Chinese startup DeepSeek has released two new open models that it claims can match frontier systems like GPT-5 and Gemini 3.0 Pro, while staying fully open source and far cheaper to run.
📌 Key Takeaways
- DeepSeek has launched V3.2 and V3.2 Speciale as open models with MIT licenses.
- The models use sparse attention to cut long-context inference costs by roughly 70%.
- Benchmarks show Speciale edging past GPT-5 and Gemini 3.0 Pro on elite math contests.
- OpenRouter already lists Speciale with a 163k context window and competitive token pricing.
- Regulators in Europe and the US are scrutinising DeepSeek’s data practices and national ties.
DeepSeek V3.2 Targets Frontier Performance In The Open
DeepSeek has rolled out two flagship models: V3.2 for everyday reasoning and tools, and V3.2 Speciale for maximum reasoning depth and agentic workflows. Both ship under the permissive MIT licence, with weights and code publicly available.
The models are positioned directly against GPT-5 and Gemini 3.0 Pro, with DeepSeek arguing that careful architecture and post-training, not just massive budgets, are enough to reach frontier-level performance. For developers, that means frontier-style capabilities without being locked into a proprietary API.
People thought DeepSeek was a one-off breakthrough, but the team is coming back much bigger. — Chen Fang
Sparse Attention And Long Context Cut Costs
At the core of V3.2 is DeepSeek Sparse Attention (DSA), a custom attention mechanism that selects only the most relevant parts of long inputs instead of brute-forcing every token against every other token. That lets the model handle long documents without exploding compute.
DeepSeek’s technical report says DSA roughly halves inference costs on long sequences compared to its previous generation. Processing a 128k-token sequence drops from about $2.40 per million tokens to around $0.70, while maintaining similar quality.
- Parameter scale: ~685 billion parameters
- Context window: 128k tokens in reports, 163,840 tokens on OpenRouter
- OpenRouter pricing (Speciale): about $0.28/M input and $0.40/M output tokens
- Optimised for: long-context reasoning, coding, and tool-augmented workflows
These numbers matter for anyone running agentic workflows, retrieval-augmented generation, or code analysis over large repositories. Frontier-class quality at materially lower long-context cost will put direct pressure on proprietary pricing models.
Benchmarks Put DeepSeek In The GPT-5 Conversation
DeepSeek backs the launch with competition-grade benchmarks rather than just synthetic leaderboards. The V3.2 Speciale variant reportedly hit a 96% pass rate on the AIME 2025 math contest, edging GPT-5-High and Gemini 3.0 Pro on the same benchmark.
On the Harvard-MIT Mathematics Tournament, Speciale reached about 99.2%, again above Gemini’s reported score. The model also achieved gold-medal level performance at the International Mathematical Olympiad and strong results at IOI and ICPC, all evaluated without external tools or internet access.
Token efficiency remains a challenge, and the model often needs longer generations to match top systems. — DeepSeek V3.2 Technical Report
In coding, V3.2 resolves over 73% of real-world bugs on SWE-Verified and significantly outperforms GPT-5-High on Terminal Bench 2.0-style multi-step coding workflows, according to reported tests. That combination of reasoning and software reliability is what many enterprises actually care about.
Open Source Strategy Meets Regulatory Walls
DeepSeek’s decision to release weights, training code, and documentation under MIT effectively hands the community an off-the-shelf frontier model. Migration scripts that accept OpenAI-style requests lower friction for teams considering a switch or hybrid strategy.
That openness, however, collides with geopolitics. European regulators have already flagged DeepSeek’s handling of EU user data, with one German authority calling transfers to China unlawful under local privacy rules. Italy has ordered the app blocked, and US officials are pushing to keep DeepSeek off government devices.
Analysts also see the launch as another data point in the cost war. DeepSeek claims training its earlier V-series models cost a tiny fraction of Western rivals, raising uncomfortable questions about whether the current spending race from US firms is sustainable if cheaper open alternatives keep closing the gap.
Conclusion
With V3.2 and V3.2 Speciale, DeepSeek is not just chasing benchmarks; it is testing whether open, high-end models can undercut closed systems on both performance and price. For developers, the immediate impact is more choice and more bargaining power.
For incumbents like OpenAI and Google, the threat is longer-term. If a Chinese lab can keep shipping frontier-class, open-licensed models with aggressive cost profiles, the economics of proprietary AI could shift faster than expected, and the real competitive moat may move from raw model quality to trust, ecosystem, and regulation.
📈 Latest AI News
2nd December 2025
- OpenAI Hits ‘Code Red’ As Gemini Puts Real Pressure On ChatGPT
- After Years Of AI Missteps, Apple Brings In A New AI Chief
- A Few Lines Of Poetry Are Enough To Defeat Many AI Safety Systems
- James Cameron Calls AI Actors “Horrifying”
- Google Deletes X Post After AI ‘Recipe Infographic’ Exposed as Stolen
For the recent AI News, visit our site.
If you liked this article, be sure to follow us on X/Twitter and also LinkedIn for more exclusive content.
