AI Daily: Anthropic Labor Report, PinchBench Model Evaluation, and New Development Tools

AI Real-World Tests and Technical Roundup: Rankings Reshuffled, Is Your Job Really at Risk?

To be honest, tracking new AI news every day can be overwhelming. Sometimes technologies claimed to be the most powerful turn out to be underwhelming in actual use. Today, I’ve compiled four significant tech developments. These include a realistic report on the labor market, unexpected PinchBench evaluation data, and new tools to ease the pressure on developers and designers. Let’s dive into these interesting findings.

Will AI Really Steal Everyone’s Jobs? Anthropic Gives an Unexpected Answer

Whenever people talk about AI, the biggest concern is always unemployment. Here’s an interesting perspective. Anthropic recently published a study on the labor market impacts of AI. They proposed a new metric called “observed exposure,” which combines the theoretical capabilities of language models with real-world Claude usage data.

To explain: many studies only look at what AI can theoretically do, but Anthropic focuses on how people actually use it. The report found that AI’s current actual coverage falls far short of its theoretical feasibility. In other words, its full potential hasn’t been realized yet. It can do a lot, but the proportion actually applied in the real world is relatively low.

So, which jobs are most affected? Data shows that computer programmers, customer service representatives, and data entry clerks have the highest observed exposure. Interestingly, workers in these high-exposure roles are typically older, have a higher proportion of women, and possess higher education and salaries.

Many might ask: Has AI already caused massive unemployment?

The answer is somewhat relieving. The report indicates that since late 2022, there has been no systematic rise in unemployment for high-exposure workers. However, there is a potential concern. For young job seekers aged 22 to 25, the recruitment speed for these high-exposure occupations has indeed slowed down. This might mean that while companies aren’t mass-firing existing staff, they are becoming more cautious about hiring inexperienced newcomers. Young graduates indeed face different challenges now, a social phenomenon that requires ongoing attention.

PinchBench Ranking Shakeup: Is More Expensive Always Better?

The next topic will definitely shock many developers. PinchBench, a platform for evaluating model capabilities, recently released its first specific test results for OpenClaw. Honestly, this data completely overturns previous assumptions.

There’s a common myth that more expensive services always mean better quality. But in this test, Google’s gemini-3-flash-preview took the top spot with a 95.1% success rate, costing only $0.72 per million tokens. In contrast, gemini-3-pro-preview, which costs twice as much, had a success rate of only 91.7%. This clearly proves that high price doesn’t necessarily mean better performance. Expensive pricing sometimes fails to reflect true technical strength.

Another stunning highlight is openai/gpt-5-nano. This model achieved an 85.8% success rate at an incredible cost of just $0.03. It’s the cheapest option on the list yet outperformed many high-priced competitors. For development teams with limited budgets, this is definitely an attractive option.

The most common question in the industry is: Which AI model offers the best value for money?

Looking at the big picture, minimax/minimax-m2.1 is arguably the best value choice currently. it ranked second with a 93.6% success rate, yet costs as low as $0.14. For comparison, Anthropic’s claude-sonnet-4.5 has a 92.7% success rate but costs $3.07—a more than twenty-fold difference.

However, some results were baffling. Everyone expected great things from Minimax 2.5, but it plummeted to 35.5%. This seems contradictory. A newer version should theoretically perform better, but it fell far behind the older version. The reason might be that the new architecture is still being tuned and hasn’t fully adapted to these specific test environments. This also serves as a reminder to always perform rigorous testing before deploying new models to production.

Coding Without Fear: Codex Security Makes Security Checks Smarter

Software development is moving faster than ever, but security often becomes a headache-inducing bottleneck. Often, development teams must compromise between speed and security. To solve this dilemma, OpenAI recently announced that Codex Security has entered the research preview stage. It’s an agent tool specifically designed for application security.

The problem is that traditional security tools often flag many irrelevant low-risk vulnerabilities, generating a lot of false positives. This forces security teams to spend massive amounts of time filtering noise. By thoroughly understanding the project’s context, Codex Security can precisely identify complex vulnerabilities that other tools easily miss.

It doesn’t just point out flaws; it also provides specific, actionable fix suggestions. In early internal testing, it successfully caught a serious cross-tenant authentication vulnerability. Over time, the tool’s precision has continued to improve, even reducing noise by 84% in some cases. This is also great news for the open-source community. OpenAI has already used this tool to help several well-known open-source projects fix critical vulnerabilities, making the entire software ecosystem safer.

Bringing Design to Life: OmniLottie Delivers a New Vector Animation Experience

Finally, let’s talk about a tool that will catch the eye of designers and frontend developers. OmniLottie is a new project built on the Hugging Face platform. It’s the first fully integrated multimodal Lottie generator family.

Readers might ask, what is Lottie? Simply put, it’s a very popular vector animation format that’s small in size and runs very smoothly on web or mobile apps. Previously, creating such animations required professional designers to spend a lot of time. Now, OmniLottie uses pre-trained vision-language models to generate complex Lottie animations directly from user instructions.

By simply inputting text, an image, or even a video, OmniLottie can automatically convert it into high-quality vector animation. The development team also released a massive dataset called MMLottie-2M (cc-by-nc-sa-4.0), containing two million animation samples with rich annotations. This provides significant help for future research in the field of vector animation generation. The project is currently open-source, and you can experience this interesting feature yourself through their online demo interface.