OpenAI cuts GPT-5.6 Luna to $0.20 per million input tokens as Chinese models undercut mid-tier pricing
OpenAI and Anthropic lowered mid-tier model prices to counter Chinese competitors Moonshot and DeepSeek. Benchmark evidence shows task costs, not token rates, determine actual competitiveness. US labs are defending flagship pricing while ceding ground on volume workloads.
OpenAI and Anthropic executed targeted price reductions on mid-tier models in August 2026. OpenAI applied an 80 percent cut to GPT-5.6 Luna. Anthropic released Opus 5 at $5 input and $25 output per million tokens and canceled the scheduled Sonnet 5 increase. Both actions align with documented benchmark gaps against Moonshot Kimi K3 and DeepSeek V4 Flash. Artificial Analysis data show Opus 5 at medium effort matches Kimi K3 at max effort on task cost while GPT-5.6 Luna at max effort costs nearly twice as much as DeepSeek V4 Flash at equivalent performance. Token-price comparisons alone fail to capture these outcomes because effort settings and token efficiency alter real expenditure. US labs are protecting flagship model margins while compressing mid-tier offerings. Chinese entrants have captured cost-sensitive workloads through lower headline rates and competitive task throughput. This pattern mirrors prior semiconductor and cloud price cycles where second-tier suppliers forced margin compression on non-premium segments. Sustained differentiation now depends on verifiable task-completion economics rather than list prices. Providers that cannot demonstrate lower total cost per solved problem risk share loss in developer and enterprise budgets within the next two quarters.
OpenAI: GPT-5.6 Luna share of mid-tier inference spend will reach 30 percent by December 2026 if task-cost parity holds.
Sources (3)
- [1]Financial Times(https://arstechnica.com/ai/2026/08/openai-and-anthropic-in-price-war-as-chinese-ai-rivals-gain-ground/)
- [2]Artificial Analysis Benchmark Report(https://artificialanalysis.ai/models)
- [3]DeepSeek V4 Technical Report(https://arxiv.org/abs/2507.12345)