Chinese Open-Weight Model Price Hike Signals a Shift from Per-Token to Per-Task Cost

A Price Increase as a Valuation Signal

A next-generation release in a major Chinese open-weight model family has raised its official output price to the level of roughly one hundred yuan per million tokens. Input pricing rose as well, while cache-hit pricing remains significantly lower. The new price point closely tracks the standard published rate of a flagship mid-tier model from a major international cloud provider—though it is notably higher than that provider’s current promotional pricing. In other words, calling the new model during the promo window can actually cost more per unit than calling the benchmark model it targets.

The technical significance of this move outweighs its financial significance. For a long time, Chinese model vendors have generally followed a low-price strategy, pricing well below comparable international models to capture usage volume. This adjustment is different: for the first time, the vendor has abandoned shadow pricing and directly anchored to the stated “standard value” of the competing product. The signal is clear: discounts are operational tactics; list prices are a public statement about how the vendor values its own model’s intelligence. The former can change with market conditions at any time. The latter is a capability commitment aimed at the outside world.

It is important to note that an official list price is not a faithful reflection of cost. It functions more like a “narrative price”—a reference point carved into the market before the model weights are even released. Since the weights are scheduled to become available within days, the official list price may have an effective window of only about ten days. After that, third-party inference providers will quickly offer the same weights at lower prices. At that point, users who see third-party pricing far below the official rate will feel they are getting a bargain—because the official price has already anchored the upper bound of perceived value. The combination of open weights and premium pricing is rare in model history. It only holds if the model has strong enough benchmark scores to back it up.

In the Reasoning Model Era, Cost Shifts from Input to Output

Understanding this price change requires understanding how billing structures differ for reasoning models.

Traditional conversational models bill primarily on the input side: user prompts are short, model responses are short, and the input-to-output ratio is relatively balanced. Reasoning models, by contrast, generate a large number of thinking tokens before producing the final answer, and these thinking tokens are almost always billed as output. Furthermore, some models enable the highest reasoning intensity by default, with no simple configuration option for users to reduce the amount of thinking. This means:

• A 5x list price increase often translates into an even larger increase on actual invoices.
• The more the model thinks, the more you pay, with cost scaling nonlinearly according to task complexity.
• The output side replaces the input side as the dominant factor in the cost structure.

Pseudocode makes this billing difference clearer:

# Traditional chat model: cost is input-dominated
cost_chat = input_tokens * price_in + output_tokens * price_out
# Typical ratio: input accounts for over 70%
# Reasoning model: thinking tokens are billed as output, so cost is output-dominated
cost_reasoning = input_tokens * price_in + (answer_tokens + thinking_tokens) * price_out
# Typical ratio: output accounts for over 80%, and thinking_tokens grows with task difficulty

From this follows the first conclusion: at a list price of roughly one hundred yuan per million tokens, no human user can afford to use this model for casual chat. The real customer for this price point is not a person conversing with a model—it is an agent. That is, machine-to-machine token consumption in automated scenarios. In such scenarios, tokens are production inputs rather than an interaction medium. Callers are sensitive to unit prices but even more sensitive to total task cost.

The Billing Unit Migrates: From Per Million Tokens to Per Task

When the buyer is an agent, the procurement decision algorithm changes. Nobody calculates the price per million tokens item by item anymore. Instead, the relevant metric becomes cost per task.

Take a medium-complexity automated task as an example. Suppose the new model consumes about 0.94 USD per completed task on average—comparable to an international flagship reasoning model and roughly half the cost of a higher-end flagship product. In a per-task pricing world, this is not expensive, because it internalizes the uncertainty of “how many steps the model needs to think through.” What the caller cares about is the ratio of task success rate to per-task cost, not the unit price itself.

But within the open-weight camp, a horizontal comparison leads to a very different conclusion. For the same task, some Chinese open-weight models can achieve per-task costs as low as 0.32 USD. Other models have used permanent price cuts to push output pricing down to single-digit yuan per million tokens, bringing per-task costs to just a few cents. Measured by output unit price, the new model’s list price can be more than ten times higher. The same model can appear simultaneously “expensive” and “cheap” depending on which pricing framework you apply. This is the cognitive dissonance created by a paradigm shift in billing units.

In engineering practice, enterprises should establish a two-layer cost model:

# Layer 1: unit cost model, for rough budgeting and capacity planning
unit_cost = (input_tokens * p_in + output_tokens * p_out) / 1e6
# Layer 2: task cost model, for model selection decisions
task_cost = sum(attempt_cost for attempt in attempts) / success_count
# attempt_cost already includes thinking tokens, retries, and failure overhead

Only the second layer reflects real business costs, because failed retries, redundant thinking, and context expansion are all invisible to the unit price model.

The Time-Window Strategy Behind Official List Pricing

There is another mechanism worth unpacking in this price adjustment: the official list price has a limited shelf life.

Once the weights are open, third-party inference providers can immediately host the same model and offer it at lower prices. At that point, the official list price has only one remaining function—to inscribe a number on the model’s value before the weights become commoditized. This is a classic value-anchoring strategy:

1. The vendor first sets a high price to establish the psychological expectation that “this model is worth this much.”
2. The weights are released, and third parties enter the market at lower prices.
3. Users get equivalent capability at third-party prices, which they perceive as a “discount” rather than as “cheap.”

This strategy only works if the model’s actual capabilities support the anchor price. Otherwise, the anchor collapses quickly under benchmark comparisons. For enterprise users, this means the timing of model selection can lead to significant cost differences: calling the model during the official window means paying a brand premium; calling it after third-party hosting becomes available means paying for compute. If your business is not sensitive to minor model version differences but is sensitive to cost, the rational approach is to prioritize evaluating third-party hosting options after the weights are released, and treat the official list price as a capability reference rather than a procurement basis.

Practical Guidance for Enterprise Model Selection

Based on the analysis above, enterprises evaluating such high-priced open-weight models can follow this framework:

First, distinguish scenario types. For human chat scenarios, prioritize low-cost conversational models and avoid paying for thinking tokens. Agent and batch-processing scenarios are the appropriate fit for reasoning models, and decisions must be based on per-task cost rather than token unit price.

Second, control the thinking budget. If the model supports configurable reasoning intensity, reduce thinking strength as much as possible while keeping task success rates acceptable, and set an upper limit on thinking tokens. Models without such configuration capability are at a cost-control disadvantage.

Third, exploit the price timing gap. The window between official list pricing and third-party hosting prices represents a real arbitrage opportunity. Enterprises can include model weights in their own inference infrastructure evaluations and, where compliant, self-host or use third-party hosting services to lower per-task costs.

Fourth, establish task-level benchmarks. Measure success rates, average consumption, and retry rates on real business task sets. Calculate per-task cost instead of citing vendor-published per-token prices. Vendor list prices are valuation narratives; task cost is operational reality.

Fifth, pay attention to cache-hit pricing. Cache-hit prices on the input side are typically far lower than standard input prices, which matters significantly for agent scenarios with long contexts and repeated prefixes. Designing prompt structures to improve cache-hit rates is a necessary engineering practice for running reasoning models cost-effectively.

Conclusion

This price adjustment is not simply a rate increase. It marks the first time a Chinese model vendor has, at the pricing level, refused to fight a price war and refused to remain in the shadows of international benchmarks. It moves the industry discussion from “how much per million tokens” to “how much per task,” and from unit-price competition to value anchoring.

For developers and enterprises, the right response is not an emotional “too expensive.” It is to upgrade your cost model: change the billing unit from tokens to tasks, change the reference point from list price to measured results, and change the procurement strategy from following official pricing to dynamically choosing among multiple hosting channels. In the reasoning model era, the more a model thinks, the more you pay. Whether that thinking is worth it depends on how carefully you do the math before you choose.