DeepSeek rolls out new AI models to rival Google, OpenAI

DeepSeek has launched DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, aiming to compete with major players like Google and OpenAI.

V3.2 matches the performance of GPT-5 on several reasoning benchmarks and can use tools such as search engines, calculators, and code executors.

The Speciale version is designed for mathematical computations, performs on par with Google’s Gemini-3 Pro, and achieved high scores on tests like the International Math Olympiad.

The company said the models integrate reasoning with autonomous tool use, supporting both thinking and non-thinking modes.

DeepSeek previously gained attention in January for advancing open-source capabilities.

The new models are available through app, web, and API, with V3.2-Speciale currently limited to API access.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Independent benchmarks back DeepSeek-V3 as competitive; third-party checks for V3.2 parity claims are pending

The Chatbot Arena leaderboard is an open, community-run platform 1. It compares large language models with head-to-head user votes plus aggregate tests such as MMLU-Pro, a broad exam for knowledge and reasoning; GPQA Diamond is a graduate-level science benchmark 1. DeepSeek-V3 beats other open-source systems and comes close to top closed-source options 2.
DeepSeek-V3.2-Exp (an evaluation-only build) matches its predecessor V3.1-Terminus, another non-public variant, in benchmarks 3. Expect small gains rather than a leap. Treat any “GPT-5 parity” claim as unproven until third-party checks on platforms like Chatbot Arena arrive 1.
DeepSeek-V3 supports a 128K token context window, which lets the model handle longer inputs in one pass 2. It posts strong math and code results 2. That aligns with the Speciale variant, which targets mathematical computations.

Engineering and data teams can save 10-30X by routing workloads to DeepSeek’s API

DeepSeek’s API pricing can be up to 30X cheaper than OpenAI or Anthropic, based on independent analysis 4. Developers and startups can cut spend by routing suitable work to this API instead of premium providers 4.
The cache-hit structure bills $0.07 per million input tokens, while cache misses cost $0.56 4. A hit occurs when the service recognizes repeated prompt segments and bills them at a lower rate. Teams can design apps to raise hit rates with repeated prompts or batched requests 4.
MLOps (Machine Learning Operations) and AI gateway vendors should add DeepSeek connectors 3. The 50%+ API price cut 3, plus tool-use support, makes it a fit for agent workflows that call calculators, search, or code executors.