Less than three weeks after launching its flagship model, OpenAI has moved again, slashing API prices for two new models by 50% in a direct challenge to DeepSeek's low-price moat. The competitive battle among large language models has shifted decisively from "capability" to "cost."

On September 22, Eastern Time, OpenAI officially released two new models, GPT-6 Sol and GPT-6 Luna, announcing API pricing that is 50% lower than the promotional rates for GPT-5.6. GPT-6 Luna's input price drops to just $0.10 per million tokens, with output at $0.50, placing it squarely in the ultra-low-price territory that DeepSeek has long claimed as its own advantage. This is a head-on assault on the Chinese competitor's core stronghold.

The price cut is far more than a simple commercial concession. OpenAI stated that the reduction stems from improved caching and inference efficiency, with the savings passed directly to users. On the same day, Anthropic launched Claude Opus 5.5, also highlighting lower operating costs. Two leading AI companies entering the market on the same day with more aggressive pricing signals a new phase of competition where performance, speed, and cost are advancing on all three fronts simultaneously.

The Product Line Expands Rapidly, Just Three Weeks After the Astra Launch

On September 3, OpenAI unveiled its flagship GPT-6 model, Astra, claiming new frontiers in computer operations, software engineering, cybersecurity, and professional work. Demand for Astra was so intense that OpenAI temporarily paused new Pro subscription registrations.

In under three weeks, OpenAI has rolled out Sol and Luna beneath Astra, quickly expanding its product line across different performance levels, costs, and use cases. OpenAI officially states that Sol and Luna "build on the technological advances behind GPT-6 Astra," bringing many of Astra's advantages into "faster, more affordable models" designed for scaled workloads.

This cadence aligns with a growing product strategy across the AI industry: flagship models push the limits of capability, while faster and cheaper versions are designed to penetrate high-frequency workflows.

API Prices Cut Directly in Half, with Luna Entering DeepSeek's Price Territory

Pricing is the clearest commercial signal of this release.

According to OpenAI's published price list, GPT-6 Sol's input price drops from the GPT-5.6 Sol promotional price of $4 to $2 per million tokens, while output falls from $20 to $10. For GPT-6 Luna, input price drops from $0.20 to $0.10, and output from $1.20 to $0.50. Both models' API costs are effectively cut by 50%.

OpenAI explains that the reductions come primarily from gains in caching and inference efficiency, not from simple subsidies. The company has also further optimized its prompt caching mechanism, allowing agents and long conversations to reuse more processed context, thereby lowering the per-call cost.

Luna's pricing is particularly aggressive, moving directly into the ultra-low-cost band that DeepSeek models have dominated, leading many to view this launch as a direct challenge to DeepSeek's price moat.

As a comparison, DeepSeek V4.1 Flash's off-peak cache-miss input price is RMB 1 per million tokens, with output at RMB 4, rising to RMB 2 and RMB 8 during peak hours. On a standard new request basis, Luna is already cheaper and avoids the cost volatility of peak/off-peak pricing.

But DeepSeek still holds its own price advantages. Its off-peak cache-hit input price is just RMB 0.02 per million tokens, far below Luna. V4.1 Flash also offers significantly higher overall capability, a 1 million token context window, and supports image understanding, tool calling, and reasoning modes.

Sol Targets Complex Workloads, Luna Aims at High-Frequency, Low-Cost Scenarios

The two models are positioned differently, not simply "the same model at different prices."

Sol is aimed at tasks requiring greater capability, including complex professional work, coding, and agent scenarios. On the AutomationBench enterprise cross-application workflow benchmark, GPT-6 Sol scored 33.2% under xhigh effort, outperforming GPT-6 Astra's 30.3% under low effort and Claude Opus 5's 26.9% under max effort. OpenAI reports that Sol's cost to complete each task is just $0.27.

Luna, meanwhile, emphasizes cost efficiency. OpenAI states that in AutomationBench high-intensity testing, GPT-6 Luna improved by 5.4 percentage points over its predecessor while cutting per-task costs by 58%. On DeepSWE v1.1, Luna scored up to 66.6%, comparable to Claude Opus 5 and Fable 5 at medium intensity, but with per-task costs 93% and 96% lower, respectively.

In terms of factual accuracy, OpenAI's internal testing shows GPT-6 Sol produces roughly half the number of errors compared to its predecessor. Luna, at higher reasoning intensity, can reach the level of GPT-5.6 Sol while costing about one-hundredth as much. OpenAI notes that these tests are based on anonymized ChatGPT conversations flagged by users for factual errors and do not represent all real-world scenarios.

Rollout Prioritizes Work and Codex, with Free Users Getting Access to Luna

In terms of product availability, OpenAI did not immediately push Sol and Luna into the standard ChatGPT chat interface.

Under the current arrangement, Plus, Pro, Business, Enterprise, and Edu users can use both models within ChatGPT Work and Codex, with API access also available. Free and Go tier users can experience GPT-6 Luna in the desktop application.

This deployment clearly favors work and development over mass consumer use. The low cost and high efficiency of Sol and Luna make them well-suited for coding tasks, agent workflows, and internal enterprise automation that require frequent model calls. At the same time, routing some high-frequency demand toward cheaper models could help OpenAI expand usage scale while alleviating infrastructure pressure.

Industry Competition Shifts from "Flagship Showdowns" to "Scale and Cost"

Notably, the launch of Sol and Luna comes amid ongoing industry debate over whether AI development should be slowed down.

Earlier this month, Anthropic CEO Dario Amodei publicly called for a slowdown in frontier AI development, warning that rapid progress could bring risks. OpenAI CEO Sam Altman subsequently expressed agreement that the industry needs to slow the pace of frontier model development and strengthen safety measures. OpenAI Chief Scientist Jakub Pachocki also suggested that coordinated deceleration of future AI development is a critical path to ensuring the safety of self-improving AI systems.

Yet, at the same time, OpenAI's own product iteration has not slowed down: just three weeks after Astra's release, it launched two new models and cut API prices by 50% outright. On the same day, Anthropic also released Claude Opus 5.5, similarly touting lower running costs.

While the industry discusses slowing down AI development, it is simultaneously accelerating model iterations and driving down call costs. The AI sector is entering a new phase where both capability and cost are rapidly descending at the same time.