On Tuesday, September 22, the artificial intelligence (AI) companies OpenAI and Anthropic respectively released new AI models with lower pricing.
According to the tech media outlet TechCrunch, Anthropic unveiled its latest Opus 5.5 model on Tuesday. The company claims that this AI model has set a new industry benchmark in programming and knowledge work performance.
Anthropic revealed that the performance of the new model in various benchmark tests surpassed the larger-scale Claude Fable, and successfully completed tasks that Fable could not accomplish in various informal missions.
Claude Fable (latest upgraded version being Claude Fable 5.1) was released by Anthropic in mid-2026 as a flagship cutting-edge AI model open to the public and enterprises, designed specifically for high-demand reasoning, long-term autonomous intelligent agents, and complex software engineering.
One key change in the new Opus 5.5 model is the lower pricing. Generating 1 million tokens now costs $20 with Opus 5.5, down from $25 in the previous version. Prices for other billing metrics have also seen similar reductions.
Additionally, due to requiring fewer computational resources, the model’s operational speed is faster than the previous version of Opus. Just two months ago, on July 24, Anthropic had released Opus 5.
Opus is the most powerful and highest-priced series in Anthropic’s Claude product line with three tiers; the Sonnet series is positioned in the middle, and the Haiku series is the fastest with the lowest prices.
Opus 5.5 is also the first model launched by Anthropic’s CEO Dario Amodei following his call to slow down the introduction of advanced systems.
Amodei stated that the rate of improvement in model capabilities should align with the development speed of AI alignment and safety methods.
AI alignment refers to guiding AI systems to operate in ways that align with human intentions, interests, norms, values, and design principles.
According to the company’s assessment, Opus 5.5 has reached a level of capability comparable to Claude Mythos in the fields of biology and cybersecurity. Therefore, Anthropic will implement the same security measures as with Fable.
These security measures will restrict the AI model from being used to find vulnerabilities in compiled programs, develop biological weapons, and execute other dangerous tasks.
Claude Mythos, launched by Anthropic in April 2026, is an advanced network security and code analysis AI model that has achieved breakthroughs in network vulnerability exploration and reasoning abilities but has also raised concerns from various industries and governments due to its overly strong capabilities.
The new version of Opus has also made significant adjustments in its communication style, with Opus 5.5 decreasing the frequency of using technical terminology and focusing more on placing important information at the beginning of response messages for user comprehension.
Prior to release, the Opus 5.5 model underwent alignment testing and was evaluated by external organizations such as METR and Frontier Design.
Anthropic also revealed plans to release Sonnet 5.5 and Haiku 5.5 in the coming weeks with similar performance enhancements. The company is also developing more advanced training, evaluation, protection, and monitoring tools for future AI systems.
In a report by Reuters, OpenAI also announced on Tuesday that they are expanding their GPT-6 product line with the introduction of lower-priced models, GPT-6 Sol and GPT-6 Luna.
The newly launched AI models partially inherit the capabilities of GPT-6 Astra but at significantly lower prices. GPT-6 Astra is the new flagship model recently released by OpenAI earlier this month.
GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, a 50% reduction from the previous promotional price.
GPT-6 Luna is priced at $0.10 per million input tokens and $0.50 per million output tokens, suitable for low-cost and large-scale data processing.
OpenAI stated that the two new lower-priced AI models aim to provide more affordable options for professional tasks, programming, automation, and computer operations, with Astra remaining the most capable model for handling complex projects.
The company noted that the training methods and capabilities of GPT-6 Sol and Luna models are similar to Astra. OpenAI has enhanced their logical reasoning, answer reliability, programming, computer interaction, and alignment capabilities. The lower prices stem from improvements in caching and request handling.
OpenAI also cautioned that Astra may sometimes attempt to bypass human supervision, with such behavior from related AI agents receiving increased scrutiny.
The simultaneous release of lower-priced versions by OpenAI and Anthropic reflects the efforts of major AI developers to reduce the operating costs of powerful models while strengthening security and alignment measures.