文｜硅碳变量，作者｜温宁
DeepSeek正式推出大模型新版本V4.1 Flash-1017，此前已有公开数据显示其技术指标，此次DeepSeek更首次自主发布帕累托前沿曲线，直观呈现模型在效率与成本维度的综合表现。
V4.1 Flash-1017在此曲线中达到行业领先水平：单位任务综合成本仅为Fable、Astra等同类模型的百分之一至二十分之一，性价比优势达四五十倍以上。
实测表明，多数场景下Token用量较前版下降逾50%，相应TCO水平亦将同步迎来新一轮实质性下探。
一次后缀看着像小版本迭代的更新，可能正在改写整个行业算账的方式。
白菜价时代结束了，账单开始说话
过去两年，国产大模型经历过一轮烈度极高的API价格战。
2024年5月，豆包大模型在价格上主打“极致性价比”，通用模型pro-32k版推理输入价格0.0008元/千tokens，较行业价格低99.3%。
随后阿里云发布降价公告，将旗下通义千问的多款商业化及开源模型进行大幅降价。其中，对标GPT-4的商业化模型Qwen-Long的API（应用程序编程接口）输入价格从0.02元/千tokens降至0.0005元/千tokens，直降97%；输出价格从0.02元/千tokens降至0.002元/千tokens，直降90%。
百度智能云也发布公告称，百度文心大模型的两款主力模型ENIRESpeed、ENIRELite将全面免费，并即刻生效。
国产大模型API一度进入“厘时代”，部分轻量模型输入侧做到1块钱可买到上百万token。
然而底层算力和推理成本存在刚性约束，价格不可能一直无限向零收敛。自3月中旬起，国内云与AI算力市场迎来一轮价格结构性震荡：算力基础设施、部分模型API上调，行业定价逻辑开始分化。
3月18日，阿里云公告称，因全球AI需求爆发、供应链涨价，行业核心硬件采购成本显著上涨。公司决定于2026年4月18日起对AI算力、CPFS（智算版）等服务价格进行调整。其中，平头哥玄铁系列算力卡上涨5%至34%，文件存储产品CPFS（智算版）上涨30%。
腾讯云上调了Tencent HY2.0 Instruct模型的输入价格，涨幅高达463%。其还宣布自5月9日起对AI算力、容器服务及弹性MapReduce相关产品价格统一上调5%。百度智能云则宣布，自4月18日起，AI算力相关产品服务价格上调约5%至30%。
在海外市场，亚马逊AWS已对用于大模型训练的EC2实例实施15%的价格上调，搭载H200芯片的顶配实例价格涨幅达15%；谷歌云更是宣布对AI计算实例价格上调20%至50%，部分地区CDN出口带宽价格涨幅高达100%。
涨价背后是清晰的商业逻辑：依靠持续补贴很难跑出正向毛利；与此同时，企业客户对大模型的定位已经转变，它不再只是用来测试体验的工具，而是逐步嵌入业务链路的生产资源。
真正撬动行业竞争逻辑发生变化的，是大模型承载任务形态的迭代。
当模型从回答问题转向闭环执行完整项目，成本结构随之改变：编程智能体要走完需求解析、工具调用、代码改写、测试验证、循环修复全流程。
行业实测数据显示，无规范优化的原生提示词存在30%-55%的无效Token消耗，长期高频调用会造成大量算力资源与资金浪费。
当单价的下探空间见顶之后，怎么花钱就比单纯看Token单价更关键。竞争主轴切换的基础正在逐步成型，而一张帕累托曲线也开始重构行业对性价比的评判标准。
帕累托前沿并不是新概念，通俗来说它是性能与成本之间的最优边界：边界上的每一个点位，都遵循取舍逻辑——想要进一步提升性能，就要承担更高成本；想要压低开销，就要接受性能折损。
落在边界内侧的方案，则还有帕累托改进空间：有机会在不牺牲性能的前提下压低成本，或是不降成本、只拉高性能。DeepSeek这次升级本质就是一次典型的帕累托改进，新版本向着前沿曲线靠近，同时实现性能抬升且单位任务成本下行。
过往厂商的通行做法大多是分开展示指标：一边放出基准跑分，一边单独挂出API价目表，很少将二者放在同一个坐标系里做综合权衡。
DeepSeek这次的特殊之处在于，其作为模型供给方主动对外发布了面向商用场景的性能—成本帕累托前沿曲线，把模型在效率和成本两个维度的综合权衡关系直观呈现出来。
在DeepSeek公布的DeepSWE v1.1评测图中，新版本的最高通过率（Pass@1）达到77.0%。横向比较，上一版本DeepSeek V4.1 Flash的最高点约为74%。新版本的最高成绩出现在每项任务平均费用约1.3元的位置，显著低于图中不少性能接近的海外模型。
这件事的商业含义，远不止模型调用更便宜。将单位任务成本确立为横向对比的标尺，本质是抛出一套新的评价框架：以往B端选型，更容易优先参考公开基准榜单的跑分，而这套新框架会推动采购方去对比，完成同一业务目标背后的实际开销。
但这条曲线由厂商自行绘制：配套的任务集、评测框架DeepSeek Harness、推理档位规则，均为内部定义。跨模型对比是否成立以及外界第三方能否复现1/100至1/20的成本差距还有待进一步验证。
至少目前来看，它更接近一份厂商给出的官方口径，而非经过第三方独立审计的结论。
标价不动，账单缩水：一次柔性的降本
DeepSeek V4.1 Flash-1017这次迭代最值得玩味的设计在于对外API调用方式和公开定价体系完全维持原状，没有下调。
模型带来的成本红利并没有通过直接修改价目表释放，而是藏在底层对Token使用效率的重构里。
在面向商用客户的解读口径中，本次迭代把Token优化归纳为三层机制：任务感知型压缩、动态上下文裁剪、自适应解码策略；其底层对应的是KV缓存压缩、上下文管理、解码调度等一系列技术组合。
实测表明，多数场景下Token用量较前版下降逾50%，相应TCO水平亦将同步迎来新一轮实质性下探。
这一模式可以看作大模型服务里一种柔性降价思路：对外公开标价保持不变，客户完成同等任务对应的实际调用账单却能够缩水。对比直接下调API单价，这套策略巧妙绕开了直接价格战附带的多重风险。
首先，它不容易直接引爆新一轮公开价格战。直接下调API单价会生成一个透明且容易跟进的数字锚点，竞争对手可以快速对标降价，最终行业容易陷入单价内卷，集体挤压毛利空间。
而效率优化带来的成本下降，沉淀在任务执行链路之中，属于模型工程层面的隐性优势，对手想要复刻和对标验证，门槛会更高。
其次，存量客户已有的合同计价体系不会受到冲击。对已经签约的企业客户而言，API公开单价维持原样，原有商务条款和预算测算框架就不需要重新调整；客户在近乎无感知的状态下，收获更低的实际任务成本，有助于稳住客户留存。
更深一层来看，厂商的竞争叙事也在随之切换。竞争优势不再局限于“每千Token多少钱”这个直观标签，转而变成任务完成率和单位任务成本共同构成的复合指标。行业比价的重心正在从Token单价，慢慢转向完成同等业务目标最终需要承担多少实际开销。
高频调用场景、Agent智能体工作流、批量自动化任务以及RAG检索增强这类业务，往往是Token冗余消耗的高发地带。上下文反复回传和模型输出冗余冗长，持续推高长期调用的综合成本。
而这次Token消耗的显著下降能够优化这类业务的成本结构，拓宽Agent工作流的落地边界，让不少自动化业务在预算维度拥有了落地的可能性。
回溯V4.1 Flash首发时，DeepSeek当时选择直接下调API定价，同时推出峰谷计费方案，闲时调用可享受半价。时隔一个多月更新的1017版本，厂商不再在公开标价上做调整，转而依托推理层、Token 链路的工程优化向外释放红利。
这一转变传递出值得留意的信号：直接降价这条路径的边际效益正在衰减，行业竞争的重心正从价目表上的数字博弈，慢慢向推理效率和Token链路优化这类底层工程能力迁移。
省Token的钱，不是白捡的
Token效率优化释放的成本红利，存在清晰的适用边界。
受益相对显著的场景，集中在编程智能体、Agent工作流、RAG检索增强、在线客服这类高频调用、上下文偏短、输出偏向结构化的业务中。
这类场景的开销瓶颈往往不在于模型本身的基础推理成本，而是冗余输出、多轮对话里历史上下文反复回传带来的无效Token消耗。任务感知压缩、动态上下文裁剪这类技术手段恰好对准了这一痛点，降本的杠杆效应会更为突出。
但行业需要认清一个现实：Token压缩带来的成本节省，从来都不是无代价的。
模型输出精简和推理链路压缩放到复杂推理、超长文档解析、多步骤链式逻辑任务里，会浮现潜在风险：压缩策略有没有可能丢失关键细节，引发推理结果前后不一致？
依靠官方基准评测得到的平均任务通过率，很难覆盖这类隐患。能在测试集里做到精简输出不等于落地到多样真实业务时，依然守住关键信息与结果可靠性。压缩是信息留存和推理成本之间的权衡，一旦平衡点偏移，成本下降的背后或许就是业务质量的损耗。
与此同时，一组绕不开的争议落在对比口径上。官方给出单位任务综合成本仅为Fable、Astra等同类模型的百分之一至二十分之一的结论，其背后藏着前提条件：Flash主打轻量高速路线，Fable 5与Astra则是旗舰级模型，定位、算力底座、能力上限并不在同一基准线上。
跨档位模型横向对比成本，其公平性需要审慎拆解，不能直接把倍数差异等同于全面的性价比碾压。
OpenAI对外阐释Astra时，也着重提到以更少Token、更低重试频次优化成本效益。海内外厂商正在共用任务成本最优这一套叙事，但各方的评测体系、任务数据集并不互通，横向对标难度很高。
由此也引出一层值得警惕的行业隐忧：如果帕累托前沿曲线慢慢变成厂商对外营销的标准工具，这套性能与成本的衡量框架，是否会催生出新一轮榜单内卷？
任务数据集、推理档位、评测运行框架都存在可调空间，各家厂商都能在自定义的坐标系里画出观感亮眼的前沿曲线。越来越多自成一套的评测标准，反而会放大行业里的信息噪声。
对B端采购方来说，厂商发布的帕累托曲线只适合当作参考素材，衡量模型价值最可靠的标尺依然是自身业务场景下的真实测试。
第一，API市场的竞争主线，或将从纸面标价转向有效Token效率。下一阶段厂商比拼的核心未必是官网公示的每百万Token单价，而是同等预算下模型能够稳定交付多少有效的业务结果。单价数字的参考价值或将持续下降，单位任务投入产出比有望成为新的核心指标。
其二，B端的选型逻辑也可能随之重塑：依托自有业务数据搭建内部基准，测算单任务成本与任务达成率或将慢慢取代通用榜单，成为模型评估的首要环节。采购方不再迷信公开跑分，更看重业务场景里的实测表现。
第三，轻量 Flash 赛道的竞争或将进一步白热化。DeepSeek、MiniMax已经将单位任务成本作为核心卖点，国产轻量化高速模型持续加码，有可能持续下压海外同类产品的价格锚点。Flash赛道也不再只是简单的平价替代选项，开始正面争夺Agent自动化场景的落地份额。
第四，工程调度能力或将获得与模型基础能力相当的权重。上下文管理、推理调度、Token 压缩策略这类工程优化手段，有望逐步成为和模型架构、基础算法同等重要的差异化壁垒。
在本次迭代里，DeepSeek同步开放自研Search与Fetch API，同时完成和腾讯CodeBuddy、WorkBuddy等Agent框架的适配，也释放出明确信号：竞争边界已经跳出模型本身，延伸至完整的任务交付全链路。
这套靠Token优化降本的商业逻辑，成立的先决条件是效率优化不能损害任务质量。一旦压缩机制在长尾复杂场景高频出错，前期省下的Token成本，很可能变成重试、人工复核、业务故障带来的隐性开销，推高项目总TCO。
帕累托前沿曲线引发讨论，其核心变化不在于单项评测指标的冲高，而在于模型价值的评价主体正在发生转移。
可见，大模型的评判权正逐步脱离榜单跑分、纸面定价这类标签，转向企业客户手中的真实业务账单。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Cyclium Silicon Carbon Variable, by Winning
DeepSeek officially launched a new version of LLM, V4.1 Flash-1017, with publicly available data on its technical indicators, and this time DeepSeek for the first time has taken ownership of the frontier curve of publish Paretto, visualizing the combination of efficiency and cost dimensions of the model.
V4.1 Flash-1017 leads the industry in this curve: the combined unit mission cost is only one-hundredth of one-hundredth of the equivalent models, such as Fable and Astra, and the value for money is more than forty-fifty times higher.
In most scenarios, Token use has been reduced by more than 50 per cent compared to the previous version, and the corresponding TCO levels will coincide with a new round of substantive downswings.
A suffix looks like a small version of an iterative update and may be rewriting the way the whole industry counts.
It's over for cabbage. The bills are starting to talk.
In the past two years, the national production of LLM has experienced an extremely strong API price war.
In May 2024, Doubao LLM played “extreme value” in price, and the generic model pro-32k version had an reasoning entry of a price of $0.0008/thousands tokens, which was 99.3 per cent lower than the industry price.
Subsequently, Alibaba Cloudpublish was announced as a price reduction, which resulted in a substantial reduction in the commercialization and open source model of Qwen under the flag. Of these, the input price for the commercialization model GPT-4 Qwen-Long API (application programming interface) dropped from 0.02 to 0.00/ktokens, or 97 per cent; the output price fell from 0.02 to 0.002/ktokens, or 90 per cent.
Baidu also declared that the two main power models of BaiduQ63LLM, ENIRESpeed, ENIRELite, would be fully free of charge and effective immediately.
The national product, LLMAPI, once entered the era of “fiction”, and some light-weight model input sided to buy millions of dollars.
However, there is a rigidity of bottom calculation and reasoning costs, and prices cannot remain unlimited to zero. Since mid-March, there has been a cycle of structural price shocks in the domestic cloud and in the AI computing market: an increase in the computing infrastructure, a partial model of the API, and a differentiation of industry pricing logic.
On 18 March, Alibaba Cloud announced that the cost of purchasing core hardware in the industry had increased significantly as a result of the global demand for AI and the price increase in the supply chain. The company decided to adjust the prices of services such as AI Calculus, CPFS (intellectual version) with effect from 18 April 2026. Among them, the CPFS (Intelligent Edition) product for document storage increased by 5 to 34 per cent.
The Tencent cloud increased the input price of the Tencent HY2.0 Institute model by 463 per cent. It also announced a 5 per cent uniform increase in prices for products related to AI computing, container services and elastic MapReduce, effective 9 May. For its part, Baidu Smart Clouds announced an increase of about 5 to 30 per cent in the price of AI-related services from 18 April.
In overseas markets, the Amazon AWS has implemented a 15 per cent price increase for EC2 examples used for LLM training, with a 15 per cent increase in the price of the top example of the H200 chip; Google Clouds has announced an increase of 20 to 50 per cent in the price of the AI calculation example and a 100 per cent increase in the bandwidth price of CDN exports in some areas.
There is a clear business logic behind the price increase: it is difficult to run straight out of the Māori with continued subsidies; at the same time, the positioning of business clients towards LLM has changed, and it is no longer merely a tool to test experiences, but rather a production resource embedded in business links.
What really drives the logic of competition in the industry is an iterative pattern of LLM missions.
When the model moves from answering questions to implementing the full project in a closed loop, the cost structure changes: programming AI agent completes the needs analysis, tool call, code rewriting, testing validation, recycling restoration process.
The industry empirical data show that the unregulated and optimized primary prompt is 30-55 per cent inefficient Token consumption and that long-term HF transfers can result in significant resource and financial waste.
It is more critical to spend money than simply looking at Token's unit price when the lower unit price is at the top. The basis for the main axis of competition is gradually being shaped, and a Paretto curve has also begun to reshape the industry ' s criteria for value for money.
The Paretto front is not a new concept; it is generally the best boundary between performance and cost: every point on the border follows the trade-off logic — higher costs to further improve performance; lower costs are accepted.
In the case of programmes that fall within borders, there is also Pareto to improve space: there is an opportunity to reduce costs without sacrificing performance, or to do so at no cost. DeepSeek's upgrade is a typical Paretto improvement, with a new version coming in close to the frontier curve while achieving performance uplifts and unit mission costs down.
It is common practice for former manufacturers to display indicators in isolation: while setting out the benchmark run points, putting up the API list on its own, they are rarely weighed in the same grid.
DeepSeek's special feature this time was that it, as a model supplier, had proactively launched publish, the performance-cost Paretto frontier curve for the commercial scene, presenting the combination of efficiency and cost dimensions of the model.
In the DeepSWE v1.1 assessment published by DeepSeek, the maximum pass rate for the new version (Pass@1) reached 77.0 per cent. For horizontal comparison, the highest point in the previous version of DeepSeek V4.1 Flash was about 74 per cent. The new version ' s highest achievement appears at an average cost of approximately $1.3 per mission, significantly lower than many of the offshore models that are close to performance in the map.
The business implications of this matter are far more than model calls. The essence of establishing unit mission costs as a horizontal comparison is to throw out a new evaluation framework: the B-end model in the past has made it easier to refer more easily to the running points of the open benchmark list, which will drive the purchaser to compare the actual costs behind the completion of the same business objective.
But this curve is drawn by the manufacturer itself: the accompanying task set, the evaluation framework DeepSeek Harness, and the reasoning slotting rule are all internal definitions. The validity of cross-model comparisons and the replicability of the cost gap of 1/100 to 1/20 by external third parties have yet to be further tested.
At least for the time being, it is closer to the official calibre given by a manufacturer than to the findings of an independent third-party audit.
The bill is shrunk: a soft drop
DeepSeek V4.1 Flash-1017 is the most interesting design for this iterative period in that the external API approach and open pricing system remain completely unchanged without downward revision.
The cost dividend resulting from the model was not released by direct modification of the price list, but was hidden in the re-engineering of the underlying efficiency of the use of Token.
In the interpretation calibration for commercial clients, Token ' s optimization is grouped into three tiers of mechanisms: mission perception compression, dynamic context tailoring, self-adaptation decodement strategies; the bottom line corresponds to a range of technical combinations such as KVcache compression, context management, decoded scheduling.
In most scenarios, Token use has been reduced by more than 50 per cent compared to the previous version, and the corresponding TCO levels will coincide with a new round of substantive downswings.
This model can be seen as a soft-price reduction in LLM services: the external public bid remains the same, while the actual call bill for the customer to perform the same job can shrink. In contrast to the direct downgrading of the API unit price, the strategy is crafty to circumvent the multiple risks associated with direct price battles.
First, it is not easy to trigger a new round of open price wars directly. A direct downgrading of API unit prices would create a transparent and easy-to-follow digital anchor, allowing competitors to quickly reduce prices to the mark, and eventually the industry would be vulnerable to falling into a single price roll and crowding out Maori space.
The cost reductions associated with optimized efficiency are embedded in the mission performance chain, which is a hidden advantage at the model engineering level, and the threshold is higher if the counterparty wants to reset and cross-check.
Second, the existing contract pricing system of the stock customer would not be affected. For already contracted business clients, the API open unit price remains the same, and the original business terms and budget measurement framework need not be readjusted; clients reap lower actual mission costs in a near-unsure state, helping to stabilize their clients ' retention.
At a deeper level, the competition narrative of the manufacturer is changing. The competitive advantage is no longer limited to “how much per thousand Token” This visual label has been transformed into a composite indicator of mission completion rates and unit cost. The focus of industry price comparisons is moving slowly from the Token unit price to the actual cost of meeting the same business objectives.
High frequency calls, Agent AI agentworkflow, bulk automation missions and RAG retrieval to enhance such operations are often high-growth areas of Token redundancy. context repeats and model output is lengthy, consistently pushing up the combined costs of long-term calls.
This significant reduction in Token consumption will optimize the cost structure of such operations, widen the landing boundary of Agentworkflow and give a number of automated operations the possibility of landing in the budgetary dimension.
Back when V4.1 Flash was launched, DeepSeek opted for a direct downward adjustment of the API pricing, while introducing the Peak Valley Billing Program, which allows for half-priced calls when available. More than a month later, version 1017 was updated, and the manufacturer no longer adjusted the open bid and released the dividends from outside on the basis of engineering optimization of the reasoning layer, the Token link.
This shift sends a noteworthy signal that the marginal benefits of the direct price reduction path are declining, that the focus of industry competition is on digital games on the price list, and that there is a slow shift towards reasoning efficiency and token chain optimization of such lower-level engineering capabilities.
Token's money is not for nothing.
Token optimizes the cost dividends released, with clear applicable boundaries.
The relatively significant benefits are concentrated in programming AI agent, Agentworkflow, RAG search enhancement, high frequency calls such as online passenger service, context was short and output was structured.
The cost bottlenecks of such scenarios are often not the underlying reasoning costs of the model itself, but rather the inefficient token consumption resulting from redundant outputs, repeated echoes of history context in multiple rounds of dialogue. Mission perception compression, dynamic context tailoring, etc., are just as relevant to this painful point, and the leverage effect of the drops is even more pronounced.
But the industry needs to understand the reality that cost savings from Token compression are never cost-free.
Simplified model output and compression of the reasoning chain into complex reasoning, hyper-long document resolution, multi-step chain logic missions would present potential risks: could the compression strategy lose key details and trigger inconsistent reasoning?
The average mission pass rate, which is measured by official benchmarks, makes it difficult to cover such risks. The reliability of key information and results is maintained when streamlining output in the test set does not amount to landing in a variety of real operations. Compression is the trade-off between the retention of information and the cost of reasoning, and once the balance has shifted, it may be that the loss of the quality of the business is behind it.
At the same time, a set of inconvenient disputes fell on the contrast caliber. The official conclusion that the combined cost of unit assignments is only one-tenth of one-hundredth of the equivalent models Fable, Astra, etc., is based on the premise that Flash is responsible for light speed routes, that Fable 5 and Astra are flagship-size models, and that positioning, computing bottoms, and capacity ceilings are not on the same benchmark.
The equity of cross-slotting models, which compare costs horizontally, requires careful disassembly and cannot be directly equated with full value for money crushing.
Open AI's external interpretation of Astra also highlighted the need to optimize cost-effectiveness with fewer Token and less retry frequency. While off- and off-site manufacturers are sharing the best set of narratives on mission costs, assessment systems, mission data sets are not mutually compatible and horizontally difficult.
This also raises an alarming industry concern: if the Paretto frontier curve slowly becomes a standard tool for the external marketing of producers, will this performance and cost measurement framework lead to a new round of internal rolls?
The task data sets, the reasoning slots, and the evaluation of the operational framework are all available, and each manufacturer is able to draw an eye-opening frontier curve in a custom system of coordinates. More and more self-contained assessment criteria tend to magnify information noise in the industry.
For the B-end purchaser, the Pareto curve of the manufacturer publish is suitable only for reference asset, and the most reliable ruler for measuring the value of the model remains a true test in its own business landscape.
First, the main line of competition in the API market or the shift from paper prices to effective Token efficiency. The core of the next stage of the manufacturer ' s match is not necessarily the value per million Token unit price advertised by the network, but rather how much effective business results can be delivered under the same budget model. The reference value of the unit price figures may or will continue to decline, and it is expected that the unit input output will become a new core indicator.
Second, the B-side selection logic can also be reshaped: Based on the availability of business data, an internal benchmark is established to measure individual mission cost and mission achievement rates or will slowly replace the generic list as the primary element of model assessment. Purchasing parties are no longer superstitious and openly score points, and are more focused on measured performance in the business landscape.
Third, the competition for light-weight Flash tracks may be further heated. DeepSeek, Mini Max, has identified unit mission costs as the core selling point, and the light quantization high-speed model of national production continues to be coded, with the potential to keep down price anchors for similar products abroad. The Flash track is no longer a simple affordable alternative and has begun to compete for the landing of the Agent automation scene.
Fourth, the engineering movement control capacity may be given a weight corresponding to the model base capacity. Engineering optimization tools such as context management, reasoning scheduling, and the Token compression strategy are expected to gradually become equally important differential barriers to modelling and underlying algorithms.
In this iterative era, DeepSeek ' s simultaneous and open self-study of Seach and Fetch API, along with the fitting of the TencentCodeBuddy, WorldBudy and other Agent frameworks, also sends a clear signal that the competitive boundary has jumped out of the model itself to complete task delivery.
This business logic, which relies on Token to optimize the downside, is premised on the fact that efficiency does not compromise mission quality. Once the compression mechanism has made a HF error in a complex long-tailed scene, it is likely that the cost saved by Token in the preceding period will become a hidden cost of retry, manual review, operational failure and push up the total TCO of the project.
The central change in the Paretto frontier curve, which triggered the discussion, was not the rush of individual assessment indicators, but the shift in the evaluation body of model values.
As can be seen, the judgement of LLM is gradually moving away from running lists, paper pricing, and towards real business bills in the hands of business clients.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.