As of this week, per-million-token pricing across the major AI labs ranges from a fraction of a cent to fifty dollars. DeepSeek’s cheapest model charges as little as $0.0028 per million cached input tokens. Anthropic’s top-tier Claude Fable 5 charges $10 input and $50 output per million. Same basic unit — a token — priced almost 4,000 times apart depending on whose API you hit.
The story is not “AI is expensive” or “AI is cheap,” but the fact that both are true simultaneously, from labs competing head-to-head for the same customers.
Mapping the battlefield
Here’s the current rate card, sorted low to high:
Grouped bar chart, input vs. output price per million tokens, log scale, across ten models from DeepSeek V4 Flash ($0.14 / $0.28) to Claude Fable 5 ($10 / $50).
The Chinese challengers:
DeepSeek’s V4 Flash costs $0.14 per million input tokens and $0.28 output — the cheapest frontier-class API on the market. Moonshot’s new Kimi K3 costs $3.00 input and $15.00 output, with a discounted $0.30 rate on cache hits.
The Western flagships:
OpenAI’s GPT-5.6 Sol runs $5/$30 per million tokens, with cheaper Terra ($2.50/$15) and Luna ($1/$6) tiers beneath it
Anthropic’s Claude Opus 4.8 sits at $5/$25, with Sonnet 5 currently at an introductory $2/$10 through the end of August before reverting to $3/$15
Look at the middle of that list, though, and the story gets less clean. Claude Haiku 4.5 ($1/$5), Sonnet’s intro pricing ($2/$10), and GPT-5.6 Terra ($2.50/$15) are all clustered within a few dollars of each other. The “American AI is expensive” framing only holds up at the very top and very bottom of the ladder — the middle tier is where the real competition is happening, and it looks a lot more like normal market convergence than a price war.
Why the Western labs stay expensive
Three real reasons, not just brand markup:
Compute cost. Training and serving frontier models on Nvidia-class hardware at scale is genuinely expensive, and that cost shows up directly in output pricing — flagship output can run six to seven times more per token than a smaller model from the same lab. Model selection, not lab selection, is still the single biggest lever on your bill.
Enterprise trust. Compliance guarantees, data handling commitments, and uptime SLAs are worth real money to buyers in regulated industries, independent of raw model quality. This is the invisible line item in every ‘why is Claude more expensive than DeepSeek’ comparison.
A tiered ladder, not a floor race. OpenAI and Anthropic both still sell a range from cheap-and-fast to expensive-and-capable. Neither has raced its entire lineup to zero, because their bet is that enterprise buyers will pay for the top of the ladder regardless of what the bottom costs elsewhere.
How DeepSeek and Kimi undercut
Two real mechanisms, one geopolitical wrinkle:
Different silicon, different economics. DeepSeek has been explicit that its strategy runs on Chinese chips rather than Nvidia — largely a function of export restrictions that pushed Chinese labs toward domestic hardware. That changes the underlying cost structure in ways that don’t map cleanly onto a ‘the’'re just discounting” narrative.
Mixture-of-experts architecture. Both DeepSeek and Kim’s models activate only a fraction of their total parameters per request. Kim’s MoE design in particular means Moonshot pays for meaningfully less compute per token than a dense model of comparable size would cost — the savings are architectural, not just promotional.
Caching that rewards volume. DeepSeek’s cache-hit pricing runs roughly 1/50th of a cache miss. That’s not a flat discount — it specifically rewards high-volume, repetitive-prompt use cases like RAG pipelines and coding agents with stable system prompts, while doing little for one-off queries.
The Kimi K3 twist
Here’s the part that breaks the “cheap Chinese mode” narrative: Kim’'s newest flagship didn’t undercut Western pricing. K3 at $3/$15 sits close to, not dramatically below, Claude Opus 4.8 ($5/$25) and GPT-5.6 Sol ($5/$30) — while matching their million-token context window and landing just a few points behind them on Artificial Analysi’'s Intelligence Index, the closest thing the industry has to a single cross-model capability score.
Scatter plot, blended price vs. Intelligence Index score, for DeepSeek V4 Pro ($2.61 blended, score 44), Kimi K3 ($9 blended, score 57.1), GPT-5.6 Sol ($17.50 blended, score 58.9), and Claude Fable 5 ($30 blended, score 59.9). See attached chart export.
Plot price against quality and the pattern is almost a straight line — the market is pricing capability reasonably efficiently right now. DeepSeek V4 Pro is the clear outlier below that line: dramatically cheap, and dramatically behind on general capability. Kimi K3 sits closest to the line itself — the “fair value” point, not the bargain-bin point. That’s a meaningfully different pitch than the one Chinese labs were making a year ago.
Not everyone’s convinced K3’s launch justified the reaction it got. One widely shared take from developer Theo Browne put it bluntly: K3” "an incredible model. It is not an incredible value”.
What buyers should weigh besides the sticker price
Where inference actually runs. Data residency matters more than headline price for regulated buyers — worth a paragraph if your audience skews toward compliance-conscious industries.
Effective cost after caching and context overhead. In multi-turn agent loops, prior context resends every turn, so “cheap” model in a long conversation can end up costing more than a pricier model with better caching behavior.
Cost-per-quality-point, not cost-per-token. The scatter plot above is the right mental model: ask what y’u're paying per unit of capability, not just per token.
Where this goes next
The open question is whether the market keeps bifurcating: a race to zero for commodity inference tasks (classification, summarization, high-volume chat) running alongside a race to converge at the top, where Fable 5, GPT-5.6 Sol, and Kimi K3 all land within a few points of each other despite wildly different pricing philosophies behind them. Kimi 3’s launch is the earliest real evidence for that second track — a Chinese lab choosing to compete on capability at a premium price, rather than on price alone.
Closing
The AI pricing war “expensive American” versus “cheap Chinese” anymore. It’s splitting into two separate races — one racing toward zero, one converging at the top — and the label on the API key tells you less than it used to about which race you’re actually in.