AI Daily: Claude Cross-Platform Upgrade, Meta Media Generation Model Debut, Microsoft Lowers Copilot Costs, Chinese Language Model Market Share Growth

AI Daily: Claude Cross-Platform Upgrade, Meta Media Generation Model Debut, Microsoft Lowers Copilot Costs, Chinese Language Model Market Share Growth

Did you know? Many fresh things happen in the field of artificial intelligence every day. From the strategic adjustments of major tech giants to the emergence of new open-source models, these technologies are quietly changing the way you and I work. Today, I have compiled the latest AI development trends for everyone. Let’s take a look at these exciting advances together.

Anthropic Extends Claude Fable 5 Promotional Access to July 12

Honestly, who doesn’t love experiencing the latest technology for free? Anthropic has heard the users’ voice. They have officially announced that the promotional experience period for Claude Fable 5 has been extended to July 12, 2026. This means existing paid subscribers have more breathing room to fully test this powerful new model.

According to the official Anthropic explanation, Pro, Max, Team, and Enterprise users with advanced seats can all use 50% of their weekly quota to use Claude Fable 5 for free during this period. Once users hit this 50% threshold, the system won’t immediately interrupt the service. You can choose to use prepaid credits to continue enjoying the powerful features of Fable 5, or easily switch back to other Claude models and continue consuming your original subscription quota.

This flexible mechanism is very helpful. Enterprises and development teams don’t have to worry about unexpected extra costs, and they can fully evaluate whether this new model meets the team’s daily work needs. If your team hasn’t tried it yet, now is definitely a good time.

Gemini API Expands Managed Agent Capabilities: Supports Background Execution and Remote MCP

Developers often encounter the annoyance of HTTP connection timeouts when handling long-running tasks. This is indeed a headache. To solve this pain point, Google has added brand-new managed agent capabilities to the Gemini API.

Through this update, developers can now start asynchronous background tasks. The system will immediately return a task ID, and the client only needs to query the progress through this ID at any time, without needing to maintain a constant connection. This approach significantly improves system stability. Furthermore, Google has added the integration capability for Remote Model Context Protocol (MCP) servers this time. Developers can directly connect agents to private databases or internal APIs, allowing AI to communicate with external tools in a secure sandboxed environment.

If you are interested in these new features, you can check out the Google Developers Blog for more practical examples. With the addition of custom function calls and credential refresh mechanisms, the practicality of the Gemini API has indeed improved significantly.

Claude Cowork Lands on Mobile and Web Versions, Taking Work With You

Imagine this scenario. In the office, you asked AI to organize a long financial report using your laptop, but you have to rush to a meeting immediately. In the past, as soon as you closed your laptop, this task might have been forced to interrupt.

The situation is completely different now. Claude Cowork has officially pushed forward to mobile and web versions, meaning your work can be taken with you anywhere. This update makes background processing a reality. You can set up a client presentation preparation task to run at 6:00 AM, and Claude will automatically comb through emails and news. By the time you drink your morning coffee, a complete draft is already waiting for you to review on your phone.

From the introduction on the official Claude blog, it can be seen that humans still hold the final decision-making power. When the model encounters steps that require manual confirmation, it will directly push the question to your mobile phone. To celebrate the launch of this feature, the official team has even doubled the usage limit until August 5th. This change truly makes AI a virtual colleague who never goes off duty.

Meta Launches Muse Image and Muse Video: Media Generation Models With Agent Capabilities

When mentioning image generation, people usually think of simple text-to-image functions. However, Meta Superintelligence Labs has brought a completely different way to play this time. They have officially released two brand-new media generation models: Muse Image and Muse Video.

What’s most special about Muse Image is that it operates like an agent. It not only faithfully follows complex instructions, but also searches for related facts online to ensure the generated images match reality. The model itself has self-correction capabilities; when it discovers a mistake in a corner of an image, it automatically performs partial modifications without needing humans to repeatedly issue instructions.

Even more anticipated is the preview of Muse Video. Based on the same training foundation, it can produce videos with extremely high visual fidelity and natively supports audio synthesis. Now, everyone can already try out Muse Image in the Meta AI application, Instagram Stories, and WhatsApp in some countries. For detailed technical demonstrations, you can refer to the official Meta announcement and their introduction on X. Meta has also added Content Seal watermarking technology this time to ensure that everyone can easily identify which images are AI-generated.

Introducing Muse Image and Muse Video, the first media generation models developed by Meta Superintelligence Labs.

Muse Image is our most advanced image generation model yet. It follows instructions faithfully, edits with precision, composes from multiple references, and draws… pic.twitter.com/byNpQZO1RW

— AI at Meta (@AIatMeta) July 7, 2026

Cohere Announces Transcribe Arabic: Open-Source Arabic Speech Recognition Model

Language diversity has always been a huge challenge in the field of speech recognition. Arabic has over 300 million native speakers and a wide variety of dialects, making it difficult for many mainstream AI systems to cope.

Cohere noticed this demand and launched a powerful open-source model: Cohere Transcribe Arabic. This model, based on a 2B parameter architecture, is specifically created to solve Arabic language problems in commercial and development environments. It can not only accurately recognize various dialects such as Gulf and Levantine, but also perfectly handle the phenomenon of “code-switching” where Arabic and English are mixed.

Judging from the test results published on the Cohere blog, its word error rate is only 25.87, significantly leading competing models like Whisper Large V3. This is definitely a huge boon for the Middle East market, which requires high-precision enterprise-level speech applications. Developers can now go to the Hugging Face model page to download weights and deploy them.

Claude Code Model Selection and Effort Level Setting Guide

When we use Claude Code for programming, we often face a difficult choice. Which model should we choose? And to what extent should the Effort Level be set?

Let me explain the difference between the two. According to official Claude documentation, model selection determines the AI’s “knowledge base” and “capacity limit.” This is like picking an expert with specific skills. Effort Level, on the other hand, controls how much time and energy the expert is willing to spend to handle your task.

When you raise the Effort Level, the model will read more files, perform more validation steps, and even try repeatedly when encountering errors. For simple daily code modifications, using a smaller model paired with the default Effort Level can save a lot of costs. When encountering architectural decisions or complex debugging tasks, switch to a larger model. This is a very practical resource allocation strategy.

Fundamentals and Applications of Agent Loops: A Claude Code Developer Guide

Besides choosing appropriate model parameters, how to effectively make agents operate automatically is also a subject of study. In the past, we were accustomed to using one-off prompts to direct AI. Now, the concept of “designing loops” is more popular in the development world.

This concept is actually quite intuitive. In a ClaudeDevs post, the team defines a loop as an agent repeatedly executing work cycles until a specific stopping condition is met. Developers can classify loops into several different types according to triggering methods and stopping conditions.

https://t.co/v694m8Eaj6

— ClaudeDevs (@ClaudeDevs) July 6, 2026

For example, “round-robin loops” are suitable for exploratory work, where you can set strict validation conditions to reduce the number of back-and-forth conversations. “Goal-oriented loops” will keep trying until a specific goal is achieved. There are also “time-driven loops,” which are very suitable for periodically summarizing Slack messages or checking code review status. Mastering these patterns allows Claude Code to help you automatically handle a large number of tedious daily tasks.

Artificial Analysis Launches Six Industry Capability Indices: Claude Fable 5 Wins

When evaluating language models, we often need to make judgments based on performance in specific fields. Artificial Analysis has launched a set of brand-new industry capability indices, covering six major fields: finance and accounting, law, healthcare, operational strategy, engineering technology, and economics.

The design of this scoring system is very close to real-world work needs. They extracted common tasks from O*NET job classifications and weighted them based on the frequency of task occurrence. According to the full report on X, Claude Fable 5 won the championship in all eight indices. This shows that it has a huge advantage in handling high-difficulty professional tasks.

Introducing six new Artificial Analysis Capability Indices for comparing model capabilities across key industry domains

The new industry indices cover Finance & Accounting, Legal, Healthcare & Medical, Strategy & Ops, Engineering, and Economics. We aim to capture the common… pic.twitter.com/hWcFv0eT5w

— Artificial Analysis (@ArtificialAnlys) July 7, 2026

Interestingly, the performance of open-source models is also quite stunning. GLM-5.2 took the top spot in five areas among open-source models. DeepSeek V4 Flash demonstrated an extreme cost advantage, with the cost of completing each task being less than $0.04. This lets us see that future enterprises can flexibly choose the most suitable AI tool based on budget and task complexity.

Reducing Costs: Microsoft Copilot Gradually Replaces OpenAI and Anthropic with Internal Models

The high cost of computing and API calls has always been a major concern for tech giants promoting AI services. Microsoft has recently begun to take action. They are gradually replacing OpenAI and Anthropic models with their self-developed MAI models in Copilot products such as Excel and Outlook.

According to a report by The Decoder, Microsoft’s internal models are already processing tens of thousands of requests per week. Microsoft AI head Mustafa Suleyman also frankly stated that their goal is to significantly reduce or even eliminate the massive fees paid to third parties.

This has triggered a phenomenon worth thinking about. Microsoft previously criticized vendor lock-in, but now they are turning to internal models to reduce costs. There is also news that future AI billing methods may shift to “pay-as-you-go.” If consumers want to use more powerful third-party models, they may need to pay extra for an upgrade.

US Companies Turn to Chinese AI Models: A New Trend Under Cost Considerations

The development cost of AI models is staggering, and usage costs cannot be underestimated either. When the model prices of top US labs such as OpenAI and Anthropic remain high, many US companies have begun to look for alternatives.

This trend of seeking high-cost-performance models has led to explosive growth in Chinese open-source models. According to data analyzed by CNBC, on the OpenRouter platform commonly used by developers, the ratio of tokens for US companies calling Chinese AI models has climbed to 46%.

Models like DeepSeek and the GLM 5.2 model developed by Z.ai not only keep up with leading US models in performance tests, but the price is also only one-tenth or even lower than competitors. When many daily tasks don’t need to use the top and most expensive models, these cheap and easy-to-use alternatives naturally become the favorites in the eyes of enterprises.

Nvidia Releases Nemotron-Labs-Audex-30B-A3B: A Powerful LLM Unifying Audio and Text

In the development of multimodal models, Nvidia is not falling behind either. Their latest released Nemotron-Labs-Audex-30B-A3B is a large language model that perfectly combines audio and text.

This model is built on the powerful Nemotron-Cascade-2 foundation and adopts a Mixture-of-Experts (MoE) architecture, with a total of 30 billion parameters, using only 3 billion parameters each time it starts. You can go to the Hugging Face introduction page to view detailed specifications. It not only expands the audio vocabulary table but also has a built-in audio encoder, capable of easily handling complex tasks such as speech recognition, speech translation, and text-to-speech.

This model even supports a context length of up to 1 million tokens. The development team also provides two operating modes, whether it’s a thinking mode requiring complex logical thinking or an instruction mode pursuing reaction speed, it can handle them freely. If your hardware resources are limited, they also simultaneously released a lightweight 2B version for developers to choose from.

Liquid AI Proposes FTPO Method: Effectively Reducing Language Model Inference Loops

Have you ever encountered this situation? When you ask AI to solve a complex math problem, it suddenly gets stuck, repeatedly outputting “Wait, let me rethink” on the screen, until it fills the entire chat window.

The industry calls this phenomenon the “Doom Loop.” This problem is particularly common when inference models process long pieces of reasoning. To solve this predicament, Liquid AI proposed a brand-new training method called Final Token Preference Optimization (FTPO). They jokingly call this method Antidoom.

Detailed technical principles can be found in the official Liquid AI blog. Traditional solutions usually involve adjusting repetition penalty parameters, but this often affects the overall performance of the model. Antidoom, on the other hand, precisely locks onto the specific token that caused the loop, training the model to choose other more reasonable continuation methods at that point in time. Experiments show that this method almost perfectly eliminates the phenomenon of repetitive loops in internal models and significantly improves various evaluation scores.

Q&A

Q1: What important updates and feature delays have occurred in the Claude ecosystem recently? A1: Anthropic announced that the free experience period for Claude Fable 5 has been extended to July 12, 2026, and paid subscribers can use 50% of their weekly quota to test it. In addition, Claude Cowork has officially landed on mobile and web versions, supporting “background processing” (such as automatically preparing client presentations in the early morning) without needing to keep the device connected. To celebrate the launch, the usage limit has also been doubled until August 5th. The official team also released an operating guide for Claude Code for developers, detailing that Effort Level is used to control how much AI reads files, validates, and expends effort, and introduced “Loops” design for agents, including automated modes such as round-robin, goal-oriented, time-driven, and active loops.

Q2: What new technology has Meta launched in the field of image and media generation? A2: Meta Superintelligence Labs released Muse Image and Muse Video, two brand-new media generation models. Muse Image possesses agent capabilities, will automatically search the internet for facts to ensure images match reality, and can perform local repainting through “self-correction” capabilities. Muse Video can produce videos with extremely high visual fidelity and natively supports audio synthesis. For security and identification, Meta also added Content Seal invisible watermarking technology.

Q3: How does Google’s Gemini API solve the pain point of developers handling long-running tasks? A3: Google expanded Managed Agents in the Gemini API, adding asynchronous background task execution. The system returns a task ID for querying at any time, without needing to maintain a constant HTTP connection. In addition, the update also adds Remote Model Context Protocol (MCP) server integration, allowing agents to connect directly to private databases, and supports custom function calls and credential refresh mechanisms.

Q4: Why are Microsoft and US enterprises gradually converting the AI models they are using? A4: Mainly to reduce huge computing and API costs. Microsoft is gradually replacing OpenAI and Anthropic models with its own MAI models in Copilot products such as Excel and Outlook. On the other hand, many US companies have begun to adopt Chinese open-source models (such as DeepSeek and GLM-5.2) in large numbers because the performance of these models has caught up with leading US models, but the price is only about one-tenth of competitors (or 60% to 90% cheaper), which has caused the proportion of tokens for Chinese models called on the OpenRouter platform to climb all the way to 46%.

Q5: In assessments of specific industry capabilities, which AI models are currently performing best? A5: According to the six major industry capability indices launched by Artificial Analysis (covering finance, law, healthcare, operations, etc.), Claude Fable 5 won the championship in all eight indices. In terms of open-source models, China’s GLM-5.2 performed brilliantly, taking the top open-source spot in five fields.

Q6: What breakthroughs have Nvidia and Cohere brought in the fields of speech and multimodality? A6: Nvidia released Nemotron-Labs-Audex-30B-A3B, a large language model with 30 billion parameters combining audio and text, possessing a context length of up to 1 million tokens, capable of handling complex tasks such as speech recognition, translation, and text-to-speech. Cohere launched Transcribe Arabic, a 2B parameter open-source Arabic speech recognition model created specifically for the Middle East market. Its word error rate is only 25.87, significantly leading Whisper Large V3, and can perfectly solve the “code-switching” problem where English and Arabic are mixed.

Q7: What is the “Doom Loop,” and what solution did Liquid AI propose? A7: The “Doom Loop” refers to the phenomenon where inference models get stuck when processing complex problems and repeatedly output the same sentences (such as “Wait, let me rethink”). Liquid AI proposed a brand-new training method called Antidoom (Final Token Preference Optimization, FTPO), which precisely locks onto the specific token that caused the loop and lets the model choose a more reasonable continuation method at that point in time, successfully reducing the loop rate of LFM2.5-2.6B models from 10.2% to 1.4%, while significantly improving evaluation scores.