Summary
OpenAI has introduced three new models under the GPT-5.6 family: Soul (flagship), Terra (balanced everyday use), and Luna (fast, high-volume tasks) — a naming shift that maps roughly to Anthropic’s Opus/Sonnet/Haiku tier structure. TheAIGRID breaks down what each tier is designed for and highlights a notable addition: GPT-5.6 Soul Ultra, a mode that spins up sub-agents to parallelize complex long-horizon work rather than relying on a single inference pass.
On benchmarks, GPT-5.6 Soul Ultra edges out Claude Mythos 5 on Terminal Bench — Stanford’s evaluation of whether AI agents can operate inside a real command line end-to-end — and substantially outperforms Claude Fable 5. On Exploit Bench, which tests whether agents can turn known vulnerabilities into working exploits, the model family clusters just below Mythos 5 while reportedly costing around 40% less than comparable Anthropic offerings. A standout data point: on long autonomous software research tasks, the model’s performance range spans 13 hours to 11,400 hours of work, with a median estimate of 71 hours — uncertainty so large the error bars break the chart, suggesting existing benchmarks can no longer contain it.
The video also covers the restricted rollout under Trump administration approval requirements and a planned Cerebras Inference integration targeting 750 tokens per second for Soul by July 2026 — a throughput figure the host argues signals a new phase of AI deployment speed.
📺 Source: TheAIGRID · Published June 27, 2026
🏷️ Format: News Analysis
