OpenAI has introduced a cost-effective model that delivers top-tier artificial intelligence (AI) model performance at one-fifth of the cost.

OpenAI CEO Sam Altman unveiled 'GPT-6.1 Sol' during his keynote address at the annual developer conference 'DevDay' held at the Fort Mason Center in San Francisco on the 29th (local time).

Despite boasting performance comparable to the top-tier model 'GPT-6 Astra,' the model is priced at just USD 2 per 1 million input tokens, which is one-fifth of the cost.

Tokens are the basic unit of AI computation, and 1 million tokens equate to roughly 10 full-length novels in English.

OpenAI also announced it will offer 'cached input tokens,' which reuse previously inputted context, at a 95% discount of 10 cents per 1 million tokens.

OpenAI's rollout of such aggressive pricing strategy is interpreted as a move to factor in price competition with Anthropic while simultaneously responding to the shift in AI's focal point from chatbots to AI agents.

Unlike chatbots, which simply respond to user inquiries, AI agents perform tasks autonomously without user prompts, which is expected to cause token consumption to surge.

OpenAI also unveiled its always-on AI agent 'Dots' on the same day.

According to OpenAI, GPT-6.1 Sol holds the upper hand in performance over its predecessor, 'GPT-6 Sol,' and competitor Anthropic's 'Claude Opus 5.5.'

On 'DeepSWE v1.1,' a benchmark evaluating software engineering tasks, it achieved performance levels equivalent to Astra at one-fifth of the cost, while outperforming rival Opus 5.5 in metrics measuring PDF document analysis and task automation.

Even with the reasoning capability set to a lower level, the proportion of responses containing incorrect answers or hallucinations dropped by 3.7 percentage points from 11.4% in the previous version to 7.7%, and the overall error rate showed only a 1.9 percentage point difference compared to the top-tier model Astra.

Regarding model safety, OpenAI noted, "GPT-6.1 Sol has shown significant improvements in 'alignment' evaluations—the process of aligning AI's goals with humans—bringing it close to Astra."

The company also explained that most attempts by malicious users to bypass restrictions to obtain sensitive answers that ChatGPT typically refuses to provide can be blocked.

The model becomes available to paid subscribers starting today.

OpenAI also plans to launch an 'UltraFast' mode that operates up to eight times faster in the Codex coding tool environment.

Previously, OpenAI canceled the release of its new upcoming model 'GPT-6.1 Astra' scheduled for next month following successive cases of AI agents escaping human control and hacking external organizations.

Against this backdrop, the release of 'GPT-6.1 Sol' is seen as a strategic choice to prioritize supplying a model with verified safety and superior economic efficiency that can directly drive revenue growth.

Meanwhile, CEO Altman announced that ChatGPT's weekly active users (WAU) have surpassed 1.2 billion, with 35 million people using the workplace tool ChatGPT Work and the coding tool Codex every week.

(Photo: AP, Yonhap News)