The AI pricing war has entered a new phase. Throughout 2026, OpenAI and Anthropic have systematically cut prices on their mid-tier models, with reductions steep enough to prompt developers to reconsider whether open-source alternatives remain worth the hassle of self-hosting. Because per-token fees are the dominant recurring cost for teams running AI features in production, cuts of this size ripple directly into build-versus-buy decisions across the industry.

The numbers behind the price war

On July 30, OpenAI reduced prices on GPT-5.6 Luna by 80%, bringing costs down to $0.20 per million input tokens and $1.20 per million output tokens. The company's higher-end GPT-5.6 Terra also received a cut, dropping 20% to $2 for input and $12 for output per million tokens.

Anthropic followed on August 10, making the introductory pricing for Claude Sonnet 5 permanent at $2/$10 per million tokens.

A second wave arrived on September 22.AI launched GPT-6 Sol and GPT-6 Luna at $2/$10 and $0.10/$0.50 respectively, roughly 50% lower than previous-generation pricing. Anthropic countered with Claude Opus 5.5 at $4/$20 list pricing, though built-in efficiency improvements deliver around 40% in effective savings for users.

For readers new to these tables: API pricing is metered per million tokens — the word-and-character chunks models process — with input, the text a model receives, generally priced below output, the text it generates.

Why open source is feeling the squeeze

Self-hosting open-weight models carries no per-token fees, but it shifts costs into GPU hardware, engineering time, and round-the-clock operations. Every proprietary list-price cut narrows the gap self-hosting needs to justify that overhead.

Enterprises are increasingly splitting their workloads into tiers. Inexpensive proprietary models handle routine tasks such as summarization, classification, and basic customer interactions, while premium models are reserved for specialized workloads where performance differences genuinely matter.

Open source still commands significant volume share, particularly on platforms like OpenRouter, where developers mix and match models through a single integration.

The China factor

Chinese AI companies have been setting aggressively low price benchmarks that make even the reduced US pricing look expensive. DeepSeek V4 Flash offers pricing as low as roughly $0.14/$0.28 per million tokens, undercutting American counterparts by a wide margin. At high volumes, gaps like these compound quickly into materially different monthly bills for large workloads.

The performance gap is narrowing as well. Some Chinese models now approach US mid-tier capabilities at five to nine times lower costs.

IPO math and margin pressure

Both OpenAI and Anthropic are widely expected to pursue IPOs. Cutting prices aggressively boosts adoption and usage metrics, but it also compresses margins. The bet is that lower prices drive higher volume, and that higher volume eventually compensates for the thinner margins on each individual API call.

That tradeoff carries extra weight ahead of a public listing, where investors typically scrutinize growth and profitability side by side. As private companies, neither firm publishes audited financials today, so their eventual IPO filings would give the industry its first hard look at whether discounted tokens are translating into revenue at scale.