DeepSeek V4 Pro 0813 - this is a historic turning point

Коротко о главном (BLUF)

В этой статье мы рассказываем о важных новостях и функциях платформы NeuralSpace. Узнайте, как наши нейросети могут помочь вам в решении повседневных задач, сэкономят время и откроют новые возможности для творчества.

Today DeepSeek shipped the final version of V4 Pro — revision 0813. No launch event and no countdown: the API simply updated. Anyone already working with deepseek-v4-pro got the new model automatically, with no code changes. The April build was a preview; this one is the GA release, and the difference shows most clearly where a model is expected to do sustained work rather than answer a chat message.

This is the day coding at Opus level became genuinely affordable. That sentence would have been marketing wishful thinking this morning; by evening it is a fact with benchmarks attached.

The model is already live on NeuralSpace: in chat, in the Code section and in the API. No "waiting for rollout" — just open it and use it.

What the model is

Technically it is the same beast as in April: a Mixture-of-Experts with 1.6 trillion parameters, 49 billion of them active at any step. Context is one million tokens, and the model can emit up to 384 thousand tokens in a single answer. For comparison, that is roughly twenty times more than most models in the neighbouring tabs.

The model can think in two modes — thinking and non-thinking — and reasoning_effort now has three steps: max, high, low. In practice a small question can be run cheaply, while a complex chain can be handled at maximum reasoning without switching models.

Benchmarks: what actually improved

Artificial Analysis re-measured the final version. Intelligence Index rose from 43.7 to 45.3. Coding Index went from 58.7 to 59.4. But Agentic Index jumped the most: from 35.3 to 37.8. That is not a coincidence — revision 0813 was tuned specifically for agentic scenarios, where a model does not answer in one shot but goes back and forth: reads files, calls tools, inspects the result and corrects itself.

Speed is about the same as the previous version: around 76 tokens per second on output, first token in about one and a half to two seconds. That is fast for a model this size. The honest downside: the model is verbose. On AA measurements it emits one and a half to two times more tokens than the market average — count on paying for extra tokens and extra seconds of waiting when it reasons at length.

Against Claude and GPT: honest arithmetic

The most striking row in the fresh table is Claude Opus 4.8, Anthropic's previous flagship. DeepSeek V4 Pro 0813 beat it in coding benchmarks on both sides: Terminal-Bench 2.1 — 87.9 versus 85.0, DeepSWE — 62.7 versus 58.0. And Opus 4.8 costs $5 per million input tokens and $25 per million output tokens — 11 and 29 times more. Here is the table:

The final version came closest to Claude Fable 5: 87.9 versus 88.0 in Terminal-Bench — a difference of one tenth of a point. In DeepSWE Fable is further ahead: 70.0 versus 62.7. Averaging all ten agentic benchmarks from DeepSeek's table, Fable leads by 5.3% on average — but DeepSeek wins on two of ten, and without the single outlier (Humanity's Last Exam without tools: 42.7 versus 53.3) the average gap shrinks to 2.8%.

Price, meanwhile, differs not by percentage points but by orders of magnitude. Fable 5 costs $10 per million input and $50 per million output tokens; V4 Pro costs $0.435 and $0.87. On a mixed rate that is about $30 versus $0.65. So for a difference of a few percentage points on tests, Anthropic asks for roughly 46 times more money. Honesty caveats: DeepSeek measured its own comparison table on its own unpublished infrastructure, two of the ten tests are internal, and there are no independent measurements of 0813 yet — the model came out today.

For scale, remember Kimi K3. A month ago it became the first Chinese model to beat Claude Opus — that was a sensation. But Kimi K3 is a trillion-parameter model that needs an expensive server to run, so at retail it is only three times cheaper than Claude Fable and only 40% cheaper than Claude Opus. Impressive, but not exactly democratization.

DeepSeek V4 Pro is a genuinely historic moment. Genuinely approaching Opus in coding, it costs 30 times less than Opus and 60 times less than Fable. Vibe coding just became available to everyone — not in the advertising sense of "available", but at the price of entry.

With GPT the picture is more nuanced. GPT-5.5 scores 56.3 on the Artificial Analysis index versus 45.3 for DeepSeek — an honestly noticeable gap. But the rates are incomparable: $5/$30 per million versus $0.435/$0.87. The cheaper GPT-5.6 Luna, however, costs $0.10/$0.60 — less than DeepSeek — and scores 52.3. So "the cheapest of the big ones" is no longer unique: on price V4 Pro competes with Luna, while on score it loses to both Terra and Sol.

Add one sobering quote from outside: in a recent assessment by the US NIST (CAISI), V4 Pro was said to trail the market leaders by about eight months. They measured the preview version, though. Whether 0813 narrowed that gap remains for independent benchmarks — and there are none yet.

The price did not change. For now

Input is $0.435 per million tokens, output is $0.87. Cache hits are still almost free: $0.003625 per million, 120 times cheaper than a normal input. For agentic chains this is the main point: project context is re-read again and again, and it is the cache that decides how much an hour of agent work costs.

One caveat: DeepSeek has already warned in its docs that it plans to raise API prices in general. When and by how much — not said. April rates still apply for now, and they are noticeably lower than comparable Western models.

Where to try it right now

• Chat: open the model page — the final 0813 is already the default.
• Code: in the Code section the model connects to a project and works with a repository, files and a terminal.
• API: the key is issued in NeuralSpace API keys, OpenAI-compatible format.

Payment is in rubles from one shared balance, with no subscription, no VPN and no foreign card. You are charged for actual usage, not a "month ahead" bundle.

What the model cannot do is look at images: input is text-only. If you need to dissect a screenshot or a photo, that is for other models. But for text, code and long agentic tasks, the final V4 Pro is one of the most honest strength-to-price deals around. Try it in chat, or grab a key on the API keys page to build it into your product.

Часто задаваемые вопросы (FAQ)

Вопрос: Как начать работу? Ответ: Зарегистрируйтесь на платформе и получите бесплатные токены для тестирования.

Вопрос: Подходят ли сгенерированные материалы для коммерческого использования? Ответ: Да, вы получаете полные коммерческие права на весь созданный контент.