DeepSeek Quietly Upgraded V4 Pro — Still 57× Cheaper Than Fable 5 on Output

A version bump that is still a launch

The official DeepSeek API docs now list DeepSeek-V4-Pro-0813 as the current V4 Pro model — a build dated 13 August 2026. The same page quietly confirms two feature upgrades that were previously reserved for the smaller V4 Flash: the Responses API, DeepSeek's native agentic interface, and an Anthropic-compatible endpoint, so the model drops into tooling originally built for Claude. The context window stays at 1 million tokens with a 384K maximum output, and thinking mode is on by default.

None of this came with a price tag change. DeepSeek-V4-Pro-0813 still bills $0.435 per million input tokens (cache miss) and $0.87 per million output tokens — the same price list DeepSeek has held since its spring discount. That stability is exactly what makes the comparison below so stark.

The gap nobody can ignore: 57× on output

Let's put the three models side by side at their current list prices, per million tokens, converted to euros at the European Central Bank reference rate of 1.1545 USD/EUR (12 August 2026):

The headline is the output column. Every million tokens Fable 5 generates costs you $50; the same volume from V4 Pro costs $0.87. Input is less dramatic but still lopsided: Fable 5 charges $10 against DeepSeek's $0.435. And the real-world gap is wider still, because DeepSeek's cache-hit input rate collapses to $0.0036 per million tokens — OpenRouter's weighted average across real traffic sits near $0.034 per million input tokens, a number that reflects just how aggressively caching and discounts drive the effective price down.

What you actually get for €0.75 per million tokens

The cheap part only matters if the model can do the work. Early signals say it can, at least on the axes developers care about. In the LMArena Code Arena WebDev board, DeepSeek-V4-Pro (Max) lands at roughly 1607 points in the preliminary AutoEval — an automatic scoring where a reward model trained on human preferences casts votes in place of live ones. That places it around eighth overall and second among open-weights models, just behind GPT-5.6 Sol (xHigh) at 1622 and behind Kimi K3 (Max) at 1674. In the Text Arena it sits at about 1465 points, on par with GLM-5.1 and Grok 4.6. One caveat worth repeating: these are early AutoEval numbers that will shift as live human votes accumulate, and the fresh build has not yet appeared on the converged leaderboard.

Independent measurement is more flattering. Artificial Analysis records 88.8% on GPQA Diamond, 37.5% on Humanity's Last Exam and 46.2% on Terminal-Bench Hard, and reports from X put the main Terminal Bench score roughly 15.8% higher than the April preview. For a model that costs a fraction of its rivals, that is the entire argument in one paragraph: you are no longer paying a premium to be in the same performance tier.

The catch: DeepSeek already warned about price hikes

There is one reason not to get comfortable with these numbers. On 6 August 2026 DeepSeek emailed its API customers that it plans "a significant increase" in API pricing, and the note still hangs under the official price list. No figures, no date. Given that the April launch price of V4 Pro was $1.74 per million input tokens — today's $0.435 is a discounted rate — a return toward that level is the obvious direction. If you want to build on V4 Pro at current prices, the window is open now and its width is unknown.

For European developers

DeepSeek-V4-Pro-0813 is available in the EU today, either directly through the OpenAI-compatible endpoint at api.deepseek.com or through OpenRouter, which accepts euro-denominated billing and exposes the model alongside a throughput of about 54 tokens per second. One practical note for European teams: DeepSeek's models remain noticeably weaker at morphologically rich European languages than at English, so they are a better fit for code, logic and data work than for long-form prose in, say, Czech or Finnish. For a deeper breakdown of what these models cost in real tasks, we compared DeepSeek against GPT, Claude and Gemini in a dedicated pricing analysis, and we covered the announced price increase here.