AI Cost Control Alert: Major Companies Imposing Traffic Limitation Orders, Flagship Models Have Become Luxury Items

A chilling industry trend of "uncontrolled AI costs" is spreading through Silicon Valley and the global tech community. From Atlassian and Adobe to Amazon, more and more top companies are starting to slow down internal AI usage, even banning employees from using the most advanced flagship models, and forcing a switch to lower-cost alternatives.

This collective action was triggered by a drastic change in AI providers' billing models. As companies move from fixed annual fees to expensive pay-as-you-go pricing, AI call costs have skyrocketed. According to leaked internal data, at least one company's monthly AI expenses have surged to three times the original amount, with monthly bills exceeding $15 million. Faced with this heavy financial pressure, companies are now re-evaluating the cost of blindly expanding AI applications.

Amid the wave of refined operations, major companies have adopted different strategies. Citibank took the most direct and firm approach: banning flagship models such as GPT-5.5, Claude Opus 4.6, and 4.7, and requiring employees to "use according to needs." Within Citibank, high-performance models are considered extremely valuable resources, and their use is strictly restricted unless absolutely necessary. Atlassian, on the other hand, introduced a cost dashboard, allowing every employee to clearly see how much money each of their "Prompts" costs. This transparency has proven effective but has also sparked collective anxiety among employees about declining work efficiency.

In contrast, GitHub has a more forward-thinking approach. They plan to shift AI quotas from "department-wide sharing" to "individual pay-as-you-go," and actively transition to open-source models, aiming to find a new balance between performance and cost. Meanwhile, companies like Adobe are no longer renewing unlimited usage agreements, giving employees a final transition period.

Industry insiders point out that this contraction trend driven by "cost economics" marks the end of the wild growth phase of the AI industry. The previous strategy of "investing without considering costs" is no longer effective. As revealed by an internal recording from Accenture, when AI is heavily used for non-core tasks such as generating PowerPoint presentations or predicting World Cup results, the inflated bubble it creates will eventually burst.

As major companies build up computing power "walls," AI developers may need to adapt to a new reality: in the future, high-performance large models will no longer be readily available general tools, but rather expensive assets that require careful budgeting and on-demand allocation. This industry "de-bubble" movement is not only a reallocation of computing resources but also a deep reshaping of the rational return of AI's commercial value.

OpenAI Upgrades GPT-5.5 Again: Better at Mind Reading, Shopping Recommendations Are More Understandable

OpenAI upgrades the GPT-5.5 Instant model, significantly enhancing intent recognition and capturing implied meanings, making complex task handling more stable. The model was released on May 5th, with a 52.5% reduction in hallucination rates in high-risk areas. Improvements in mathematics, science, and visual reasoning are evident. The update on May 30th further optimized the naturalness of responses.

Volcano Engine Launches Doubao Audio Generation Model 1.0: Generate Cinematic-Quality Audio with One Sentence, Character Voices Stay Consistent for 10 Minutes

Volcano Engine launches Doubao Audio Generation Model 1.0, which supports text or audio input and generates complete audio works end-to-end. The core breakthrough is that a single Prompt can simultaneously generate dialogue, sound effects, and background music, eliminating the need for traditional multi-track editing. This technology greatly simplifies the audio production process, allowing users to efficiently produce professional-quality audio like an 'audio director,' completely eliminating the complex post-production work of manual alignment and mixing.

Wikimedia Founder Says No to AI: The 'Hallucination' Problem Remains Severe, Editing Work Still Needs Humans

Jimmy Wales, co-founder of Wikipedia, stated that the current 'hallucination' problem of AI remains serious, as it often outputs false information despite appearing confident, and is still not trustworthy. Therefore, he has banned AI from directly editing entries. However, he believes AI can serve as an alarm clock, providing reminders and assistance.

Volc Engine Launches Doubao Seedance 2.5 Video Model, the Azuris Large Model Service Has Over 1.1 Million Users

On June 23, 2026, Volcano Engine launched its Seedance2.5 video generation model at the Summer FORCE Conference, set to release in July. It offers three key breakthroughs: direct generation of 30-second native videos, joint generation from up to 50 multimodal assets, and local editing with visual consistency. President Tan Dai noted that video generation is crucial to world models.....

Amazon Launches Beta Testing of Generative AI Assistant Alexa+ in Hindi in India

Amazon has launched the beta testing of the Hindi version of Alexa+ in India. It invites selected users by email to fill out a form before June 22 to optimize the localized experience. This product is a generative AI launched in 2025. After gradual promotion in the United States, it was fully opened in February of this year and has now expanded to markets including the UK, Canada, Brazil, Mexico, Italy, and Germany.