Strengths
- DeepSeek's largest open model, 1.6 trillion parameters under the permissive MIT license.
- Low price for a model this size: $1.32 per million input tokens and $3.96 per million output tokens at peak, half that off-peak.
- A 1 million token context window, answers of up to 384K tokens and three thinking levels (low, high, max).
- DeepSeek reports 87.9 on Terminal Bench 2.1 and 60.0 on Humanity's Last Exam with tools.
Weaknesses
- Text only. Unlike the cheaper Flash model, it cannot handle images.
- On DeepSeek's DeepSWE coding test, the newer and cheaper V4.1 Flash scores higher (77.0 on DeepSWE v1.1 versus 62.7), although the test versions differ.
- Mistral says its Large 4 beats V4 Pro on Coding Agent Index and AutomationBench. That is Mistral's own claim.
- The app and API share the data protection issues of all DeepSeek services: data in China and a NÚKIB warning rating the threat as “high”.
Best for
- complex text reasoning at a low API price
- self-hosting by organisations with large GPU clusters
Not ideal for
- images and scanned documents
- personal data in the DeepSeek app or API
- Czech public administration
How to use it in Europe
V4 Pro is available in the app, on the web and in the API (deepseek-v4-pro), and it works in Slovakia and Czechia. Data is stored in China, though, Czech state administration may not use it for official work, and Czechia's NÚKIB rates the threat as high. DeepSeek has said it will keep the V4 Pro API running after 14 September 2026, with no end date. The weights are open, but self-hosting takes data centre hardware such as a node with four NVIDIA GB300 chips. Local use with no link to DeepSeek is outside the NÚKIB warning.
What is coming
Confirmed by vendor When DeepSeek released V4.1 Flash in September 2026, it mentioned a V4.1 Pro model that has yet to launch. It has not published a date, prices or specs. When: no date yet →
Rumor, unconfirmed According to The Information, DeepSeek is training a model with about 2 trillion parameters. References to V5 reportedly appeared in DeepSeek's code on 27 October. It is unclear whether this is the same model, and the company has announced nothing. When: no date, AIBase expects November 2026 rather than October →
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| Terminal Bench 2.1 | 87.9 | huggingface.co ↗ (vendor-reported), August 2026 |
| Humanity's Last Exam, without / with tools | 42.7 / 60.0 | huggingface.co ↗ (vendor-reported), August 2026 |
| DeepSWE | 62.7 | huggingface.co ↗ (vendor-reported), August 2026 |
| Toolathlon-Verified | 74.1 | huggingface.co ↗ (vendor-reported), August 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- DeepSeek V4.1 Flash: DeepSeek's cheap, fast open-weight model with a 1 million token window. In the app and API, though, your data goes to China.
Similar models from other vendors
- Claude Opus 5.5 (Anthropic): Anthropic's main model for demanding work, coding and long documents, and the one Anthropic recommends starting with.
- Gemini 4 Carbon (Google): Google's most capable model, announced on 28 October 2026. So far only selected cybersecurity experts have access.
- Gemini 4 Argon (Google): Google's most capable publicly available model, out in October 2026. Built for demanding analytical work where accuracy matters and made-up answers are costly.
- Muse Wonder (Meta): Meta's strongest model for long-running agents and coding, still released as a preview.
- Mistral Large 4 (Mistral AI): Mistral's largest model, still in preview. Built for demanding tasks, long documents and security analysis with data kept in the EU.
- Mistral Large 3 (Mistral AI): Mistral's previous large model with open weights under Apache 2.0. Cheap in the API and suited to your own powerful server.