AI Programming Rate Limiting DeepSeek-V4 Experience | Momei Blog | 2026 Week 19 Strawberry Weekly Report

草梅友仁 — Sun, 10 May 2026 14:05:33 GMT

This article is published and updated on CaoMeiYouRen’s Blog and synchronized across multiple platforms. For updates, please refer to the version on the blog. You can also check the latest version via the Original Link at the end of the article.

Preface

Welcome to the CaoMei Weekly! This is a weekly digest curated by CaoMeiYouRen with the help of AI, aiming to provide you with the latest blog updates, GitHub activities, personal updates, and recommendations from other weekly publications.

Open-Source Updates

Development of Momei Blog continues steadily this week.

You can try the demo site here: https://demo.momei.app/

Use the demo admin account with email admin@example.com and password momei123456.

Or register on the official website: https://momei.app/

You can also visit the documentation site to learn about the project’s overall plan and future roadmap: https://docs.momei.app/

Momei Blog has officially released version 1.16.0. Below are some screenshots of the new pages and features.

Added creator statistics functionality—now you can view individual creator stats in the backend.

New remote repository synchronization feature allows syncing blogs to specified GitHub/Gitee repositories (using Hexo-style markdown files).

For more features and pages, visit the official website or check previous blog posts for additional screenshots.

We welcome all users to try it out and share your feedback and suggestions.

Development will continue according to the roadmap and backlog. Stay tuned for more updates!

Certainly, the Momo Blog still has many details to refine, and its features are not yet fully developed. If you have any suggestions or feedback, feel free to raise them in the project’s GitHub issues.

If you’re also interested in Momo Blog, you’re welcome to participate in its development and testing.

AI Coding Plans Are Widely Implementing Rate Limits

Recently, while using GitHub Copilot, I’ve frequently encountered session limits and weekly limits. This means that even with a paid Copilot Pro subscription, rate limiting is unavoidable.

The only issue is that these limits don’t account for users who pay extra. Previously, when exceeding the quota, I could simply purchase additional usage—for example, in March, I spent an extra $10, and my usage even reached 200% (nearly 100% beyond the original limit).

Now, however, this extra paid usage is also counted toward the rate limit, leading to an absurd situation where even the included quota might not be fully utilized (whereas before, usage would likely have already exceeded 80%).

Although GitHub Copilot will switch to usage-based billing starting next month (see: GitHub Copilot Is Moving to Usage-Based Billing), could the user experience be slightly improved during this final transition period?

Of course, the throttling of Coding Plans isn’t an issue unique to GitHub Copilot. Whether it’s international players like Claude, OpenAI, Gemini, or domestic AI model providers like GLM and MiniMax, all have started imposing limits on Coding Plans without exception.

Some have gone as far as setting extremely low usage caps, even at the cost of unilaterally modifying subscription terms and angering users.

For example, GLM outright discontinued old plans and forced users to switch to new ones. (Reference: Old Plan Migration and Compensation Notice)

Meanwhile, Alibaba Cloud went further by completely shutting down Coding Plan Lite, making it impossible to renew or upgrade. (Reference: Notice on Discontinuation of Coding Plan Lite Renewals and Upgrades)

So before getting upset with these AI providers, it’s worth considering a new question: Why are they all uniformly throttling Coding Plans?

The reason is simpler than you might think—providing subscription services for large AI models is actually a capital-intensive business.

Where does the heaviest cost lie? In computing power, or more specifically, GPUs. The marginal cost reduction here is minimal—meaning that for every additional user, the provider must supply corresponding computing resources. In contrast, traditional internet services benefit from significant economies of scale: serving the first user requires substantial resources, but by the 10,000th user, the incremental cost per user is much lower.

With Coding Plans, however, the computing power required for the first user is nearly the same as for the 10,000th, making it difficult to reduce costs significantly.

Moreover, high-usage users can truly push plans to their limits. The more subscribers there are, the more such users emerge, driving up computing demands and keeping costs high.

This is why more and more AI providers are cutting back on plan allowances.

There is another similar issue that has recently come to light. Many have heard about the decline in performance of the once-popular video-generation large model, Seedance2.0.

Although it’s still possible to access a “full-powered” version through some paid subscriptions, the experience is noticeably less impressive compared to its initial release.

This highlights one of the biggest challenges faced by large AI models: when a new model first emerges, computational resources are relatively abundant, leading to a better user experience. However, as more users are drawn in, computational resources become strained, the user experience deteriorates, and eventually, users start leaving—until the next major model arrives, repeating the cycle.

This pattern holds true whether we’re talking about text-based large models, image-generation models, or video-generation models.

As a result, AI models that rely solely on brute-force computational power for high performance are unsustainable. The reason is that performance gains from increased computational resources are roughly linear, whereas the new demands generated by users grow exponentially. Therefore, the improvement in AI model performance can hardly keep up with the surge in new demands.

That’s why, even though DeepSeek itself evaluates DeepSeek-V4-Pro as still having a certain gap in reasoning capability compared to Opus 4.6, I remain more optimistic about DeepSeek’s approach.

DeepSeek has consistently strived to develop AI models that deliver better performance while requiring less computational power. I believe this is the future of large AI models.

(It would be even better if DeepSeek could offer a cost-effective DeepSeek-V4-Pro subscription plan in the future.)

Experience Using DeepSeek-V4

During this period, I’ve personally used DeepSeek-V4-Pro and DeepSeek-V4-Flash extensively for programming, consuming approximately 550 million tokens and spending around 45 RMB. On average, this translates to about 12 million tokens per RMB.

This includes some usage of V4-Flash, but since the volume was minimal, it was merged into the overall calculation.

The reason it can be so cost-effective lies in DeepSeek’s mysterious caching mechanism, which ensures a high cache hit rate. In programming—a context where repetitive content is common—this proves to be an excellent cost-saving strategy.

Of course, it must be acknowledged that there is still a significant gap between DeepSeek-V4-Pro and the current top-tier AI models, such as GPT-5.5 and Claude Opus 4.6. This became evident to me after practical use, where I noticed considerable oversights in the coding outputs of DeepSeek-V4-Pro. So, expecting DeepSeek-V4-Pro to catch up to the leading AI models right away is still unrealistic.

However, I remain optimistic about DeepSeek—after all, at this price point, what more could you ask for?

Just look at how expensive Anthropic’s Claude API is: Opus 4.7 charges $5 (~34 RMB) per million tokens for input and a staggering $25 (~170 RMB) per million tokens for output.

As for subscription plans, frequent account suspensions and stricter KYC verification have made usage increasingly cumbersome, driving costs up rapidly.

So, while in theory, one should use the most powerful AI model to maximize productivity and offset rising expenses with higher revenue, in reality, no one can afford to ignore costs.

As the saying goes, the rich have their ways, and the poor have theirs. What matters most is tailoring the most cost-effective solution to your actual needs.

GitHub Release

rss-impact-server

v1.18.2 - 2026-05-02 20:28:53

Summary:

v1.18.2 Release Notes

Testing: Added unit tests for the caching service.
Bug Fixes: Updated dependencies and optimized cache management implementation.

momei

v1.16.0 - 2026-05-09 20:42:55

Summary:

1.16.0 (2026-05-09)

✨ New Features

admin:

v1.15.0 - 2026-04-25 20:36:49

Summary:
v1.15.0 Release Notes:

New Features:

Admin panel AI module now supports task detail loading
API enhancements for type safety and error handling
Comment system adds translation functionality and cross-language prompts
Friend links module now includes localized support for meta-info interfaces, along with friend link management and internationalization
Code governance: Completed initial legacy code comment cleanup and standardized JSDoc templates; Added ESLint/type debt governance documentation and configurations; Added comment drift governance documentation
Added Hexo-style article repository synchronization and Hexo repository sync functionality
Internationalization: Enhanced auditing features; Improved admin-posts module translations; Updated multi-language support and management interface text
Markdown editor now supports editor features and code copying
Paginated queries now include secure parsing
Article list queries now support language fallback filters
SEO: Added page SEO-related features and structured data with SEO optimizations for tag pages
Others: Added AI-powered comment translation; Unified string list parsing logic

Bug Fixes:

Article management module now includes management features and internationalization support
Email templates now support runtime fields
External Feed panel optimized test case structure
External Feed now handles items without titles
Completed second round of ESLint/type debt tightening
Internationalization: Optimized repository sync info localization; Added internationalization support for user agreement management
Settings module improved field parsing logic and added external source editor functionality

Code Refactoring:

AI module standardized trailing slash handling logic
ASR module now uses the toWebSocketBuffer function for data transmission
Database now includes readiness checks
Governance: Refactored Hexo repository sync into remote repository sync
Internationalization refactored dynamic key handling logic
Settings API enhanced type handling logic
Third-party settings optimized external source model binding
Others: Improved effect mounting conditions

picgo-plugin-optimization

v1.1.1 - 2026-05-10 00:30:35

Summary:
Version 1.1.1 (2026-05-09):

Fix: Refactored tsdown dependency configuration.
Refactor: Modified neverBundle dependencies for tsdown, added picgo and axios.

Latest GitHub Starred Repositories

CaoMeiYouRen starred SoftwareCopyright-Skill - 2026-05-10 18:00:14
This tool, named “China Software Copyright Application Material Generator,” is developed in Python. It automatically generates complete Word-format copyright application materials by reading local projects. The project is fully open-source, eliminating the need for additional paid application services, and has received 989 stars on GitHub.
CaoMeiYouRen starred LaughterCan - 2026-05-05 19:26:10
This repository has no description. The primary language is Python, and it has received 5 stars.
CaoMeiYouRen starred awesome-deepseek-agent - 2026-05-04 12:13:27
This GitHub repository has no description, and the primary language is unspecified. It has received 1,022 stars.
CaoMeiYouRen starred deepseek_v4_rolepaly_instruct - 2026-04-30 23:25:30
This is a special control instruction guide for DeepSeek-V4 role-playing. The primary language is specified, and the repository has 1,790 stargazers.
CaoMeiYouRen starred hermes-agent - 2026-04-28 11:13:15
An agent that grows with you. The primary programming language is Python, and it has received 141,827 stars.