deepseek-flash; legacy V4 Flash names route to V4.1 Flash, and V4 Pro has also moved beyond the preview release. Pricing is now dynamic by peak and off-peak periods. This review checks official documentation and model cards only; we did not call the paid API, download weights, or reproduce benchmarks.
TL;DR: three points to check
- 1.6T MoE with 49B activated: those are the V4-Pro figures in the official model card, alongside a 1M context window.
- DSA and token-wise compression target long-context cost: the official report says V4-Pro uses 27% of V3.2's per-token FLOPs and 10% of its KV cache at 1M context.
- The current products have changed: V4.1 Flash replaced the old V4 Flash API, V4 Pro was updated, and April prices should not be treated as current.
1. Context: where V4 sits in DeepSeek's 2026 roadmap
On 2026-04-24, DeepSeek released V4-Pro and V4-Flash in preview. Official Hugging Face cards list open weights under the MIT License. The release added DSA sparse attention, token-wise compression, and a 1M context limit.
The preview card reported results such as 93.5 on LiveCodeBench and 3206 on Codeforces for Pro-Max. These are vendor results. They should be read with the stated reasoning effort, harness, and sampling settings, then checked against a task-specific regression set.
By September, DeepSeek had released the V4 Pro update and V4.1 Flash. The remaining sections preserve the April analysis where useful, while the operational details use current official documentation.
2. Specs at a glance
3. The architectural innovation: what is DSA Sparse Attention?
DeepSeek presents DSA as a major V4 architecture change. Attention compute and KV-cache use rise quickly with context length, so a 1M-token limit is only useful when latency and memory remain manageable.
- Sparse attention computation: each token attends only to a dynamically selected "salient subset" of other tokens, not the full context. Subset selection is learned by the model through content-driven sparsity, rather than a fixed local window or a MoE-style routing classifier.
- Token-wise compression: at the KV-cache layer, historical tokens are compressed adaptively based on content, so older tokens occupy less memory while preserving key semantics.
Combined effect: V4-Pro uses 27% of V3.2's per-token FLOPs and 10% of its KV cache; V4-Flash is even more aggressive at 10% FLOPs and 7% memory. On the same hardware, V4 can serve roughly an order of magnitude more concurrent requests.
Conceptually DSA shares ground with Mistral's Sliding Window Attention, the long-context moves Anthropic and others have made for Claude, and the state-space approach in Mamba. All trade full-connectivity attention for long-context practicality. What's distinctive about DeepSeek's version is how thoroughly the sparse assumption is baked into the entire pipeline: the model is trained, not just inference-patched, under the sparse regime. That's why V4 ships 1M context as a default capability rather than a special configuration.
4. Benchmarks: V4-Pro vs GPT-5.4 vs Claude Opus 4.6 vs Gemini 3.1-Pro
Below is V4-Pro (in highest-reasoning "Pro-Max" mode) compared head-to-head with the three closed-source frontier models. Numbers are from DeepSeek's official technical report cross-checked against published third-party numbers. Green marks where V4-Pro takes the best (or tied-best) result; orange is close; red is a clear gap.
| Benchmark | DeepSeek V4-Pro | Claude Opus 4.6 | GPT-5.4 | Gemini 3.1-Pro |
|---|---|---|---|---|
| LiveCodeBench | 93.5 | 88.8 | Not listed | 91.7 |
| Codeforces Rating | 3,206 | Not listed | 3,168 | Not listed |
| SWE-bench Verified | 80.6 | 80.8 | Not listed | 80.6 |
| MMLU-Pro | 87.5 | 89.1 | 87.5 | 91.0 |
| GPQA Diamond | 90.1 | 91.3 | 93.0 | 94.3 |
| Terminal-Bench 2.0 | 67.9 | 65.4 | 75.1 | 68.5 |
| HLE (Hard Logic Eval) | 37.7 | Not listed | Not listed | Not listed |
How to read this
- Coding and competitive programming: V4-Pro is the top open-source result and ties or beats closed-source frontiers. LiveCodeBench 93.5 leads both GPT-5.4 and Gemini 3.1-Pro, and Codeforces 3206 edges out GPT-5.4. SWE-bench Verified is a three-way tie with Claude Opus 4.6 and Gemini 3.1-Pro.
- Knowledge-heavy and scientific reasoning are still gaps. MMLU-Pro trails Gemini by 3.5 points; GPQA Diamond trails by 4.2 points. DeepSeek itself acknowledges a "3–6 month lag" behind the SOTA frontier on these.
- Command-line agentic work (Terminal-Bench): V4 is on par with Claude/Gemini, but materially behind GPT-5.4. GPT-5.4's 75.1 remains the single best score for tool-using agents.
The preview table shows competitive coding and agent results, but it is still a vendor comparison across models with different inference budgets and harnesses. Use it to select candidates, not to replace an internal evaluation.
5. Current pricing and availability, checked in September 2026
| Current API model | Cache-miss input (off / peak) | Output (off / peak) | Context | Notes |
|---|---|---|---|---|
deepseek-v4-pro |
$0.66 / $1.32 | $1.98 / $3.96 | 1M | V4 Pro; no image input |
deepseek-flash |
$0.15 / $0.30 | $0.60 / $1.20 | 1M | V4.1 Flash; image input |
All values are USD per million tokens. Off-peak rates are half of peak rates, and cache-hit input has a separate lower rate. Pricing and model routing can change, so estimate cost with actual token counts, cache-hit rates, and calling hours.
6. Selection guide: when to use V4, and when not to
Test V4 Pro for complex text and agent tasks
Use real repository and long-document tasks to measure success rate, tool calls, first-token latency, and total cost. A 1M limit does not mean sending an entire codebase in every request is efficient.
Test V4.1 Flash for throughput and vision input
The current API is cheaper, has a higher concurrency limit, and accepts images. Sample batch outputs before scaling, and verify structured output, retries, and failure handling.
- No regression set: define pass conditions before switching models.
- Strict compliance requirements: review data handling, terms, logs, region, and supplier contracts first.
- A stable model identity is mandatory: legacy names can be routed to newer models, so monitor versions and keep a rollback path.
7. Three ways to access V4
A. DeepSeek's official API (most direct)
It's OpenAI-compatible; change the base_url:
B. Anthropic-compatible API
DeepSeek also documents https://api.deepseek.com/anthropic. Check model mapping, tool calls, reasoning content, and SDK versions before moving production traffic.
C. Self-hosting via Hugging Face + vLLM
The official Hugging Face cards provide weights and deployment notes. Hardware depends on the checkpoint, precision, parallelism, context length, and concurrency. We did not run a local deployment and do not claim a fixed GPU count.
8. How V4 fits into this week's bigger trends
Our weekly highlights covered multimodal generation and systematic agent training. The V4 updates leave two concrete areas to track:
- Long context is now a standard service specification: the current pricing page still lists 1M context, but production value depends on recall, latency, and cost tests.
- Agent APIs keep expanding: the GA release added Responses API support and new agent benchmarks. The scores still depend on the published harness and sampling settings.
The multimodal status has changed: deepseek-flash now maps to V4.1 Flash, which the current pricing page lists with image support. V4 Pro remains text-only there.
9. How to compare cost
Dividing a public benchmark score by token price does not produce a useful cost-performance metric. Total cost includes success rate, retries, output length, cache hits, tool calls, latency, review time, and migration work.
A better comparison fixes a set of real tasks and measures first-pass success, tokens per successful task, and wall-clock time. Apply both peak and off-peak rates to those traces.
10. What to keep checking
The verifiable changes are DSA, token-wise compression, a 1M context limit, open weights, and a changing API product line. The April preview is no longer the whole current state.
Four follow-ups worth tracking:
- API routing: whether legacy names continue mapping to newer models.
- Pricing: changes to peak periods, cache rates, and concurrency limits.
- Model cards: checkpoint, dependency, license, and deployment changes.
- Regression tests: whether each update improves your own tasks.
Frequently asked questions
What's the core difference between DeepSeek V4 and V3?
V4 adds DSA (DeepSeek Sparse Attention) and token-wise compression, while raising the official service context limit to 1M tokens. The official technical report says V4-Pro uses 27% of V3.2's per-token FLOPs and 10% of its KV cache at 1M context.
How do I choose between V4 Pro and V4.1 Flash?
The current API maps deepseek-flash to the vision-capable V4.1 Flash and deepseek-v4-pro to V4 Pro. Both list a 1M context window. Flash has a higher concurrency limit and lower price, while Pro does not accept images. Choose with task-specific quality, latency, and cost tests.
How is the current DeepSeek V4 API priced?
DeepSeek currently uses peak and off-peak rates and separates cache-hit input, cache-miss input, and output tokens. Prices change, so this article no longer treats its April 2026 snapshot as current. Check the official Models & Pricing page before deployment.
Is DeepSeek V4 open-source? What license?
DeepSeek publishes V4-Pro and V4.1-Flash weights on Hugging Face, and their model cards list the MIT License. Before deployment, review the model card, dependency versions, hardware requirements, and license text.
How do I call the current DeepSeek V4 API?
Use https://api.deepseek.com with the OpenAI-compatible interface and select deepseek-flash or deepseek-v4-pro. The legacy deepseek-v4-flash name currently routes to V4.1 Flash. Recheck the official change log before production releases.
Want to track new AI models and trending papers daily? Paper Collector pulls Hugging Face papers and open-source projects every day with summaries.
Visit Paper CollectorPrimary sources: DeepSeek's V4 preview announcement, official change log, current pricing page, V4-Pro model card, and V4.1-Flash model card. Benchmark results are vendor-reported. We did not call the paid API, download weights, or reproduce tests.