GPT-5.6 Crushes Claude Fable 5 in TerminalBench – What This Means for AI

OpenAI has just released GPT-5.6, and its TerminalBench 2.1 results put it ahead of Anthropic’s Claude models: Sol Ultra scores 91.9%, Sol 88.8%, Claude Mythos 5 88.0%, while both Claude Fable 5 and GPT‑5.6 Terra reach 84.3%.

TerminalBench evaluates a model’s ability to operate in a terminal environment—reading the environment, selecting commands, executing them, handling failures, and iterating—rather than answering questions or solving single‑file coding tasks, making it a key indicator of AI programming capability.

Claude Fable 5, launched on June 9 and paused three days later, serves as a dramatic contrast; its brief availability highlights the competitive pressure of OpenAI’s new release.

The GPT‑5.6 family is organized into three tiers:

Sol : flagship model for the most difficult, long‑running, and complex tasks.

Terra : balanced tier that delivers GPT‑5.5‑level performance for everyday work.

Luna : cheap, fast tier aimed at high‑frequency calls and bulk workloads.

Pricing reflects this hierarchy:

Sol – $5 per M input tokens, $30 per M output tokens.

Terra – $2.5 per M input tokens, $15 per M output tokens.

Luna – $1 per M input tokens, $6 per M output tokens.

Sam Altman notes that Terra matches GPT‑5.5 performance at half the price, while Sol costs the same as GPT‑5.5 but is smarter and more efficient, indicating OpenAI’s strategy to cover multiple price points rather than showcasing a single flagship.

Despite the strong scores, GPT‑5.6 is only available via limited preview, not the open access originally planned. OpenAI classifies Sol, Terra, and Luna as “High capability” for network‑security and biochemistry domains, capable of vulnerability discovery but not yet able to autonomously execute full‑scale attack chains.

On ExploitBench, Sol’s curve rises sharply, achieving roughly one‑third of the output tokens needed to reach the performance of Anthropic’s Mythos Preview, demonstrating competence beyond simple code generation and into security‑focused reasoning.

OpenAI also disclosed that more than 700 000 A100‑equivalent GPU hours have been spent on automated red‑team testing to hunt for universal jailbreak methods, underscoring a rigorous “factory stress test” for high‑capability models.

Three key observations emerge:

The AI‑programming battlefield has been reclaimed, with GPT‑5.6 overtaking Claude Fable 5 on the same benchmark.

Model products are entering a tiered era, prompting users to choose Sol, Terra, or Luna based on task demands, cost, and reliability.

Release processes for powerful models are becoming infrastructure‑like: staged rollouts, trusted‑access controls, continuous safety monitoring, and extensive red‑team testing.

Thus, GPT‑5.6 combines headline‑grabbing performance with a cautious, layered deployment strategy, reflecting how future high‑capability AI releases may look.