OpenAI, Anthropic, xAI cut model costs 40–50%, prioritize AI safety
In brief
- OpenAI's GPT-6 Sol dropped 50% to $2/$10 per million tokens; Luna tier at $0.10/$0.50
- Anthropic's Claude Opus 5.5 priced 40% lower at $4/$20 per million tokens; xAI's Grok 4.7 at $2/$6
- Chinese AI models undercut U.S. offerings by up to 9× per-token basis, forcing repricing
- Altman, Amodei, and Musk called for deliberate slowdown to prioritize safety research
- Usage-based pricing drives lab profitability directly; margin pressure cascades through cloud providers
The price cuts
OpenAI's GPT-6 Sol model is priced at approximately $2 per million input tokens and $10 per million output tokens, representing a 50% reduction compared to GPT-5.6. The company also released GPT-6 Luna at $0.10 input and $0.50 output per million tokens, targeting cost-sensitive workloads. Anthropic's Claude Opus 5.5 lands at $4 input and $20 output per million tokens, a roughly 40% reduction from its predecessor. xAI's Grok 4.7, launched around September 21–22, 2026, entered at $2 input and $6 output per million tokens.
The timing is not accidental. By mid-2026, Chinese AI models were undercutting comparable U.S. offerings by as much as nine times on a per-token basis. Silence wasn't an option. OpenAI, Anthropic, and xAI moved aggressively to reclaim margin and market share while signaling that frontier-grade AI no longer commands a premium purely on capability alone.
Safety and economic sustainability
The repricing also maps onto a broader shift in how AI labs frame their mission. Sam Altman, Dario Amodei, and Elon Musk have each made public statements in 2026 calling for a deliberate slowdown in frontier AI development to allow safety research and regulatory frameworks to catch up. Lower model costs reduce the financial pressure to release models faster and cheaper to justify R&D spend.
Amodei has specifically flagged the emergence of autonomous AI agents as a near-term risk category, with warnings that without adequate safeguards, dangerous capability gaps could materialize within a six to twelve month window. Cheaper models don't solve that risk directly, but they do allow labs to fund safety research without sacrificing revenue. The math changes when you're not forced to race.
Ripple effects across the stack
The usage-based pricing model that has become standard across the industry amplifies this dynamic. Revenue now scales with actual consumption rather than license seats, so labs benefit directly when enterprises run more queries. Lower per-token costs mean higher absolute volume. AWS, Google Cloud, and Azure each offer managed AI inference products that sit on top of these foundation models. When underlying model costs drop 40–50%, the margin structure of the entire inference stack gets renegotiated.
Mid-tier AI companies that built business models around the gap between frontier and commodity models face pressure as that gap narrows. Specialists that sold fine-tuning or optimization layers on top of older, expensive models now compete with cheaper frontier offerings. The industry consolidates further.
The repricing is a bet. It assumes that sustainable AI development requires room to breathe, and that room costs less than the alternative: a chaotic race that blows past safety checkpoints and invites regulatory backlash.
