Free GPT-5.6 Luna vs DeepSeek, MiniMax, and ByteDance: The Price War Intensifies

At the end of July, OpenAI slashed 80% off the API prices of its consumer model GPT-5.6 Luna. New price: $0.20 per million input tokens, $1.20 for output. Three weeks later, the logical consequence follows: Luna becomes the default free ChatGPT model for "Free" and "Go" users. Unlimited text chats are on the menu.

This isn't just another price drop. It's a strategic shift that reconfigures the entire "cheap models" layer of the market—where DeepSeek, MiniMax, and ByteDance have dominated the conversation so far with their open weights and aggressive pricing.

Luna's New Positioning

With $0.20/M input, Luna slides into the lower tier of the market, populated by small Chinese open-weight models and Google's frugal offerings (Gemini 3.5 Flash at $0.15/M) and Xiaomi. The shift is clear:

• $0.20 vs $0.435 for DeepSeek V4 Pro (with its permanent 75% promotion): Luna takes the lead on input
• $1.20 vs $0.87 on output: DeepSeek remains cheaper on output

The "build vs buy" calculation for many enterprise uses—ticket sorting, first draft generation, simple classification—shifts. Where self-hosting a DeepSeek open-weight had a real economic advantage, it's now worth a second look.

A month ago, no proprietary model from a major American lab was competitive against DeepSeek. Today, Luna is—and it comes with OpenAI's infrastructure, ecosystem, and SLA.

The Chinese Response: Almost Instantaneous

DeepSeek responded the same day with the launch of V4 Flash, a 284 billion parameter MoE variant (13 billion active per token), aggressively positioned in terms of price-to-quality ratio. MiniMax and ByteDance Seedance 2.5 hold their ground in the video and multimodal segments, with prices 5 to 7× lower than Western equivalents.

According to several analysts cited by MLQ News and VentureBeat, we're entering a prolonged price war for the second half of 2026. China bets on openness (downloadable weights, permissive licensing), OpenAI bets on vertical integration and massive distribution via ChatGPT.

Distribution as a Weapon

Making Luna free in ChatGPT is a move deeper than a mere marketing operation. OpenAI thus captures usage data of its model on hundreds of millions of new daily requests—something no Chinese model hosted by a third party can do.

Two expected effects:

1. Continuous implicit fine-tuning. Each interaction with Luna (feedback like/dislike, rephrasing, abandonment) enriches OpenAI's RLHF corpora. The model improves even without a new version.
2. Ecosystem lock-in. A user accustomed to using Luna for free via ChatGPT won't switch to DeepSeek for personal use. The B2C battle is about experience, not benchmarks.

On Code Benchmarks: China Still Holds

The DeepSWE comparison conducted by Together.ai shows: on long-context code tasks, DeepSeek V4 Flash still leads Luna at a comparable cost. Translation: on sophisticated developer workloads, the Chinese ecosystem hasn't lost.

It's rather on general public and generalist B2B uses (support, content, classification, extraction, summarization) that Luna becomes unbeatable. Market segmentation intensifies.

Winners and Losers

• OpenAI gains in user base—clear objective: consolidate a billion active users and prevent DeepSeek from encroaching on the general public.
• Anthropic isn't in this war and acknowledges it: Claude Sonnet 5 remains at $3 / $15 per million, positioned as premium quality for enterprise uses.
• Google is trapped: Gemini 3.5 Flash is already aggressive but the delay of Gemini 3.5 Pro prevents it from counterattacking in the high-end segment.
• Cloud providers (Together.ai, Fireworks, DeepInfra) see their "host an open-weight model for you" proposition crumble on large volumes.

Key Takeaways

There are now three well-defined fronts:

• Free general public—dominated by ChatGPT (Luna) and Gemini, with Claude as a premium outsider
• Dev / code / agent workloads—where DeepSeek and Claude maintain the qualitative edge
• Public video and multimodal—dominated by Chinese models (Seedance, MiniMax H3) due to the price differential

One thing is certain: the unit margin of the AI industry in 2026 is collapsing. The economic winners won't be those who make the best model, but those who know how to distribute massively and integrate deeply. OpenAI is playing this card to the fullest.