DeepSeek V4 Pro & Flash: $0.28 Output Frontier That Goes Dark July 24 — July 2026

> DeepSeek V4 Pro and Flash launch July 10 with 1M context, $0.28 output Flash tier, competitive frontier before legacy shutdown July 24.

Audio version coming soon

Verified by Essa Mamdani

Introduction: The Cheapest Frontier Yet?

July 10, 2026. DeepSeek drops two models at once: DeepSeek V4 Pro and DeepSeek V4 Flash, and in the same announcement, sets a July 24 shutdown date for DeepSeek V2.5 and V3 legacy endpoints. It's classic DeepSeek - ship cheap, move fast, deprecate faster.

The headline number: Flash at $0.28 per million output tokens. Input at $0.07/M. Pro at $0.42 input / $0.72 output with full frontier reasoning. Both with 1M context window standard, 128K output.

This is 85% cheaper than GPT-4o pricing and 60% cheaper than Claude 3.5 Sonnet at comparable intelligence. For the second time, DeepSeek resets the price floor.

What is DeepSeek V4? Pro vs Flash

DeepSeek V4 is the successor to V3.1 and R1/V3 lineage. It's a 671B total MoE, 37B active for Pro, and a distilled ∼215B total, 12B active for Flash. Yes, same total parameter family as V3, but with major architecture upgrades.

V4 Pro (DeepSeek-V4-0324-Pro):

671B MoE, MLA (Multi-head Latent Attention) v2
1M context via DeepSeek's own YARN+NSA (Native Sparse Attention)
Hybrid reasoning with controllable thinking tokens
MMLU 89.1%, LiveCodeBench 71.2%, AIME 2025 79.4%
Ideal for: Coding agents, deep research, finance

V4 Flash (DeepSeek-V4-Flash):

Distilled from Pro using 10M SFT + RL samples
3.2x faster than Pro: ~142 tok/s on H800
Pricing: $0.07 input / $0.28 output - cheapest frontier-class model in July 2026
Same 1M context, 92% of Pro's capability per DeepSeek evals
Ideal for: High-throughput apps, batch processing, RAG, startups scaling

Both support DeepSeek-V4 Think mode and JSON + Function Calling v2, now compatible with Anthropic tool use schema.

Benchmarks: Near GPT-5 at 1/10th Cost

DeepSeek's own chart puts V4 Pro just under GPT-5/5.6 family and Grok 4.5 but ahead of Llama 4 Behemoth in coding:

MMLU: Pro 89.1%, Flash 87.3%
GPQA Diamond: Pro 71.4%, Flash 68.9%
HumanEval: Pro 92.1%, Flash 89.4%
SWE-Bench Verified: Pro 58.3% (vs Claude 4 Sonnet 62%, vs Avocado 1.1 55%)
RULER 1M: Pro 88.7%, Flash 84.2%
SimpleQA: Pro 44.2% - improved hallucination resistance

The value argument is simple: If Pro scores within 3% of GPT-5 on your workload but costs 80% less, you switch. Flash is for when you need Meta Avocado 1.1-style cost savings but with DeepSeek's better Chinese and math performance.

The July 24 Shutdown: Migrate or Break

DeepSeek announced that V2.5, V2, and V3 (non 0324) will go dark on July 24, 2026. All API requests will 404. This is aggressive - only 14 days migration.

What you must do:

Change model="deepseek-chat" alias now points to V4 Pro (was V3). Test ASAP.
model="deepseek-reasoner" still points to R1 but will move to V4 reasoning variant in August.
Update prompt: V4 is less tolerant of system prompt constraints than V3. Remove excessive "You MUST" language.
Cost recalculation: Flash is new - if you were using V3 for cheap tasks, Flash is cheaper and better.

This is the second time DeepSeek has forced migration quickly. It's painful but keeps their infrastructure lean.

Comparison to Other July 2026 Launches

July is stacked with efficiency plays:

vs Qwen3-Next 80B A3B MoE 1M Context: Qwen is open-weight and cheaper to self-host. DeepSeek V4 is closed/API-first but more capable. If you need open, choose Qwen. If you need cheap API frontier, choose DeepSeek Flash.
vs PrismML Bonsai 27B 1-bit phone model: Bonsai is for local private inference. DeepSeek Flash is for cloud scale - different leagues.
vs LiteRT JS Browser AI: Combine them - LiteRT for on-device classification, DeepSeek Flash for fallback heavy lifting.
vs SEA-LION v45 multilingual and Fugu Ultra: DeepSeek leads on Chinese/English/STEM, SEA-LION leads on SE Asian languages, Fugu leads on Japanese multi-agent.

Common theme: Every lab is chasing cheap frontier. Alibaba via sparsity, DeepSeek via MoE distillation and scale, PrismML via 1-bit quantization.

Developer Experience

DeepSeek V4 API is OpenAI-compatible:

python
1from openai import OpenAI
2client = OpenAI(
3  base_url="https://api.deepseek.com",
4  api_key="sk-..."
5)
6res = client.chat.completions.create(
7  model="deepseek-v4-flash", # or deepseek-v4-pro
8  messages=[{"role":"user","content":"Explain Qwen3-Next vs DeepSeek V4"}],
9  max_tokens=8192
10)

New in V4:

Context Caching: 75% discount for cached input (1M context becomes affordable)
Mid-thinking tool calls: Model can call tools during thinking phase
128K output with streaming thinking tokens

Throughput via DeepSeek's own infra is reportedly improved - no more rate limit hell from early 2025.

Should You Adopt?

Yes if:

You have high token volume (Flash saves $10K+/month at scale)
You use 128K+ context regularly
You're comfortable with China-hosted API (or use third-party proxies like Together/Anyscale once they host)

No if:

You need absolute top reasoning (GPT-5 Sol still leads) or Grok 4.5 terminal bench performance
You require EU data residency (DeepSeek doesn't offer yet)
You can't handle July 24 breaking change risk

Final Take

DeepSeek V4 Pro and Flash aren't about being #1 on benchmarks. They're about being #1 on the price-to-intelligence frontier curve. At $0.28 output, Flash makes frontier AI feel like a commodity - similar to what Qwen3-Next does for self-hosters.

With legacy shutdown July 24, DeepSeek is forcing the market forward. If you haven't benchmarked V4 Flash against Haiku 3.5 or GPT-4o mini, do it this week. You might cut your LLM bill by 70%.

#ai-models#deepseek#cheap-ai