智东西（公众号：zhidxcom）
编译 | 孙小雯
编辑 | 程茜
OpenAI和Anthropic为了抢用户，最近又杠上了！
10月30日，OpenAI刚将部分Pro套餐的ChatGPT Work和Codex额度翻倍；次日，Anthropic取消了已验证Ultra账户的5小时用量限制，周限额保留。11月1日，OpenAI又在帮助文档中向用户预告，将Work功能整合进普通聊天，并计划从12月起，让Plus、Pro和Business套餐的ChatGPT与Codex共享使用额度。
订阅套餐轮番加码，与此同时，年价值上亿美元的企业订单，也成了两家争抢的目标。
智东西11月2日消息，据美国科技媒体The Information于10月24日报道，在OpenAI提供超过30%的模型价格折扣后，美国AI编程平台Replit将大部分AI支出从Anthropic转向OpenAI。据该媒体援引知情人士消息，Replit与OpenAI的合同年价值超过1亿美元（约合人民币6.71亿元）。
为了抢客户，OpenAI的模型报价低至六折、送免费额度、让工程师进群“贴身服务”。一家银行刚开始考虑试用Claude，OpenAI就开出ChatGPT账户翻倍、年费不变的条件。
目前，包括美国AI编程平台Replit、人力资源软件公司Rippling和IT支持自动化公司Atomicwork等在内的多位Anthropic老客户，都或多或少选择“弃Anthropic投OpenAI”。
据The Information援引知情人士消息，Replit此前主要向Anthropic采购模型服务，原本打算继续合作。但在合同谈判中，Anthropic要求Replit承诺更高的采购支出，才愿意延续此前的折扣。
与此同时，OpenAI给出了超过30%的模型价格折扣。Replit最终接受了OpenAI的报价。据上述知情人士披露，双方合同年价值超过1亿美元（约合人民币6.71亿元）。Replit仅用几小时就完成了从Anthropic切换至OpenAI所需的代码修改，产品调整用了几天，比此前与Anthropic谈合同的时间还短。
美国人力资源软件公司Rippling也增加了OpenAI模型在编程工作中的使用比例。该公司的工程师此前就同时使用Anthropic和OpenAI的模型，但主要依赖Anthropic。
据The Information报道，今年5月，Rippling数百名工程师的编程工作中，Anthropic模型承担57%，OpenAI模型承担16%；截至10月24日该报道发布时，Anthropic模型的占比降至18%，OpenAI模型的占比升至52%。
OpenAI的抢客户策略已经初见成效。除了折扣，OpenAI还用了哪些办法争取订单？Anthropic又如何应对？
一、Anthropic季度营收反超OpenAI，头部模型智能指数仅差1分
据The Information报道，今年第二季度，Anthropic的营收首次超过OpenAI。Anthropic于10月26日公开提交的IPO注册文件显示，公司当季营收约115.55亿美元（约合人民币775.87亿元）；路透社此前披露，OpenAI同期收入为67亿美元（约合人民币449.88亿元），但路透社同时指出，两家公司的收入统计口径并不完全一致，OpenAI未计入部分云合作伙伴销售收入。
▲Anthropic季度经营数据，2026年第二季度营收为115.55亿美元（图源：Anthropic IPO注册文件）
据The Information分析，编程模型的表现，让Anthropic拥有了较强的定价能力。Anthropic的IPO注册文件显示，公司第二季度算力及基础设施支出约为55.58亿美元（约合人民币373.20亿元），不到同期营收的一半。
OpenAI的情况则不同。据The Information报道，OpenAI此前预计，今年算力支出将明显超过收入。该媒体认为，大幅折扣正在帮助OpenAI争取客户，但也会加大其收支压力。
不过，The Information采访的多名应用开发者称，OpenAI近几个月推出的新模型比此前更能与Anthropic竞争。
从最新跑分来看，两家的头部模型已经排得很近。第三方大模型评测机构Artificial Analysis于10月27日发布的智能指数v5显示，最高思考档位下，Claude Opus 5.5获得58分，GPT-6 Astra为57分，仅差1分；GPT-6.1 Sol则获得55分。
▲Claude Opus 5.5与GPT-6 Astra的智能指数成绩相差1分（图源：Artificial Analysis）
在同一评测的真实代码任务测试FrontierCode 2.0中，Claude仍然领先。在最高思考档位下，Opus 5.5成绩为46%，GPT-6 Astra和GPT-6.1 Terra均为42%，相差4个百分点。
▲FrontierCode 2.0测试中，Opus 5.5获得46%，GPT-6 Astra和GPT-6.1 Terra均为42%（图源：Artificial Analysis）
二、OpenAI模型报价低至六折，银行客户每账户月费约6.67美元
为了争取企业订单，OpenAI给出多项优惠：模型报价低至六折，提供免费额度，有的老客户甚至拿到了账户翻倍、年费不变的条件。
首先是大额模型折扣。
据The Information报道，两家公司都会向大客户提供优惠，但Anthropic会在客户达到合同用量上限后终止折扣。
总部位于美国的职业培训公司DataCamp，此前在Anthropic模型上投入较多。据The Information报道，DataCamp CEO乔纳森·科内利森称，OpenAI愿意对其内部产品开发所用模型提供40%的折扣。
不过，想拿六折优惠，编程工具就得“换门庭”：DataCamp须采用Codex，替代Claude Code、OpenCode等竞品。
截至该报道发布时，DataCamp仍在考虑这一报价，并与OpenAI协商，希望面向客户的应用也能享受同样的折扣。科内利森称，OpenAI模型的原价就低于Anthropic，若完成切换并获得折扣，相关成本预计可降低80%至90%。
其次是提供免费额度。
据The Information报道，美国IT支持自动化公司Atomicwork CEO维杰·拉亚帕蒂称，公司获得OpenAI及美国投资机构Khosla Ventures提供的免费额度后，开始培训新工程师使用Codex，替代Claude Code。另据该媒体援引知情人士消息，Khosla Ventures是OpenAI的重要投资方，会向所投创企提供这类额度。
拉亚帕蒂预计，截至2027财年的第一个月，Atomicwork在Codex上的支出接近30万美元（约合人民币201万元），上一财年则几乎为零；同期Claude Code支出预计约为200万美元（约合人民币1343万元）。
免费额度之外，OpenAI还向部分客户提供工程师支持。据The Information报道，一家获得美国风投机构Andreessen Horowitz投资的创企，与OpenAI工程师、销售人员建有Slack群，帮助解决Codex使用问题。该公司与Anthropic销售人员的类似沟通群中，则没有工程师参与。
第三，在合同费用不变的基础上，给已有的ChatGPT客户增加账户。
据The Information援引一家美国银行高管消息，该行营销团队使用ChatGPT生成广告，HR团队用其撰写职位说明。该行原本开始考虑将部分工作转交Claude，OpenAI随后允许该行将ChatGPT账户数量翻倍至500个，年度合同费用仍为约4万美元（约合人民币26.86万元）。
折算下来，每个账户每月约6.67美元（约合人民币45元），远低于通常每人每月超过20美元（约合人民币134元）的价格。这项优惠让该行继续使用ChatGPT。
三、Anthropic下调新模型API价格，Claude重回OpenRouter支出榜首
Anthropic在部分企业合同谈判中坚持采购门槛，但也在新模型发布时下调了公开API价格。
9月22日发布的Opus 5.5，每百万Token输入、输出价格分别为4美元、20美元（约合人民币26.86元、134.29元），较Opus 5均降低20%，缓存读取价格降低60%。
Opus 5.5上线后，Anthropic重新成为美国模型聚合平台OpenRouter上客户支出最多的模型供应商。据The Information报道，OpenAI此前曾超过Anthropic；在该报道发布前的一周左右，平台客户在Anthropic模型上的支出约为OpenAI的两倍。
到了10月底，两家公司又进一步加码订阅权益。
10月29日，OpenAI将Pro 100、Pro 500套餐的Work和Codex额度翻倍，并取消Pro 200此前计划的额度削减。Pro 200原本额度较低的用户，用量提升至此前的两倍；已有较高额度的用户，额度保持不变。
OpenAI还宣布，12月起，Plus、Pro和Business套餐的ChatGPT与Codex将共享使用额度，并按Token计量。额度消耗取决于所选模型和任务，Instant聊天不扣除这一额度。
Anthropic则将原Max 20x套餐更名为Ultra，取消了已验证Ultra账户的5小时用量限制，月费仍为200美元（约合人民币1343元），保留周限额。
结语：模型能力差距缩小，企业大客户更有议价权
OpenAI凭折扣争取到Replit的大单，Anthropic则在更新模型后重新取得OpenRouter客户支出领先。两组变化表明，企业采购既看模型效果，也看实际报价，领先厂商同样需要反复争取客户预算。
随着模型能力差距缩小、切换供应商更方便，企业大客户有了更多议价空间。模型效果之外，合同折扣、免费额度和技术支持也在影响采购选择。对模型厂商而言，竞争对手一旦推出效果接近、价格更低的产品，原有客户就可能转向。技术领先带来的高价，越来越难长期维持。
来源：The Information、路透社、Anthropic、OpenAI
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。
Intellectual (WeChat Official Account:zhidxcom)
Compiled by Sun Xiaoxuan
Editor, Concubine.
OpenAi and Anthropic have been fighting for the user!
On 30 October, Open AI just doubled ChatGPT Work and CodexZ81Z for part of the Pro package; the following day, Anthropic removed the five-hour limit for certified Ultra accounts and retained the weekly limit. On November 1st, Open AI also warned users in the help document about the integration of the Work function into regular chats and planned to share quota with Codex in ChatGPT from December for the Plus, Pro and Business packages.
Subscriptions were added, while business orders worth hundreds of millions of dollars a year were targeted.
According to Jiji, on November 2, United States technology media The Information reported on October 24 that after Open AI provided more than 30 per cent model price discounts, the US AI programming platform Refrit shifted most of AI expenditure from Anthropic to Open AI. According to sources quoted by the media, the contract with Open AI was worth over $100 million per year (approximately RMB 671 million).
In order to rob clients, Open AI's model offers are as low as 60% off, delivers free quota and allows engineers to enter the group of “close services”. As soon as a bank started considering the trial of Claude, Open AI introduced the condition of doubling the ChatGPT account at constant annual fees.
Currently, many of the older clients of Anthropic, including the American AI programming platform Reflit, the Human Resources Software company Ripping and the IT Support Automation Corporation Atomicwork, have more or less chosen “to abandon Anthropic for Open AI”.
According to the Information sources, Refrit had previously procured model services mainly from Anthropic and had intended to continue its cooperation. However, in the contract negotiations, Anthropic asked Replit to commit to higher procurement expenditures in order to be willing to continue the previous discount.
At the same time, Open AI offered model price discounts of over 30 per cent. Refrit eventually accepted Open AI ' s offer. According to the above-mentioned sources, the annual value of the contracts of the parties exceeds $100 million (approximately RMB 671 million). Reflit completed the code changes required to switch from Anthropic to Open AI in just a few hours, and the product adjustments took several days, less time than before to negotiate the contract with Anthropic.
The United States human resources software company Ripping also increased the proportion of Open AI models used in programming. The company ' s engineers had previously used both the Anthropic and Open AI models, but relied mainly on Anthropic.
According to The Information, in May this year, in the programming work of hundreds of Rippling engineers, the Anthropic model carried 57 per cent and the Open AI model 16 per cent; by 24 October, when publish was reported, the share of the Anthropic model fell to 18 per cent and the share of the Open AI model increased to 52 per cent.
Open AI’s customer-raiding strategy has begun to bear fruit. In addition to discounts, what other methods does Open AI use to obtain orders? What about Anthropic?
One: Anthropic takes a quarter against Open AI, with only one point off of head model intelligence.
According to The Information, for the first time in the second quarter of this year, Anthropic won more than Open AI. Anthropic publicly submitted IPO registration documents on 26 October show that the company received approximately $11,555 million (approximately $77,587 million RMB) during the season; Reuters previously disclosed that Open AI earned $6.7 billion (approximately $44,988 million RMB) for the same period, but Reuters also stated that the income calculus of the two companies was not fully consistent and that Open AI did not include some cloud partner sales income.
Anthropic quarterly operating data received $11,555 million in the second quarter of 2026 (source: Anthropic IPO registration document)
According to the Information analysis, the performance of the programming model gave Anthropic a strong pricing capability. Anthropic's IPO registration document shows that the company's second quarter of its calculations and infrastructure expenditure was approximately $5,558 million (approximately RMB 37,320 million), less than half of the amount collected during the same period.
Open AI is different. According to The Information, Open AI had previously predicted that this year ' s computing expenditure would significantly exceed revenue. According to the media, significant discounts were helping Open AI to compete for clients, but would also increase its income and expenditure pressure.
However, several app developers interviewed by The Information stated that the new models launched by Open AI in recent months were more competitive with Anthropic than before.
According to the latest running scores, the head models of the two families are already close. The third party, Artificial Analysis of publish, on 27 October, showed that under the highest level of reflection, Claude Opus 5.5 scored 58 points, GPT-6 Astra was 57 points, only 1 point less; GPT-6.1 Sol received 55 points.
One point difference between Claude Opus 5.5 and GPT-6 Astra ' s intelligence index performance (source: Artificial Analysis)
Claude is still in the lead in FrontierCode 2.0 for the same evaluation. At the highest level of reflection, the achievement of Opus 5.5 was 46 per cent, and GPT-6 Astra and GPT-6.1 Terra were 42 per cent, a difference of 4 percentage points.
For the FrontierCode 2.0 test, Opus 5.5 received 46%, GPT-6 Astra and GPT-6.1 Terra, both 42% (source: Artificial Analysis)
ii. Open AI model offers are as low as 60% discounts, with bank customers paying approximately $6.67 per month for their accounts Dollar
In order to obtain an enterprise order, Open AI offered a number of advantages: the model offer was as low as a six-cent discount, provided free quota, and some older customers even received double-counting and constant annual fees.
First is the large model discount.
According to The Information, both companies offer concessions to large customers, but Anthropic terminates the discount once the customer reaches the contract ceiling.
The United States-based vocational training company, Datacamp, had previously invested more in the Anthropic model. According to The Information, DataCamp CEO Jonathan Cornelison stated that Open AI was willing to provide a 40 per cent discount on models for its internal product development.
However, in order to get a six-percent discount, the programming tool would have to be “replaced”: DataCamp would have to use Codex instead of Claude Code and OpenCode.
At the time of the coverage of publish, DataCamp was still considering the offer and, in consultation with Open AI, hoped to benefit from the same discount for client-oriented applications. Konelison states that the original cost of the Open AI model was lower than that of Anthropic, and that the associated costs were expected to be reduced by 80 to 90 per cent if the switch were completed and discounts were obtained.
The second is the provision of free quota.
According to the Information report, Atomicwork CEO Rajapati, an American IT support automation company, began training new engineers to use Codex as a replacement for Claude Code, after obtaining Z14Z from Open AI and the United States investment agency Khosla Ventures. The media also quoted sources that Khosla Ventures was an important investor in Open AI and offered such quota to the initiative.
By the first month of the 2027 fiscal year, Lajapati projected that Atomicwork spent nearly $0.3 million (approximately RMB 2.01 million) on Codex, compared with almost zero in the previous fiscal year; Claude Code spent approximately $2 million (approximately RMB 13.43 million) during the same period.
In addition to free quota, Open AI provided engineer support to some of its clients. According to The Information, an innovative venture that obtained the investment of the American wind projector Andressen Horowitz, with a Slack group of Open AI engineers and salesmen, helped solve the problem of Codex use. The company was not involved by engineers in a similar community of communication with Anthropic sales personnel.
Third, the existing ChatGPT clients were given additional accounts based on the constant cost of the contract.
According to The Information, a senior United States bank executive was quoted as using ChatGPT for advertising, which was used by the HR team to write job descriptions. The Bank had begun to consider transferring part of its work to Claude, and Open AI had subsequently allowed the Bank to double the number of ChatGPT accounts to 500, at an annual contract cost of approximately $40,000 (approximately RMB 2686).
Each account was converted to approximately $6.67 (approximately 45 yuan) per month, which is far below the normal price of over $20 (approximately 134 yuan) per person per month. The concession allowed the bank to continue using ChatGPT.
iii. Anthropic downgrades the API price and Claude returns to the top of OpenRouter's spending.
Anthropic insisted on the procurement threshold in contract negotiations for some enterprises, but also reduced the open API price at the time of the new model publish.
On 22 September, publish Opus 5.5, with an input and output price of $4 per million Token, $20 (approximately RMB 26.86 and RMB 134.29), was 20 per cent lower than Opus 5 and cache reading prices 60 per cent lower.
When Opus 5.5 went online, Anthropic re-established itself as the most expensive model supplier on the United States model polymer platform OpenRouter. According to The Information, Open AI had previously exceeded Anthropic; about the week before the report publish, platform clients spent about twice as much on the Anthropic model as Open AI.
By the end of October, two companies had further added subscriptions.
On 29 October, Open AI doubled the work of the Pro 100, Pro 500 sets and Codexquota and cancelled the quota cuts that Pro 200 had previously planned. Pro 200 original users of quota were twice as high as previously; there are already higher users of quota, and quota remained the same.
Open AI also announced that, starting in December, ChatGPT from the Plus, Pro and Business packages would be shared with Codex and measured by Token. quota consumption depends on the selected model and tasks, and the Instant chat does not deduct this quota.
Anthropic changed the original Max 20x set to Ultra, removed the five-hour limit on the certified Ultra account and remained at a monthly fee of $200 (approximately RMB 1343) and retained the weekly limit.
Conclusion: Reduced modelling capacity gaps and greater bargaining power for large corporate clients Rights
Open AI won a large bill for Reprit at a discount, while Anthropic, after updating the model, regained the lead in OpenRouter customer spending. The two sets of changes suggest that firm procurement depends on both model effects and actual offers, and that leading manufacturers also need to press for client budgets repeatedly.
As model capacity gaps narrow and suppliers are more easily switched, large corporate clients have more bargaining space. In addition to the model effects, contract discounts, free quota and technical support are affecting procurement choices. For model producers, the original customer may shift once the counterparty has introduced products that are nearer and cheaper. The high prices associated with technological leadership are becoming increasingly difficult to sustain in the long term.
Source: The Information, Reuters, Anthropic, Open AI
Auto-fetched and translated by AI Atlas. All rights belong to the original author.