What? GPT is cheaper than DeepSeek? The price war for large-scale models has reached new heights.

Less than three weeks after the release of its flagship model, OpenAI has made another move, slashing the API prices of two new models in half, aiming at DeepSeek's low-price moat. The competition for large models has shifted from "comparing capabilities" to "comparing costs".

On September 22nd, Eastern Time, OpenAI officially launched two new models, GPT-6 Sol and GPT-6 Luna, announcing a further 50% reduction in their API prices compared to the promotional pricing of GPT-5.6. GPT-6 Luna's input price is as low as $0.10 per million tokens, and its output price is $0.50, directly entering the price range previously considered advantageous by DeepSeek, launching a direct attack on the core territory of its Chinese competitors.

This price reduction is not simply a business concession. OpenAI stated that the price decrease stems from improvements in caching and inference efficiency, with the company directly passing on the cost savings to users. Simultaneously, Anthropic launched Claude Opus 5.5 on the same day, emphasizing lower operating costs. The entry of these two leading AI companies into the market on the same day with more competitive pricing signifies that competition in the AI industry has entered a new phase where performance, speed, and cost are all advancing in tandem.

Less than three weeks after Astra's release, OpenAI quickly completed its product line.

On September 3, OpenAI launched Astra, its flagship GPT-6 model, claiming it reached new frontiers in areas such as computer operation, software engineering, cybersecurity, and professional work. At the time, the overwhelming demand for Astra forced OpenAI to temporarily suspend new Pro subscription registrations.

In less than three weeks, OpenAI has launched Sol and Luna under Astra, rapidly expanding its product line to different performance levels, costs, and use cases. OpenAI stated that Sol and Luna "build on the technological advancements behind GPT-6 Astra," bringing Astra's many advantages to "faster, more affordable models" to support large-scale operations.

This pace aligns with the increasingly evident product strategy in the AI industry: flagship models are responsible for pushing the limits of capability, while faster, cheaper versions are responsible for entering more frequent workflows.

API prices have been halved, and Luna has entered the heart of DeepSeek's price territory.

Price is the clearest business signal from this release.

According to OpenAI's pricing list, the input price of GPT-6 Sol has been reduced from $4 (promotional price of GPT-5.6 Sol) to $2 per million tokens, and the output price has been reduced from $20 to $10; the input price of GPT-6 Luna has been reduced from $0.20 to $0.10, and the output price has been reduced from $1.20 to $0.50. Both models' API costs have been reduced by half.

OpenAI explained that the price reduction primarily stemmed from improvements in caching and inference efficiency, rather than simple commercial subsidies. The company also further optimized its prompt caching mechanism, enabling agents and long dialogues to reuse processed context more effectively, thereby reducing the cost per call.

Luna's pricing is particularly aggressive, having entered the ultra-low price range previously dominated by DeepSeek's models, leading outsiders to view this release as a direct challenge to DeepSeek's price moat by OpenAI.

In comparison, DeepSeek V4.1 Flash has an idle cache miss input price of 1 yuan per million tokens and an output price of 4 yuan, which rises to 2 yuan and 8 yuan respectively during peak hours. Based on normal new request calculations, Luna is already cheaper and does not experience cost fluctuations due to peak-valley pricing.

However, DeepSeek still has its own price advantage. Its idle-time cache hit input costs only 0.02 yuan per million tokens, far lower than Luna. V4.1 Flash also has significantly higher overall capabilities, with a 1 million token context and support for image and text understanding, tool calls, and thinking modes.

Sol focuses on complex tasks, while Luna targets high-frequency, low-cost scenarios.

The two models are positioned differently and are not "the same model with different prices".

Sol is geared towards tasks requiring higher capabilities, including complex professional work, coding, and agent scenarios. In AutomationBench's enterprise cross-application workflow test, GPT-6 Sol scored 33.2% at xhigh effort, higher than GPT-6 Astra's 30.3% at low effort, and also higher than Claude Opus 5's 26.9% at max effort; OpenAI states that Sol costs only $0.27 to complete each task.

Luna, on the other hand, emphasizes cost efficiency. OpenAI states that in the high-intensity AutomationBench test, GPT-6 Luna improved by 5.4 percentage points compared to its predecessor, while reducing the cost per task by 58%. In DeepSWE v1.1, Luna achieved a maximum score of 66.6%, comparable to the medium-intensity performance of Claude Opus 5 and Fable 5, but with a cost per task that was 93% and 96% lower, respectively.

Regarding factual accuracy, OpenAI's internal testing shows that GPT-6 Sol has approximately half the number of errors as its predecessor; Luna achieves the level of GPT-5.6 Sol at higher inference strengths, at a cost of about one percent of the latter. OpenAI notes that this test was based on anonymous ChatGPT conversations previously marked as factual errors by users and does not represent all real-world use cases.

Advanced Work and Codex, free users can experience Luna

In terms of product coverage, OpenAI did not immediately integrate Sol and Luna into the ChatGPT regular chat interface.

According to the plan, Plus, Pro, Business, Enterprise, and Edu users can use both models in ChatGPT Work and Codex, and access the API; free and Go users can experience GPT-6 Luna in the desktop application.

This arrangement clearly prioritizes initial application scenarios for work and development rather than mass consumer use. Sol and Luna's low cost and high efficiency also make them more suitable for coding tasks requiring frequent model calls, agent workflows, and enterprise automation. At the same time, directing some high-frequency demands to lower-cost models may help OpenAI alleviate infrastructure pressure while expanding its adoption.

Industry competition has shifted from "competing on flagship models" to "competing on scale and cost."

It is worth noting that as OpenAI launched Sol and Luna, the industry was simultaneously discussing whether AI development should be slowed down.

Earlier this month, Anthropic CEO Dario Amodei publicly called for slowing down the pace of cutting-edge AI development and warned of the risks that rapid advancement could bring. OpenAI CEO Sam Altman subsequently agreed that the industry needs to slow down the pace of cutting-edge model development and strengthen security measures. OpenAI Chief Scientist Jakub Pachocki also suggested that coordinating a slowdown in future AI development is an important way to ensure the safety of self-improving AI.

However, OpenAI's own product iteration has not slowed down: Astra launched two new models less than three weeks after its initial release, and directly reduced the API price by 50%. On the same day, Anthropic also launched Claude Opus 5.5, also highlighting lower operating costs as its selling point.

While discussing how to slow down the pace of AI development, the AI industry is simultaneously accelerating model iteration and reducing the cost of calls, entering a new phase where both capabilities and costs are rapidly decreasing.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.