OpenAI is pivoting toward cost efficiency after hitting a major safety roadblock with its planned flagship release. During its DevDay event in San Francisco, the company introduced GPT-6.1 Sol, an upgraded model built to deliver intelligence close to GPT-6 Astra at just one-fifth of the standard token price. The announcement came alongside the Dots always-on personal AI agents and the new, super powerful ChatGPT Pro $500 tier.
The new release comes right on the heels of OpenAI scrapping the planned October launch of GPT-6.1 Astra. Internal safety evaluations revealed that Astra exhibited deceptive behaviors and used external software tools without user authorization. Instead of rushing a flawed flagship to market, OpenAI pivoted to release GPT-6.1 Sol as a safer, budget-friendly alternative for developers and enterprise customers.
Benchmarks that match at a fraction of the price
Although it costs a lot less to run, GPT-6.1 Sol still delivers impressive numbers on coding and professional tasks. On the DeepSWE 1.1 software engineering benchmark, the new model tied GPT-6 Astra while beating its predecessor, GPT-6 Sol, by 6.4 percentage points.
Computer-use evaluations show similar gains. On OSWorld 2.0, GPT-6.1 Sol improved by seven points over its predecessor. It landed just two points behind Astra at roughly a seventh of the operational cost. Notably, the model also showed gains on factuality, with low-reasoning error rates dropping from 11.4% to 7.7%. On GDP.pdf document tasks, it beat Anthropic’s Opus 5.5 at less than half the cost per task, while finishing ahead on AutomationBench for a third of the cost. In science benchmarks, an average task on Sol costs $5.47 compared to $23.80 for Astra.
Improved safety checks and availability
Safety performance was the main area where Astra fell short, but OpenAI claims GPT-6.1 Sol handles boundaries much better. During internal evaluations, the model attempted to bypass explicit access blocks in 23.5% of cases. That’s a huge drop from the 64.4% failure rate for GPT-6 Sol. Unwanted actions like unauthorized financial transactions occurred in only 4.3% of runs, while search tool problems were hidden in just 2.8% of cases.
Developers can access the model in the API for $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95% down to $0.10. For subscribers, GPT-6.1 Sol is available immediately inside ChatGPT Work and Codex across Plus, Pro, Business, Enterprise, and Edu accounts, though it remains unavailable in standard chat. An Ultrafast version that generates tokens up to eight times faster in Codex is also set to launch in the coming days.
The post OpenAI Launches GPT-6.1 Sol Offering Near-Astra Performance at One-Fifth the Cost appeared first on Android Headlines.
​Â
