A Model Explosion: GPT 5.6 Sol, Grok 4.5 and Meta Muse Rewrite the Rules

2026-07-10 · Source: AI Explained · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Advanced, long

Summary

OpenAI recently released three new models: GPT-5.6 Soul, Terror, and Luna, which demonstrate strong performance across various benchmarks, often at a third of the cost of Anthropic's Claude series. GPT-5.6 Soul scored 54% on Agent's Last Exam, surpassing Fable's 45%, and achieved 80 on the Artificial Analysis Coding Index compared to Fable's 77. While Grok 4.5 leads on the Suey Marathon benchmark, and Meta's Muse Spark 1.1 offers competitive "vibe coding" performance (72% vs Soul's 81%) at 35 times less cost, the market is seeing a "model explosion" with cost-efficient alternatives. A significant concern emerged regarding GPT-5.6 Soul's security, as the UK AI Security Institute found it easier to jailbreak than Fable, even with universal jailbreaks. OpenAI also introduced a real-time voice agent and claims internal self-improvement, though these are hard to verify. The author concludes that model improvement, with potential for quadrillion-parameter models, is far from over.

Key takeaway

For Machine Learning Engineers evaluating new frontier models, you should prioritize performance-per-dollar metrics over raw benchmark scores, as models like GPT-5.6 Soul, Grok 4.5, and Meta Muse Spark offer near-frontier capabilities at significantly reduced costs. Be aware of potential security vulnerabilities, such as the jailbreaking ease found in GPT-5.6 Soul, and factor this into your risk assessment. Explore specialized benchmarks like Agent's Last Exam for real-world economic value and consider how these cost-efficient options can accelerate your project timelines.

Key insights

Cost-efficient frontier models like GPT-5.6 Soul, Grok 4.5, and Meta Muse Spark are rewriting AI performance-to-price rules.

Principles

• Performance-per-dollar is a critical metric.
• Benchmarks like Agent's Last Exam reflect economic value.
• Verifiable domains are susceptible to AI "crushing" tasks.

In practice

• Evaluate models on performance-per-dollar.
• Consider cheaper alternatives like Grok 4.5 or Muse Spark.
• Prioritize verifiable domains for AI automation.

Topics

• Large Language Models
• AI Benchmarking
• Model Cost Efficiency
• GPT-5.6 Soul
• Grok 4.5
• Meta Muse Spark
• AI Security

Best for: CTO, VP of Engineering/Data, AI Architect, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• I Tested GPT-5.6 Sol, Claude Fable 5, and Grok 4.5 for a Week — Here’s the Real Winner — Top-tier LLMs now specialize, balancing raw intelligence with cost and speed for diverse applications.
• [AINews] not much happened today — AI model stratification and agentic orchestration are advancing rapidly, but usability and cost transparency remain critical challenges.
• Oh no (the new Grok model is good) — Grok 4.5 offers near-frontier intelligence for coding and agentic tasks at a significantly lower cost and higher token efficiency.
• 🤖 AI Agents Weekly: GPT-5.6 Family, Meta Muse Spark 1.1, Grok 4.5, SWE-1.7, Robostral Navigate, The Harness Effect, and More — New AI models are increasingly specialized for agentic tasks, offering tiered capabilities and long context windows.
• GPT-5.6 IS OUT! GLM 5.5 Is Mythos Level, U.S Governement Banning AI Cause of Dario?, & Grok 4.5! — Frontier AI models like GPT 5.6 and Grok 4.5 are rapidly advancing in coding and cybersecurity, but face significant national security scrutiny and access limitations.
• Grok 4.5 is so cheap compared to Fable 5 and GPT 5.5 that benchmark gaps may not matter much — Grok 4.5 offers competitive performance at a significantly lower cost, challenging higher-priced LLMs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Explained.