DeepSeek V4 Pro Goes GA: 2% of Fable 5's Price
DeepSeek V4 Pro 0813 exits preview quietly. 1.6T params MoE, 49B active, 1M context, priced at $0.87/M output — less than 2% of Claude Fable 5.
I noticed it at 11pm last night. The DeepSeek API pricing page had one small change: the version string next to deepseek-v4-pro flipped from preview to DeepSeek-V4-Pro-0813. No blog post. No tweet. No press release. Just a version number bump on a pricing table.
That one-line change means their flagship model — running in preview since April 2026 — is now officially production-ready.
The DeepSeek API pricing page — V4 Pro version number quietly updated to 0813.
Specifications
| Spec | Value |
|---|---|
| Total Parameters | 1.6 trillion (1,600B) |
| Active Parameters per Token | 49B |
| Architecture | Mixture of Experts (MoE) |
| Context Window | 1,000,000 tokens |
| Max Output | 384,000 tokens |
| Concurrency Limit | 500 |
| License | MIT |
| API Formats | OpenAI + Anthropic dual-format |
This puts it in the top tier globally. The preview version has already been downloaded 1.4 million times on HuggingFace. Every third-party benchmark published so far was run on that preview.
Pricing: Less Than 2% of Claude Fable 5
The GA pricing matches the off-peak rate — which is the permanent price reduction from May 2026 (cut to 1/4 of the original).
| Item | DeepSeek V4 Pro | Claude Fable 5 | Ratio |
|---|---|---|---|
| Input (/1M tokens) | $0.435 | $10.00 | 4.35% |
| Output (/1M tokens) | $0.87 | $50.00 | 1.74% |
| Cache-hit Input | $0.003625 | — | — |
On output pricing, it’s 1.74% of Fable 5. Two full orders of magnitude cheaper.
Per-task cost is even more dramatic. One benchmark institution measured $3.15 per task on Claude Fable 5 versus $0.03 on DeepSeek V4 Flash — a 100x gap. V4 Pro 0813 hasn’t been independently benchmarked per-task yet, but the pricing structure is unchanged.
Why It’s This Cheap
The architecture uses two attention variants:
- Compressed Sparse Attention — reduces computation per token
- Heavy Compressed Attention — dramatically shrinks KV cache
DeepSeek claims this combination cuts per-token inference compute to 27% of V3.2, and KV cache to 10% of the previous generation. At 1M context length, the savings compound massively. Lower inference cost → lower pricing floor.
DeepSeek’s homepage — characteristically understated, no launch announcement anywhere.
Performance: Neck and Neck With Fable 5
DeepSeek ran internal benchmarks across 10 agent evaluations against Claude Fable 5. Dropping the widest gap (Humanity’s Last Exam, where V4 Pro trails by 10.6 points), the remaining 9 benchmarks show an average gap of just 2.8%, with DeepSeek winning 2 of them.
Important caveat: these are self-reported numbers using an internal evaluation framework. Neither the framework nor the test sets are public. The real-world performance of 0813 needs independent verification.
Three Inference Modes
| Mode | Description | Use Case |
|---|---|---|
| Standard | Normal inference | Chat, simple tasks |
| High Reasoning | Enhanced CoT | Complex reasoning, code gen |
| Maximum Reasoning | Push model limits | Math, long-chain reasoning, agents |
The API supports both OpenAI and Anthropic formats natively. Developers already building on either ecosystem can switch with zero code changes. Full support for tool calls and structured JSON output.
The V4 Product Line
| Model | Positioning | Input | Output |
|---|---|---|---|
| V4 Flash 0731 | High-volume, low-cost | $0.14/M | $0.28/M |
| V4 Pro 0813 | Complex reasoning, agents | $0.435/M | $0.87/M |
Flash handles throughput. Pro handles depth. Many AI teams are already running both in production — routing simple completions to Flash and complex agent tasks to Pro.
What to Watch
Price increase incoming. The pricing page explicitly states “a significant price adjustment” is planned. No timeline or specifics given. The current window may not last.
Compute. Founder Liang Wenfeng disclosed roughly 20,000 H100-equivalent GPUs. Small compared to Western labs’ clusters of hundreds of thousands. But they’re actively acquiring more and building a GW-class domestic compute center.
Weights. MIT-licensed, but GA weights aren’t released yet. Given the V4 Flash pattern (API first, weights follow), likely just a matter of time.
Third-party evaluation labs are queuing to benchmark V4 Pro 0813. Real performance data incoming.
What This Means for Developers
If you’re building AI agents or processing long contexts:
- Cost drops two orders of magnitude — same-quality tasks go from dollar-level (Fable 5) to cent-level (V4 Pro)
- Zero migration effort — supports both OpenAI and Anthropic API formats
- Flash + Pro combo — route simple tasks to Flash ($0.28/M), complex to Pro ($0.87/M)
- Wait for verification — self-reported benchmarks look strong, but run your own evals before production commitment
For more on V4 Flash benchmarks, see our DeepSeek V4 Flash deep dive.
FAQ
How does V4 Pro compare to V4 Flash?
Flash is optimized for throughput at minimal cost. Pro is optimized for reasoning depth — better at multi-step agent tasks, math, and code generation. Flash costs 3x less but Pro scores significantly higher on complex benchmarks.
Is it really 2% of Fable 5’s price?
On output tokens: $0.87 vs $50.00 = 1.74%. Yes. But DeepSeek has warned of an upcoming price increase.
Should I switch from Claude/GPT to V4 Pro?
Run your own eval first. The self-reported numbers are strong but unverified by third parties. If your workload is cost-sensitive and the quality holds up on your tasks, the savings are massive.
When will weights be released?
No official date. MIT license confirmed. Based on the V4 Flash precedent (API → weights within weeks), expect it relatively soon.
What about the price increase?
The pricing page says “significant increase expected.” No timeline. Lock in current rates if your usage is predictable.
