems while reaching for Soul 95% of the time due to speed and cost. The pareto frontier is real: multiple companies are growing revenue and accelerating growth simultaneously, even as relative market share shifts because the overall market is expanding faster than any single player.
Self-training marks a shift in how capability scales. OpenAI's blog noted that Soul autonomously post-trained the smaller 5.6 Luna model. This decouples model improvement from raw pretraining compute, opening a different dimension for scaling.
Model numbering has lost semantic meaning. Version numbers no longer map cleanly to a single training phase. Post-training, reasoning-specific scaling, and autonomous optimization make it hard to bake capability into a single integer. The numbers are drifting toward release year instead—a "5.6 in 2026" rather than a strict architectural ranking. Google's Gemini lineup already reflects this confusion: 3.5 Flash and 3.1 Pro coexist with unclear hierarchy.
The benchmark progress and self-training capability are tangible. The fragmented frontier and numbering drift suggest OpenAI is entering a phase where raw version superiority matters less than task-specific fit and cost-performance trade-offs.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.