Our take: For everyday users, the useful question is whether AI can do their work well at a lower total cost. A narrowing benchmark gap is interesting, but “3%” cannot stand in for your own experience.
What does the report actually say?
An RT Chinese video on Bilibili, published on October 5, 2026. It summarizes Bloomberg reporting on Bloomberg Intelligence analyst Robert Lea: following DeepSeek’s September release of V4.1 Flash, the benchmark score gap between leading Chinese models and US competitors had narrowed to about 3%, compared with roughly 9% in May and 15% earlier in the year. Sources: original video page, Bloomberg report, and licensed Bloomberg News edition.
Those are figures attributed to a particular benchmark comparison. We checked the video page and a published edition of the Bloomberg byline, but did not rerun the models. The figures should not be presented as a live ranking of every model on October 7.
A 3% gap is not a measure of the whole AI industry
Two people can earn similar exam marks and still perform differently at work. The same applies to models. Drafting a Chinese email, reading a long contract, cleaning a spreadsheet, writing code, and operating software can expose different strengths and weaknesses. One aggregate score compresses those differences and may not cover the task you care about.
LiveBench describes regularly refreshed questions and evaluation against verifiable answers, intended to address issues including test contamination. That gives benchmarks value, but they still measure specified tasks. They do not directly measure your editing time, compatibility with your software, or whether your workflow finishes smoothly. Read the LiveBench project documentation.
The reported figures support a narrower gap in that comparison. They do not establish equal performance on every task, and certainly do not mean that chips, computing capacity, services, and the entire AI industry are only 3% apart.
What everyday users might gain: a wider choice
Our interpretation is that more capable alternatives could let users choose by need and become less dependent on one brand. That is a promising direction, rather than a price cut demonstrated by this report.
For office workers, preserving an email’s important conditions and avoiding invented details can matter more than a ranking. For a small business, accurate product copy and a controllable tone may matter more than technical specifications. For a developer, the useful result is code that works in the actual project and remains maintainable. Success on one test does not automatically establish suitability for any of these jobs.
Compare the total cost of finishing a task
A cheap model that needs repeated attempts and manual repair may not save money. A more expensive tool that works reliably and fits your workflow might be better value. Count subscription or usage fees, waiting, retries, and human review. This is a method for choosing tools, not a measured claim about any particular product’s pricing or performance.
Try five tasks you genuinely do each week. Give two candidates the same materials and instructions, then record completion, errors, correction time, and actual spending. Start with public or anonymized material. Whether sensitive information can be entered depends on the tool’s data terms and your workplace rules. Choose the tool that reduces your burden, or use different tools for different tasks.
Why we are cautiously optimistic
A narrower gap could help make useful AI more accessible through competition. But a capable model and a dependable product are separated by speed, availability, language experience, integrations, and service. The market share implications in the reporting are an analyst’s expectation, not an outcome already demonstrated.
It is reasonable to add DeepSeek and other alternatives to a shortlist and reconsider what you pay for a familiar brand. There is no need to move every workflow because of one headline. Use rankings to decide what to try; use your own tasks to decide what to keep.
Original screenshots and sources
An editorial commentary by ToolAI. Reported facts are attributed to their sources; selection advice and implications are our analysis, without hands-on product benchmarking.