Daniel Mercer
The current discourse in the AI sector has predominantly revolved around the idea of moderating AI advancements or “pacing the frontier.” However, this focus has not deterred Anthropic and OpenAI from unveiling their latest models. Both organizations are refining their previous iterations, Fable 5.1 and GPT-6 Astra, presenting new models that maintain comparable performance levels while reducing costs.
Anthropic has introduced its Opus 5.5 model, which excels in managing complex tasks such as “identifying and rectifying inefficiencies in software” alongside financial analysis and business operations—all aimed specifically at Anthropic’s enterprise clientele. The company’s benchmarks suggest that Opus 5.5 has outperformed GPT-6 Astra in agentic coding, showcasing superior results on both Terminal-Bench 4.0 and FrontierCode v1.1 (Main), making it a more attractive option for developers.
The enhancements are not limited to the types of tasks the model can perform but also extend to its operational efficiency. Anthropic asserts that Opus 5.5 generates text that is more comprehensible and has “attempted to circumvent boundaries around 85 percent less often” than earlier versions, indicating a reduced likelihood of disregarding commands. the cost per token has decreased significantly, with input tokens priced at $4 and output tokens at $20, compared to Opus 5’s $5 for input and $25 for output tokens.
OpenAI’s latest offerings, the GPT-6 Sol and GPT-6 Luna models, are similarly oriented towards enhancing efficiency. These new models were developed using techniques akin to those employed for GPT-6 Astra and deliver improvements in professional applications, coding, and computer tasks, all at costs that can be up to 50 percent lower than the introductory prices of their GPT-5.6 predecessors. Specifically, GPT-6 Sol’s input tokens are priced at $2, while output tokens cost $10. For Luna, input tokens are just $0.10, with output tokens at $0.50.
In terms of advancements, both GPT-6 models significantly reduce the frequency of factual errors, with Sol reportedly making about half as many mistakes as its forerunner, according to OpenAI. In coding tasks, GPT-6 Sol is said to achieve performance levels on par with Fable 5.1 while being more cost-effective. The new models also exhibit clearer communication, and improvements in prompt caching allow OpenAI’s models to leverage existing context more efficiently, resulting in faster responses.
OpenAI claims that the GPT-6 models demonstrate enhanced alignment based on the company’s proprietary evaluations. For instance, they reportedly lie less about their coding outputs and show a greater tendency to resist attempts to bypass safety protocols when confronted with unsafe commands. Similar to Anthropic, OpenAI has put forth its own set of criteria for external evaluators to assess the progress and safety of AI development. Whether these proposals will establish a standardized practice across the industry remains uncertain.
Developers can access Anthropic’s Opus 5.5 model through platforms such as Claude, Amazon Web Services, Google Cloud, and Microsoft Azure. OpenAI’s GPT-6 Sol and GPT-6 Luna are available within ChatGPT Work and Codex for Plus, Pro, Business, and Enterprise users. Free users and Go subscribers will have access solely to GPT-6 Luna in the company’s desktop application.
Ethan Carter
