OpenAI releases GPT-6.1 Terra with a 50% price cut—can Chinese large models still hold onto their low prices?

Anthropic had just submitted its prospectus, and OpenAI promptly released a new model right after—the giants are making move after move, all vying for attention.

On October 27 Beijing time, OpenAI released GPT-6.1 Terra. OpenAI claims that Terra's Agent capabilities are enough to go toe-to-toe with the flagship Astra, yet its price is only one-tenth of the latter's.

This number itself is nothing novel—after all, over the past year, news of large-model price cuts has long grown tiresome. But what makes this different from the past is that, compared with the previous generation GPT-5.6 Terra, the new model's standard API price is directly cut by 50%, while it shows clear improvements in complex task execution, factual accuracy, and coding capability.

OpenAI's new move is drawing particular attention in China—after all, what it wants to seize is exactly the territory where Chinese large models are most comfortable.

01 OpenAI Is Now Doing the "Economics Math" Too

If you look only at the top scores on leaderboards, Terra is hardly a bombshell. But put the pricing page and the benchmark chart side by side, and the signal is more worth reading than the scores: what OpenAI is selling this time is not the capability ceiling, but unit cost.

According to OpenAI officially, GPT-6.1 Terra's API pricing is $1 per million input tokens and $5 per million output tokens, 50% lower than the previous-generation Terra and the same-generation Sol. At medium reasoning intensity, its score on Agents' Last Exam is 15 percentage points higher than GPT-5.6 Terra, and the per-task cost is about one-third of the latter's.

Many users in the comments reacted first with: "That's it?" "Feels not much more impressive than the previous generation." The official positioning is very blunt: to complete complex computer-operation tasks at lower cost. Under this positioning, "whether it is the strongest" takes a back seat; "good enough and cheap" is the selling point.

Looking only at OpenAI's own product line, you would feel this is just a normal generational upgrade. But bring your gaze back to the current large-model market, and another reference line cannot be avoided: domestic models such as Kimi K3.1 Pro and DeepSeek have already dragged down the price anchor for high-performance models.

Kimi K3.1 Pro's API pricing is $2 per million input tokens and $10 per million output tokens. This is a model with 2.8 trillion total parameters (weights are planned to be open-sourced in early November), with overall intelligence above the median of its peers; Moonshot AI says its price is "only one-fifth of GPT-6 Astra."

Looking at absolute numbers alone, Terra's output price of $5 per million tokens is indeed lower than K3.1 Pro's $10. But the two are not on the same tier: Terra is the "senior worker" in OpenAI's matrix, with Sol and Astra above it taking on the most complex tasks; K3.1 Pro is a product in the open-source camp going head-to-head with frontier closed-source flagships. The former targets "doing daily chores, good enough and cheap," while the latter targets "close to the frontier, and cheap."

More critically, Chinese models have pushed down cost-effectiveness not by burning money on subsidies. UBS Securities analysts believe that the training cost of Chinese large models is only one-tenth that of overseas leading vendors, and API pricing is 10% to 20% of comparable overseas models; this low price does not come from losses—vendors can still maintain 20% to 40% gross margin on the API business.

The source of the cost advantage is that the mixture-of-experts architecture makes the number of parameters activated per inference extremely low, and leading vendors' GPU utilization reaches over 70%. The MoonClip optimizer used by Moonshot AI on the K3 series is officially said to achieve "the effect of 40T with 20T training data"; the linear attention mechanism then suppresses the cost inflation brought by long context. They have compressed the cost structure itself through architecture and engineering efficiency.

OpenAI's logic is clear: the flagship model shows off muscle, Terra handles volume. The volume price directly benchmarks all "cheap and good enough" models worldwide.

02 The Hard Boundary of Compute and the "Quota" Problem

The most critical point is that Terra is not merely a defensive price cut—hard capability is improving in tandem.

According to data published by OpenAI, in the Terminal-Bench 4.0 command-line coding test, Terra's success rate at the highest reasoning intensity reaches 50.3%, versus only 21.5% for the previous generation; in internal factual tests, the error rate at low reasoning intensity dropped from 53.6% to 16.7%, and at the highest reasoning intensity to 4.3%. The work is done better, and the price is cheaper.

But Chinese large models cannot avoid a pain point: quota. Less than 48 hours after Kimi K3.1 Pro's release, Moonshot AI issued an apology notice, admitting that after the new model went online it suffered rate limiting and abnormal quota consumption, and stated bluntly that "the problem is on us, and we should have fixed it earlier," while urgently compensating subscribers with 25% of their monthly total usage.

This is not an isolated case. The DeepSeek platform has also suffered multiple large-scale outages due to an imbalance between surging users and compute expansion; its official service status page shows that since the beginning of 2026 there have been dozens of service anomalies. Algorithms and architecture can compress unit cost to extremely low levels, but under the total-quantity constraint of compute supply, scale itself is a hard boundary. UBS analysts also point out that the core bottleneck for some model vendors' commercial expansion "mainly lies in insufficient compute supply, not market demand."

Of course, Terra is not the strongest model; it just makes the "good enough plus cheap" tier more aggressive. And this tier is precisely the main battlefield of domestic models. So the challenge domestic models need to face is not technology, but: when "cheap" is no longer an exclusive weapon, and opponents also start competing on cost-effectiveness, what is left in hand?

B-end scenarios, data compliance, localized deployment, and government-enterprise channels—these are things OpenAI cannot provide in the short term.

In the future, what developers and enterprises will ask is probably no longer "which model is the smartest," but "which model is the most cost-effective for completing this task." The division of labor among vendors is also becoming clearer: flagships tackle the capability ceiling, while the middle tier is responsible for stuffing AI into workers' daily workflows.

The real contest is no longer "model versus model," but a calculation beyond the model: whoever can make cheapness a sustainable cost structure, and whoever embeds models into scenarios others cannot enter, will be the one who laughs last.

English text is being prepared.