- Published on
Just In: GPT-6 New Model Shakes the Table! Crashes into DeepSeek's Territory at Rock-Bottom Prices
- Authors
- Name
- DeepHorus
OpenAI has officially released two new models, GPT-6 Sol and GPT-6 Luna, with API pricing slashed to roughly half that of their predecessors, directly challenging DeepSeek, which is known for its low prices. Among them, GPT-6 Luna is priced at $0.1 per million tokens for input and $0.5 for output, equivalent to approximately ¥0.7 and ¥3.5 in Chinese yuan. By comparison, DeepSeek V4.1 Flash charges ¥1 per million tokens for cache-miss input during off-peak hours, rising to ¥2 and ¥8 during peak hours, meaning Luna has already achieved a price reversal over DeepSeek in ordinary new-request scenarios.
DeepSeek still maintains a significant advantage in cache-hit scenarios, where its off-peak cache-hit input costs only ¥0.02 per million tokens, far below Luna. In addition, V4.1 Flash features a 1-million-token context window, supports image and text understanding, tool calling, and thinking mode, giving it clearly superior overall capabilities. However, in ordinary requests with cache misses, Luna has already been the first to drag DeepSeek into the same price range.
Earlier this month, OpenAI released its most capable model, GPT-6 Astra, targeted at difficult tasks, priced at $10 per million tokens for input and $50 for output. Sol and Luna inherit Astra's training methodology but are aimed at everyday tasks with faster speeds and lower prices. GPT-6 Sol's pricing has been reduced from the previous generation's $4 and $20 to $2 and $10, with OpenAI CEO Sam Altman emphasizing that the metric that truly matters is the cost of completing tasks.
Official tests show that GPT-6 Sol achieved a score of 33.2% on AutomationBench, higher than Claude Opus 5's 26.9%, with a per-task cost of only 9% of the latter's. In the Agents' Last Exam benchmark, Sol reached 56.4%, with per-task costs approximately 60% lower. In the software engineering benchmark DeepSWE v1.1, Sol scored 68.8%, close to Claude Fable 5's 69.9%, at roughly 80% lower cost.
Developer Flavio Adamo said after trying it that tasks that previously took an hour can now be completed in about 20 minutes, with token consumption less than half of before. However, some evaluations have noted that in terms of visual presentation, Claude Opus 5.5's completeness and design taste are still clearly superior to GPT-6 Sol. Artificial Analysis pointed out that the overall intelligence level of the two new models is roughly on par with their predecessors, and the truly significant change comes from the cost reduction.
This wave of price cuts marks a shift in OpenAI's competitive strategy — extending from competing for the highest scores on leaderboards to how much work can be completed per dollar. GPT-6 Sol and Luna are now available starting today on ChatGPT Work and Codex for Plus, Pro, Business, and other users. The continued evolution of low-priced model capabilities could reshape the pricing landscape of the entire AI model market.
