On April 24, 2026, DeepSeek released V4 Pro — a 1.6T-parameter open-weight model that ties or beats GPT-5.6 Sol, Claude Opus 5, and Gemini 3.1 Pro on coding benchmarks, at roughly one-thirtieth the price. On May 22, the 75% launch discount became permanent.
DeepSeek V4 Pro is the best open-weight coding model released so far. It matches Claude Opus 5 on SWE-bench Verified (80.6% vs 80.8%), leads every model on LiveCodeBench (93.5%) and Codeforces (3,206), ships a 1M-token context window, and costs $0.87 per million output tokens — about 29× cheaper than Opus and 34× cheaper than GPT-5.6 Sol. Weights are open under MIT license on Hugging Face. Free access via Abacus AI ChatLLM, chat.deepseek.com, and OpenRouter.
DeepSeek V4 vs GPT-5.6: Coding Benchmarks
| Benchmark | DeepSeek V4 | GPT-5.6 | Winner |
|---|---|---|---|
| SWE-bench | 80.6% | 78.2% | DeepSeek |
| Codeforces | 3206 | 3100 | DeepSeek |
| MMLU | 89.4 | 90.1 | GPT-5.6 |
Key insight: DeepSeek V4 beats GPT-5.6 on coding benchmarks while being completely free and open-source. You can self-host it or use the free API. The only cost is if you need Claude-level general reasoning.
What Is DeepSeek V4 Pro?
V4 Pro is the flagship in DeepSeek's V4 family, released April 24, 2026 by the Hangzhou lab behind V3 and R1. It's a 1.6 trillion parameter Mixture-of-Experts model that activates only 49 billion parameters per token — the knowledge capacity of a giant model at the inference cost of a much smaller one. A leaner sibling, V4-Flash (284B total, 13B active), covers high-volume, latency-sensitive workloads.
What sets it apart from prior open-weight releases isn't just price — it's capability. On SWE-bench Verified it scores 80.6%, just 0.2 points behind Claude Opus 5 (88.6% SWE-bench) and 5.7 ahead of GPT-5.6 Sol (74.9%). On LiveCodeBench, built specifically to resist training-set contamination, its 93.5% is the highest any model has posted. On Codeforces, its 3,206 rating beats GPT-5.6 Sol (3,168) and Gemini 3.1 Pro (3,052).
Training ran on 32+ trillion tokens using Huawei Ascend 950 and Cambricon accelerators rather than Nvidia GPUs — the first frontier-class model trained entirely outside the Nvidia ecosystem. The practical upshot: weights are open under MIT license, and you can self-host if you have the hardware. The full FP8 checkpoint runs ~862GB, needing ~900GB of VRAM in production (8×H100 or MI300X); quantized builds fit on 2–4 consumer GPUs.
Three Innovations That Make 1M Context Viable
Million-token context is a common marketing bullet point; using it economically usually isn't, since standard attention scales quadratically. V4 solves this with three innovations detailed in its technical report. The result: at 1M tokens, V4 Pro uses just 27% of predecessor V3.2's per-token compute and 10% of its KV-cache memory.
Hybrid attention (CSA + HCA)
Tiers memory: recent tokens get full attention, mid-range tokens get compressed attention, distant tokens get heavily compressed summaries. This is what makes long-context inference a standard feature rather than a premium add-on.
mHC connections
Constrains residual mixing matrices to the Birkhoff Polytope via Sinkhorn-Knopp, capping signal amplification at 1.6× instead of the 3,000× blowups seen in earlier experiments — what made stable training at 1.6T parameters possible.
Muon optimizer + MoE routing
Replaces AdamW with an orthogonalization-based optimizer, paired with a two-stage post-training pipeline — independent expert specialization, then unified consolidation — for cleaner reasoning chains than vanilla RLHF.
V4 Pro also ships three reasoning modes — Non-Think, Think High, and Think Max — plus tool calls that now work inside thinking mode, a fix over the R1 series that matters for agentic workflows.
Where V4 Pro Wins — and Where It Trails
DeepSeek is unusually candid about where V4 Pro falls short. Here's a depth read on five benchmarks — closer to the surface means a stronger relative score, teal markers are wins, coral markers are where a rival leads:
GPQA and MMLU-Pro list V4 Pro's own score — Gemini 3.1 Pro leads both. Terminal-Bench is GPT-5.6 Sol's clearest advantage (75.1% vs 67.9%).
| Benchmark | V4 Pro | Opus 4.8 | GPT-5.6 Sol | Gemini 3.1 Pro | Winner |
|---|---|---|---|---|---|
| SWE-bench Verified | 80.6% | 80.8% | 74.9% | 80.6% | Opus / V4 |
| LiveCodeBench | 93.5% | ~89% | ~86% | 91.7% | V4 Pro |
| Codeforces rating | 3,206 | N/R | 3,168 | 3,052 | V4 Pro |
| Terminal-Bench 2.0 | 67.9% | 65.4% | 75.1% | 68.5% | GPT-5.6 Sol |
| GPQA Diamond | 90.1% | 91.3% | 93.0% | 94.3% | Gemini |
| HLE (Pass@1) | 37.7% | 40.0% | 39.8% | 44.4% | Gemini |
| Input price ($/M) | $0.435 | $5.00 | $5.00 | $3.50 | V4 Pro |
| Output price ($/M) | $0.87 | $25.00 | $30.00 | $21.00 | V4 Pro |
| Open weights | MIT ✓ | ✕ | ✕ | ✕ | V4 Pro |
The pattern: V4 Pro is the undisputed coding champion, Gemini 3.1 Pro holds the knowledge and reasoning edge, and GPT-5.6 Sol leads the hardest agentic terminal tasks. For everyday code generation, debugging, and reasoning, V4 Pro is at or near the top at a price that makes the comparison feel almost unfair.
The 75% Cut That Became Permanent
V4 Pro launched at $1.74/M input and $3.48/M output — already aggressive next to Opus's $5/$25 and GPT-5.6 Sol's $5/$30. DeepSeek applied a 75% promotional discount immediately, set to expire May 31. On May 22, 2026, DeepSeek confirmed the discount would not roll back — the promotional rate is now the permanent list price, with no expiry.
Concretely: a pipeline processing 10M output tokens a day costs roughly $261/month on V4 Pro, versus about $7,500/month on Opus 4.8 and $9,000/month on GPT-5.6 Sol. One caveat: independent reviewers note V4 Pro is verbose, burning 4–5× the median output tokens of rivals on equivalent tasks — the real bill narrows the gap slightly, though it stays dramatic.
Where to Run It — Free & Paid
V4 Pro runs as "Expert Mode" on DeepSeek's web chat — free for individuals, no API key or setup. V4-Flash is available as "Instant Mode."
V4 Pro is on the ChatLLM Teams platform, selectable from the model dropdown alongside GPT, Claude, Gemini, and dozens more.
V4 Pro and V4 Flash are live with multi-provider routing, useful for failover and price comparison, at the same $0.435/$0.87 rates.
Open weights under MIT at deepseek-ai/DeepSeek-V4-Pro. Full FP8 checkpoint ~862GB; quantized builds fit 2–4 consumer GPUs.
OpenAI- and Anthropic-compatible endpoints — a drop-in swap for Claude Code, Cursor, or any OpenAI SDK app by changing the base URL.
Day-0 NIM containers on build.nvidia.com. Blackwell GB200 NVL72 delivers 150+ tok/s/user at 30× better perf/watt than H200.
Pros & Cons
Pros
- Best coding model on the market — leads or ties every major coding benchmark
- 29–34× cheaper output than Opus/GPT, with near-free cached-context input
- Open weights (MIT) — self-host, fine-tune, audit, deploy on-prem
- 1M-token context that's economically viable, not just a spec sheet number
- Tool calls now work inside thinking mode — unlike the R1 series
- Free access through chat.deepseek.com, Abacus AI, and OpenRouter
Cons
- Text-only — no vision or audio, unlike GPT-5.6 Sol, Claude, or Gemini
- Less reliable on 100+ turn autonomous agent loops than Claude Opus
- Verbose — burns 4–5× the output tokens of rivals on equivalent tasks
- Trails Gemini 3.1 Pro on GPQA, HLE, and SimpleQA knowledge benchmarks
- Labeled preview, not GA — behavior may shift before stable release
- High hallucination rate (94% on AA-Omniscience) when uncertain
Who Should Use It — and Who Shouldn't
Use V4 Pro if you're…
- A cost-sensitive developer or startup where LLM spend is a real line item
- Doing heavy coding, PR review automation, or algorithmic reasoning
- Running high-volume agent pipelines where per-token cost dominates
- Needing data sovereignty via self-hosting on your own infrastructure
- Working with long documents where cached 1M-token context pays off
Skip it if you need…
- Multimodal capability — image, screen-OCR, or audio input
- The longest, most error-resilient agentic loops (Claude still leads there)
- Research-grade knowledge work (Gemini 3.1 Pro is the stronger pick)
- Deep integration with the OpenAI ecosystem (Assistants API, real-time voice)
- Guaranteed production stability — this is a preview, not GA
Frequently Asked Questions
The Bottom Line
DeepSeek V4 Pro is the most important AI release of 2026 so far — not because it's the smartest model, but because it breaks the price-capability curve the LLM market has followed since GPT-4. A model with open weights, an MIT license, and frontier-tier coding performance now costs less than a dollar per million output tokens. It's verbose, it hallucinates when uncertain, and Claude still wins the longest agentic loops — but for the coding and reasoning work that fills most developers' daily usage, it's at or near the top, at a price that makes the comparison feel broken.
You scroll to the benchmark tables one last time, the 80.6% SWE-bench score still sinking in.
🔬 Testing Methodology
- Test period
- 7+ days of hands-on use
- Pricing
- Manually verified at time of writing
- Affiliate
- Some links may earn a commission — full disclosure