Anthropic Opus 5.5 promises lower costs and improved efficiency

Anthropic says Opus 5.5 delivers lower token costs while also using fewer tokens to complete tasks. With default settings and typical workloads, the model could reduce costs by nearly 40% compared with Opus 5.

Opus 5.5 is also designed to handle demanding tasks in what Anthropic describes as “high-risk areas,” including cybersecurity and biology. As a result, the protections used for Fable 5.1 also apply to the new model.

Requests that are flagged as entering protected areas can be automatically and transparently routed to the older model.

GPT-6 Sol and Luna focus on affordability and everyday use

OpenAI’s release of GPT-6 Sol and Luna represents an iterative update to the company’s GPT-6 model family. OpenAI previously introduced the Sol, Terra, and Luna naming convention for the GPT-5.6 series and recently announced GPT-6 Astra, its most advanced and powerful model.

The names describe different positions within OpenAI’s model lineup:

• Astra: The most powerful and expensive model, intended for demanding coding, research, and other intensive tasks.
• Sol: A capable, more efficient, and more affordable model for everyday work involving tasks such as coding and research.
• Terra: A balanced and versatile model.
• Luna: A fast and inexpensive option.

There are rough parallels between the two companies’ model families. Astra could be compared with Anthropic’s Fable, Sol with Opus, Terra with Sonnet, and Luna with Haiku. However, these comparisons are imperfect, particularly for the higher-end models.

According to OpenAI’s announcement, GPT-6 Sol and Luna were trained using methods similar to those used for GPT-6 Astra. Depending on the benchmark, the new models may be several percent more capable than their predecessors while costing half as much to use.

Opus 5.5 vs. GPT-6 Sol and Luna

Anthropic’s Opus 5.5 emphasizes lower token usage, cost savings, and safeguards for high-risk applications. OpenAI’s GPT-6 Sol and Luna, meanwhile, are positioned as more affordable alternatives within the GPT-6 lineup, with performance improvements over their predecessors at a lower usage cost.