Anthropic and OpenAI launched new models targeting professional uses, combining improved performance with lower model operating costs. This comes at a time when economic efficiency and the reliability of models for coding and enterprise tasks are becoming increasingly important, alongside the ongoing debate over the safety of advanced models.
Opus 5.5 Focuses on Complex Tasks
Anthropic unveiled the Opus 5.5 model, which it says improves the handling of complex tasks, software bug detection and repair, financial data analysis, and business task execution. The company’s tests indicate that the model outperforms GPT-6 Astra on agentic coding tasks in the Terminal-Bench 4.0 and FrontierCode v1.1 (Main) benchmarks.
Anthropic says the model produces clearer text and is 85% less likely to attempt to circumvent boundaries and safeguards compared with previous models. The company also reduced the price of one million input tokens to $4 and one million output tokens to $20, compared with $5 and $25, respectively, for the Opus 5 model.
Three Different Cost Levels at OpenAI
OpenAI launched the GPT-6 Sol and GPT-6 Luna models, with improvements in professional work, coding, and computer use. The company says the cost of the two models is up to 50% lower than the promotional prices of the GPT-5.6 models.
The price of one million input tokens in GPT-6 Sol is $2, compared with $10 for output. GPT-6 Luna costs 10 cents per million input tokens and 50 cents per million output tokens. According to OpenAI, GPT-6 Sol makes approximately half as many errors as the previous model, while achieving a level of coding performance comparable to Fable 5.1 at a lower cost.
The improvements also include greater clarity in communication and support for temporarily storing prompt context, allowing more context to be reused and responses to be accelerated.
What Changes in Practice?
The new prices give developers and companies greater scope to use models for repetitive or invocation-intensive tasks, while the more capable models target coding, analysis, and agentic execution. However, the figures cited regarding superiority, accuracy, and safety come from the companies’ own tests and therefore do not by themselves establish general superiority across all uses or environments.
Safety and Availability Improvements
OpenAI says GPT-6 models are less likely to provide incorrect information about the results of coding tasks and are more likely to reject attempts to bypass protective procedures when receiving unsafe commands. The two companies are working to propose standards that external evaluators can use to measure model progress and safety, but it is not yet clear whether these proposals will become a unified standard.
Opus 5.5 is available through Claude, as well as Amazon Web Services, Google Cloud, and Microsoft Azure. GPT-6 Sol and GPT-6 Luna are available through ChatGPT Work and Codex for customers on Plus, Pro, Business, and Enterprise plans, while free-plan users and Go subscribers receive GPT-6 Luna only through the company’s desktop application.