Here's the cleanest sign yet that AI pricing is being re-measured: the same week Anthropic quietly scrapped a planned 50% price hike on Claude Sonnet 5 and OpenAI slashed GPT-5.6 Luna by 80%, DeepSeek told its API users to expect a "significant" increase. Two of the most important pricing signals in the industry, moving in opposite directions at the same time.
The fork itself is the story — not the percentages. Beneath it, the real shift is that AI is no longer priced per token. It's being priced per successful task, and that single change is rewriting who can charge what for intelligence.
The fork: US models go cheaper, Chinese models go dearer
On one side, the frontier players are undercutting themselves. On July 30, OpenAI cut GPT-5.6 Luna by 80% (Terra got smaller cuts, Sol held its $30 output price), while simultaneously resetting paid-tier quotas on Codex and ChatGPT Work — effectively a price cut on both API and subscription. On August 10, Anthropic cancelled Sonnet 5's scheduled 50% increase, keeping it at $2 per million input tokens / $10 output as a permanent price.
On the other side, China's strongest models are raising. DeepSeek, still the cheapest frontier-class player at roughly $0.14/$0.28 per million tokens (V4-Flash), told users to expect a notable rise. And Moonshot launched Kimi K3 in mid-July at a deliberately premium band — ¥20 input / ¥100 output per million tokens. After two straight years of falling model prices, the curve has forked.
Why the US is cutting: it's not about margins, it's about workloads
OpenAI and Anthropic are fighting over the same thing: Agent workloads. Claude Code and Codex target the same developers and enterprises, and each extra real task run exposes where agents fail, retry, and call tools — feedback that gets wired back into evaluation, product, and reasoning strategies. The competitive loop is explicit: lower the entry barrier → more real tasks run → more failure modes surface → better agents → more workloads. For a workhorse Agent model, price directly controls call frequency, which is exactly why Anthropic abandoned a 50% hike on its main agent model.
The pressure from below: open-weight models are eating the middle
But the US price cuts only make sense against a second force: Chinese open-weight models are rapidly absorbing everything that doesn't need frontier intelligence. On OpenRouter, the token share handled by open models climbed from 34% in January to 65% by June, with the platform's four most popular models all coming from China.
The Stanford Digital Economy Lab's Enterprise AI Playbook (April) adds a sobering data point from 51 successful deployments across 41 organizations: 42% of projects said the underlying model was fully replaceable — rising to 71% for routine, rule-defined tasks, versus just 18% for high-complexity, high-stakes reasoning.
This is the real structural squeeze. Enterprises now route planning and complex reasoning to Claude or GPT, and shunt classification, extraction, simple code edits, and format conversion to DeepSeek, Qwen, or GLM. One task that used to send 100% of its tokens to a single flagship now splits across a router. Chinese open weights are compressing the surface on which US flagships can charge a premium — which is why OpenAI held Sol's price while slashing Luna's by 80%.
The unit of pricing just changed: cost per successful task
Agents are what made the old metric obsolete. A chat-era question was one or a few calls; a coding agent reads files, plans, hits the terminal, edits code, runs tests, reads errors, and loops — dozens to hundreds of calls behind a single instruction. The real economics now look like:
Task cost = token price × token count × agent steps × retries
Community-run benchmarks like DRadar make the consequences visible: higher reasoning tiers deliver better task performance, but at marginal costs that grow several to tens of times faster. A cheap model like DeepSeek V4-Flash finishes a meaningful share of real software-engineering tasks at a tiny fraction of the cost. The metric that actually matters to enterprises is now cost per successful task — a $0.50 model that needs 40 tries can be pricier than a $5 model that finishes in 8 steps. Stanford's research found the same pattern in productivity: when systems run in high-autonomy "escalation" mode (AI completes 80%+ of tasks, humans handle exceptions), median productivity gains hit 71%, versus ~30% for human-in-the-loop modes.
Why DeepSeek can raise — and what it's really testing
DeepSeek doesn't have pricing power yet. It has the right to test pricing power. Its installed base was built on being absurdly cheap, so a 2–3x increase is now a live experiment: if usage largely holds, global adoption was never just about price; if workloads flee to Qwen, Kimi, or GLM, then cheapness carried more weight than assumed. Moonshot is running a different bet entirely — open weights for distribution, premium API pricing to harvest high-value, long-horizon coding and knowledge work.
The tell from Stanford's research is that enterprises already treat models as swappable parts. One operations lead interviewed for the playbook built his own multi-LLM gateway, re-judging every request across cost, accuracy, relevance, and latency: "does this task really need deep search, or is a mini model enough?" Another team switches models freely — "whoever gets better or cheaper wins our traffic." Customers remain; model loyalty is collapsing.
What this means for you
- Stop budgeting in tokens. Start tracking cost per successful task in your own workloads. Token price alone tells you almost nothing about Agent economics.
- Build a routing layer. A simple multi-model gateway lets you match each request to the cheapest model that can do it — planning goes to frontier, execution goes to cheap. That's where most of the savings actually live.
- Evaluate the combination, not the model. New benchmarks like Agents' Last Exam test model + harness together, because a frontier model inside the wrong harness can underperform a cheap one in the right setup. Run your own evaluations on your real tasks before committing.
- Watch the DeepSeek experiment. If adoption holds after the hike, "cheap + open" graduates from a pricing strategy to a brand position — and every open-weight roadmap shifts with it. If it doesn't, the floor under Chinese model pricing just got re-negotiated.
The two price curves are converging from both directions — frontier premiums get squeezed from above and below — toward a single equilibrium: the price of getting a real task done. The models are becoming replaceable parts. The task is what's being priced.