DeepSeek says R1 now rivals OpenAI and Gemini models

DeepSeek’s R1 upgrade aims straight at OpenAI and Gemini

DeepSeek has launched its upgraded R1 model R1-0528, and it’s not playing catch-up anymore. The new version significantly gains mathematical reasoning, logical accuracy, and code generation.

It’s being hailed as one of China’s most advanced AI efforts yet, with benchmark results placing it just behind OpenAI’s o3 and Google’s Gemini 2.5 Pro.

As global AI competition heats up, DeepSeek’s low-profile but high-impact release shows it’s ready to contend at the top of the leaderboard.

87.5% accuracy in AIME math, a real leap

The R1-0528 model made headlines with its performance on the AIME 2025 benchmark, achieving 87.5% accuracy, up from just 70%. This jump highlights serious advances in the model’s ability to handle complex problem-solving and multi-step reasoning.

It also reflects DeepSeek’s deeper token utilization, more than 23,000 tokens per math question on average. That’s not just brute force, it’s more innovative architecture and refined optimization behind the scenes.

Hugging face release shows quiet confidence

Rather than a flashy press conference, DeepSeek quietly uploaded the R1-0528 model to Hugging Face. This low-key strategy mirrors its earlier releases, but hasn’t stopped the tech community from noticing.

Developers now have access to open-source weights and API access, making R1-0528 one of the most accessible top-tier LLMs on the market. It’s a subtle flex from a company letting results and benchmarks do the talking.

Function calling and JSON output support added

R1-0528 isn’t just about better math. DeepSeek has added enhanced support for function calling, structured data generation, and JSON output, which are crucial for developers working on automation and enterprise use cases.

These upgrades make building AI-powered tools and apps directly on the model easier, reducing post-processing work and increasing integration speed. For many, this aligns the model with top Western LLMs regarding real-world deployability.

Logical reasoning gets a serious boost

Where older versions of DeepSeek models faltered in multi-step reasoning, R1-0528 shines. Its logical capabilities have caught up with, and in some cases surpassed, what’s seen in popular open-source LLMs.

The model now delivers more consistent, interpretable, and coherent outputs from coding challenges to data analysis queries. It’s a significant win for researchers and devs looking for a smart, affordable LLM that doesn’t compromise depth or precision.

Benchmarks put R1-0528 just behind GPT-4.5

According to third-party evaluation from LiveCodeBench, DeepSeek’s R1-0528 now performs just a step below OpenAI’s o3 and Gemini 2.5 Pro models. That’s a staggering accomplishment for a model built on a much smaller budget, estimated at under $10 million.

With this release, DeepSeek has moved from “promising” to “serious threat” status among the LLM elite, especially for organizations looking to adopt advanced AI without paying GPT-tier subscription fees.

Hallucination rates drop significantly

One of the biggest criticisms of LLMs is their tendency to “hallucinate,” generate convincing but inaccurate responses. DeepSeek claims R1-0528 has made a substantial dent in this issue. Tests show improved fact retention, better citation handling, and fewer nonsensical outputs.

For use cases in finance, medicine, or law, where hallucinations can be costly or dangerous, this update makes R1-0528 a far more trustworthy option.

Chinese innovation fueled by export restrictions

Ironically, U.S. export controls may have accelerated DeepSeek’s innovation. Blocked from high-end chips and Western training tools, Chinese companies had to get creative. DeepSeek’s rapid progress with fewer resources proves that necessity breeds invention.

Founder Liang Wenfeng’s rise as a national tech figure reflects growing confidence in China’s domestic AI ecosystem and a warning to global competitors that resource restrictions aren’t enough to curb advancement.

R1’s success helped influence chinese AI policy

The release of the original R1 earlier this year sent shockwaves through China’s tech and political circles. It was a surprise hit that challenged assumptions about what China’s AI sector could deliver.

R1’s performance helped spotlight the importance of home-grown innovation, and R1-0528 builds on that momentum. Liang Wenfeng’s appearance at President Xi’s economic forum underscores how seriously China is backing DeepSeek and its implications for national AI strategy.

R1-0528 powers distillation into other models

Beyond its standalone performance, R1-0528 is also powering advances elsewhere. DeepSeek used its reasoning engine to improve Alibaba’s Qwen 3 8B Base model via distillation, boosting performance by over 10%.

This modular reuse shows how foundational R1-0528 could become across the Chinese AI landscape. It’s not just a model; it’s a platform others can build from, accelerate training with, and scale rapidly.

R1-0528’s cost efficiency redefines AI economics

One of R1-0528’s most impressive feats isn’t just performance, it’s price. Industry estimates suggest the model was trained for under $10 million, a fraction of the budgets used by OpenAI or Google.

This efficiency flips the narrative that only Big Tech can build cutting-edge AI. DeepSeek proves that with intelligent optimization, tight engineering, and targeted resources, it’s possible to deliver top-tier results without breaking the bank, disrupting the economics of AI development.

Nvidia felt the R1 shockwave too

Nvidia’s stock dipped briefly after R1’s launch—what Wikipedia calls ‘the largest single-company decline in U.S. stock market history’—reflecting investor concern over rising non‑U.S. AI competition.

Chipmakers, if models like R1 prove viable on fewer or locally sourced resources. R1-0528 builds on that momentum, showing China can produce competitive AI even under semiconductor restrictions. It’s a wake-up call for U.S. firms long confident in their hardware-led lead.

U.S. firms respond with discounts and mini models

DeepSeek’s rise hasn’t gone unnoticed. In the wake of R1-0528, OpenAI dropped prices and introduced the O3 Mini model to stay competitive. Google followed with new discounted Gemini tiers.

These moves suggest Western firms are now watching China’s LLMs not just out of curiosity, but concern. DeepSeek’s cost-performance balance is influencing pricing strategies across the AI industry.

Developers praise coding capabilities

R1-0528 isn’t just good at logic puzzles, it’s strong at real-world coding, too. Developers using the model report smoother performance in Python, JavaScript, and front-end frameworks.

Enhanced function calling, JSON structuring, and fewer syntax errors make it a serious tool for building apps and automations. It may not yet beat GPT-4 Turbo in creative writing, but it’s becoming a trusted coding companion for engineers.

DeepSeek’s chat platform gains traction

Beyond the model, DeepSeek also builds user-facing tools. Its in-house chatbot interface, powered by R1-052,8, has quietly gained users in China and beyond.

With competitive response quality and faster output than some Western bots, it’s integrated into search tools, productivity apps, and enterprise support platforms. This vertical integration could help DeepSeek establish a lasting presence in consumer-facing AI.

And DeepSeek’s influence might not stop there, its tech could soon be fueling even bigger players: DeepSeek Could Power Meta’s AI Ambitions.

From underdog to contender in six months

Six months ago, DeepSeek was barely known outside of China. Today, its models are being compared to OpenAI’s best, and its founder shares the stage with China’s tech royalty.

R1-0528 marks not just an upgrade, it’s a declaration. With every benchmark crossed and every price barrier broken, DeepSeek shows that the AI race is far from over. It’s just getting interesting.

And if DeepSeek’s rise has you watching closely, you’ll want to see what its rivals are up to: Worried About DeepSeek? Google Gemini Tracks You More.

What do you think DeepSeek’s R1 model outperforms Gemini or ChatGPT? Please share your thoughts and drop a comment.

Read More From This Brand:

• Criminals Spread Malware Disguised as DeepSeek AI
• Why Are Companies Banning DeepSeek AI?
• DeepSeek and Gemini AI Go Head to Head

Don’t forget to follow us for more exclusive content right here on MSN.

If you liked this story, you’ll LOVE our FREE emails. Join today and be the first to get stories like this one.

This slideshow was made with AI assistance and human editing.