OpenAI Uses Over 30% Discounts and Free Credits to Poach Anthropic Customers; Replit Annual Contract Exceeds $100 Million

OpenAI and Anthropic have clashed again recently in a bid to grab users!

On October 30, OpenAI had just doubled the ChatGPT Work and Codex quotas for some Pro plans; the next day, Anthropic removed the 5-hour usage limit for verified Ultra accounts, while keeping the weekly limit. On November 1, OpenAI again announced to users in its help documentation that it would integrate the Work feature into regular chat, and planned starting in December to let ChatGPT and Codex share usage quotas across the Plus, Pro, and Business plans.

Zhidongxi November 2 news, according to a October 24 report by the US tech media outlet The Information, after OpenAI offered model price discounts of more than 30%, the US AI coding platform Replit shifted most of its AI spending from Anthropic to OpenAI. According to the outlet, citing people familiar with the matter, the annual value of Replit's contract with OpenAI exceeds $100 million (about 671 million yuan).

To win customers, OpenAI's model quotes were as low as 60% of list price, it offered free quotas, and it sent engineers into group chats for "hands-on service." A bank had just begun considering trialing Claude when OpenAI offered a deal doubling ChatGPT accounts with the annual fee unchanged.

Currently, multiple longtime Anthropic customers, including the US AI coding platform Replit, the HR software company Rippling, and the IT support automation company Atomicwork, have more or less chosen to "abandon Anthropic for OpenAI."

According to The Information, citing people familiar with the matter, Replit had previously mainly purchased model services from Anthropic and originally intended to continue the cooperation. But during contract negotiations, Anthropic required Replit to commit to higher procurement spending before it was willing to extend the previous discount.

At the same time, OpenAI offered model price discounts of more than 30%. Replit ultimately accepted OpenAI's offer. According to the people familiar with the matter, the annual value of the contract between the two sides exceeds $100 million (about 671 million yuan). It took Replit only a few hours to complete the code changes needed to switch from Anthropic to OpenAI, and product adjustments took a few days—shorter than the time previously spent negotiating the contract with Anthropic.

The US HR software company Rippling also increased the share of OpenAI models used in coding work. The company's engineers had previously used both Anthropic and OpenAI models, but relied mainly on Anthropic.

According to The Information, in May this year, among the coding work of hundreds of Rippling engineers, Anthropic models accounted for 57% and OpenAI models for 16%; as of the publication of that report on October 24, Anthropic models' share had fallen to 18% and OpenAI models' share had risen to 52%.

OpenAI's customer-grabbing strategy has begun to show results. Besides discounts, what other methods has OpenAI used to win orders? And how is Anthropic responding?

1. Anthropic's quarterly revenue overtakes OpenAI; top models' intelligence index differs by only 1 point

According to The Information, in the second quarter of this year, Anthropic's revenue exceeded OpenAI's for the first time. Anthropic's IPO registration filing, publicly submitted on October 26, shows that the company's revenue for the quarter was about $11.555 billion (about 77.587 billion yuan); Reuters previously disclosed that OpenAI's revenue for the same period was $6.7 billion (about 44.988 billion yuan), but Reuters also noted that the two companies' revenue accounting standards are not entirely consistent, and OpenAI did not include some cloud partner sales revenue.

▲Anthropic quarterly operating data; second-quarter 2026 revenue was $11.555 billion (Source: Anthropic IPO registration filing)

According to The Information's analysis, the performance of coding models has given Anthropic relatively strong pricing power. Anthropic's IPO registration filing shows that the company's compute and infrastructure spending in the second quarter was about $5.558 billion (about 37.320 billion yuan), less than half of its revenue for the same period.

OpenAI's situation is different. According to The Information, OpenAI had previously expected that its compute spending this year would significantly exceed its revenue. The outlet believes that large discounts are helping OpenAI win customers, but will also increase pressure on its income and expenditure.

However, multiple application developers interviewed by The Information said that the new models OpenAI launched in recent months are more able to compete with Anthropic than before.

Judging from the latest benchmark scores, the two companies' flagship models are already ranked very close to each other. The intelligence index v5 released by third-party large-model evaluation organization Artificial Analysis on October 27 shows that at the highest thinking level, Claude Opus 5.5 scored 58, GPT-6 Astra scored 57, a gap of just 1 point; GPT-6.1 Sol scored 55.

▲Claude Opus 5.5 and GPT-6 Astra differ by 1 point on the intelligence index (Image source: Artificial Analysis)

In FrontierCode 2.0, the real-world coding task test of the same evaluation, Claude still leads. At the highest thinking level, Opus 5.5 scored 46%, while GPT-6 Astra and GPT-6.1 Terra both scored 42%, a difference of 4 percentage points.

▲In the FrontierCode 2.0 test, Opus 5.5 scored 46%, while GPT-6 Astra and GPT-6.1 Terra both scored 42% (Image source: Artificial Analysis)

2. OpenAI model quotes as low as 60% of list price; bank customer pays about $6.67 per account per month

To win enterprise orders, OpenAI has offered multiple concessions: model quotes as low as 60% of list price, free quotas, and some old customers have even received terms with doubled accounts at an unchanged annual fee.

First, large-volume model discounts.

According to The Information, both companies offer concessions to major customers, but Anthropic terminates discounts after customers reach their contracted usage limits.

DataCamp, a US-based vocational training company, had previously invested heavily in Anthropic models. According to The Information, DataCamp CEO Jonathan Cornelissen said OpenAI was willing to offer a 40% discount on the models used for its internal product development.

However, to get the 60%-of-list-price discount, the coding tool has to "switch allegiance": DataCamp must adopt Codex to replace competitors such as Claude Code and OpenCode.

As of the publication of that report, DataCamp was still considering the offer and negotiating with OpenAI, hoping that customer-facing applications could enjoy the same discount. Cornelissen said OpenAI's models are already cheaper than Anthropic's at list price, and if the switch is completed and the discount obtained, related costs are expected to fall by 80% to 90%.

Second, free quotas.

According to The Information, Vijay Rayapati, CEO of US IT support automation company Atomicwork, said that after the company received free quotas provided by OpenAI and US investment firm Khosla Ventures, it began training new engineers to use Codex instead of Claude Code. According to the outlet, citing people familiar with the matter, Khosla Ventures is a major investor in OpenAI and provides such quotas to startups it backs.

Rayapati expects that by the first month of fiscal year 2027, Atomicwork's spending on Codex will be close to $300,000 (about 2.01 million yuan), compared with almost zero in the previous fiscal year; over the same period, Claude Code spending is expected to be about $2 million (about 13.43 million yuan).

In addition to free quotas, OpenAI also provides engineer support to some customers. According to The Information, a startup backed by US venture capital firm Andreessen Horowitz has a Slack group with OpenAI engineers and sales staff to help resolve Codex usage issues. In a similar communication group the company has with Anthropic sales staff, no engineers participate.

Third, adding accounts for existing ChatGPT customers while keeping contract fees unchanged.

According to The Information, citing an executive at a US bank, the bank's marketing team uses ChatGPT to generate advertisements, and its HR team uses it to write job descriptions. The bank had initially begun considering handing some work to Claude, after which OpenAI allowed the bank to double its number of ChatGPT accounts to 500, while the annual contract fee remained about $40,000 (about 268,600 yuan).

Converted, that is about $6.67 per account per month (about 45 yuan), far below the usual price of more than $20 per person per month (about 134 yuan). This concession led the bank to continue using ChatGPT.

3. Anthropic cuts API prices for new models; Claude returns to the top of OpenRouter spending

Anthropic insists on procurement thresholds in some enterprise contract negotiations, but it has also lowered public API prices when releasing new models.

Opus 5.5, released on September 22, is priced at $4 and $20 per million tokens for input and output, respectively (about 26.86 yuan and 134.29 yuan), both 20% lower than Opus 5, with cache read prices down 60%.

After Opus 5.5 launched, Anthropic again became the model provider with the most customer spending on US model aggregation platform OpenRouter. According to The Information, OpenAI had previously surpassed Anthropic; in the week or so before that report was published, platform customers' spending on Anthropic models was about twice that on OpenAI.

On October 29, OpenAI doubled the Work and Codex quotas for the Pro 100 and Pro 500 plans, and canceled the previously planned quota reduction for Pro 200. For Pro 200 users who originally had lower quotas, usage was raised to twice the previous level; users who already had higher quotas kept their quotas unchanged.

OpenAI also announced that starting in December, ChatGPT and Codex in the Plus, Pro and Business plans will share usage quotas, measured by tokens. Quota consumption depends on the selected model and task, and Instant chat does not deduct from this quota.

Anthropic, meanwhile, renamed its original Max 20x plan to Ultra, removed the 5-hour usage limit for verified Ultra accounts, kept the monthly fee at $200 (about 1,343 yuan), and retained the weekly limit.

Conclusion: As the model capability gap narrows, large enterprise customers have more bargaining power

OpenAI won Replit's big order with discounts, while Anthropic regained the lead in OpenRouter customer spending after updating its models. The two sets of changes show that enterprise procurement looks at both model performance and actual quotes, and leading vendors likewise need to repeatedly compete for customer budgets.

As the model capability gap narrows and switching suppliers becomes easier, large enterprise customers have more room to bargain. Beyond model performance, contract discounts, free quotas and technical support are also influencing procurement choices. For model vendors, once a competitor launches a product with similar performance and a lower price, existing customers may switch. The high prices brought by technological leadership are increasingly difficult to sustain over the long term.

Sources: The Information, Reuters, Anthropic, OpenAI