On July 24, 2026, Anthropic released Claude Opus 5. For a company that built its reputation on "best-in-class" intelligence, the announcement had an unusual headline. It wasn't about raw capability. It was about cost.
"Opus 5 comes close to the frontier intelligence of Claude Fable 5 at half the price," Anthropic wrote in its announcement. The model is priced at $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8's pricing. Fable 5 charges $10 and $50.
The benchmark data backs the claim. On Frontier-Bench v0.1, a test for real software engineering tasks, Opus 5 scored 43.3% — more than double Opus 4.8's 18.7%, and ahead of Fable 5's 33.7%. On CursorBench 3.2 at maximum effort, Opus 5 landed within 0.5% of Fable 5's peak — at half the cost. On ARC-AGI 3, a test designed for fluid intelligence in novel environments, Opus 5 scored 30.2%. The next-best model, GPT-5.6 Sol, scored 7.8%. On OSWorld 2.0, Opus 5 outperformed every model at any price point, beating Fable 5's best result at just over a third of the cost.
Anthropic also emphasized that Opus 5 uses fewer tokens than Opus 4.8 to complete tasks, reducing average cost by about 30%. The company called it "designed to be used every day" — a signal that US AI labs are shifting their focus from chasing capability to chasing token efficiency.
But the pricing still sits firmly in the enterprise premium tier.
—— The Other Track: Capability Parity at a Fraction of the Cost ——
While Anthropic optimizes within a premium frame, Chinese AI labs are playing a completely different game.
DeepSeek V4 Flash charges $0.14 per million input tokens (cache miss) and $0.28 per million output tokens. V4 Pro runs at $0.435 input and $0.87 output. Cache hits push input costs as low as $0.0028 per million tokens. Compare that to Opus 5's $25 output — the gap is 89x. Not incremental. Structural.
Kimi K3, Moonshot AI's open-source model, charges $3 per million input and $15 per million output — still cheaper than Opus 5's $25, while Moonshot claims top-tier global performance.
UBS analysts estimated that training China's leading models costs roughly one-tenth of US counterparts. Inference costs run 10% to 20% of comparable US models, while maintaining 20% to 40% gross margins. As one Fortune analysis put it: "DeepSeek-V4-Pro's output costs about $0.87 per million tokens, while Z.ai GLM-5.2 costs $4.40."
By June 2026, Chinese models surpassed US models in weekly token usage for the first time. OpenRouter data shows Chinese open-source models run 60% to 90% cheaper than their top US equivalents from Anthropic and OpenAI.
Stanford's 2026 AI Index Report documented that the performance gap between US and Chinese models has effectively closed, with the two sides trading the lead multiple times since early 2025. A Fortune analysis noted that even US startups and Fortune 500 companies are "quietly plugging the models into their operations to rein in spiraling AI budgets."
—— Two Tracks ——
Track One: US labs optimizing within the premium tier. Anthropic, OpenAI, and Google compete to deliver more intelligence per dollar — from a pricing floor that stays high. Opus 5 is a real step forward in efficiency, but at $25 per million output tokens, it remains firmly in the enterprise premium category. Strategy: make the best model cheaper, but keep prices elevated.
Track Two: Chinese labs competing on affordability. DeepSeek, Moonshot, and Z.ai deliver models good enough for most enterprise workloads at prices 1/10th to 1/50th of US alternatives. UBS noted that Chinese developers have been "highly focused on efficiency improvements" rather than "pursuing the most powerful models at any cost." Strategy: make the model so cheap the cost question disappears entirely.
The two tracks reflect different assumptions. US labs assume enterprises will pay a premium for the best intelligence. Chinese labs assume enterprises will optimize for total cost of ownership — and that when the price gap is large enough, "good enough" beats "best."
—— Bottom Line ——
On raw capability, US labs still lead. Fable 5 and GPT-5.6 Sol are the most powerful publicly available models. The frontier is still American.
On cost efficiency, the gap has narrowed fast. Opus 5 delivers Fable-level performance at half the price — real progress, but within a frame where $25 per million output is considered a "discount." DeepSeek's $0.28 is not in the same universe.
On affordability, Chinese labs are winning decisively. When the price difference is measured in multiples of 50 to 100, the enterprise decision shifts from "which model is smarter" to "which model can I afford to run at scale."
The question is not which track is faster. It's which track builds a sustainable business. US labs bet on premium pricing. Chinese labs bet on volume at thin margins. The winner is whoever reaches escape velocity first.
The analysis above is based on publicly available data as of July 27, 2026. All benchmark scores, pricing, and performance claims are sourced from the respective companies' published materials. I am not affiliated with any of the companies mentioned unless explicitly stated. For the most current information, please visit the official sources linked throughout this article.