DeepSeek has opened the public beta of its V4-Flash API to all developers, releasing the same week OpenAI slashed GPT-5.6 Luna prices by 80%. The model keeps the 284-billion-parameter mixture-of-experts architecture from the preview but has been fully retrained with a focus on agentic tasks, the most contested corner of the developer market.
🔍 THE BOTTOM LINE
DeepSeek’s V4-Flash costs $0.14 per million input tokens and $0.28 per million output. OpenAI’s newly discounted Luna sits at $0.20 input and $1.20 output. Even after the steepest single-day price cut in OpenAI’s history, DeepSeek still undercuts Luna by 30% on input and more than four times on output. Anthropic’s cheapest model, Haiku 4.5, charges $1 input and $5 output — seven times DeepSeek’s rate. The gap is narrowing. It has not closed.
What V4-Flash Actually Delivers
The V4-Flash-0731 build ships with a one-million-token context window, 384K maximum output, and an OpenAI-compatible endpoint at api.deepseek.com. Any team already calling GPT-5.6 can switch by changing one config line. The model also supports the Responses API and Codex-style workflows, which matter for agentic coding use cases where a model writes, tests, and iterates on code autonomously.
On SWE-bench Verified, a benchmark that measures whether an AI can fix real GitHub issues, DeepSeek reports V4-Flash-Max at 79%, up from 69% for the previous V3 generation. That is within range of Western frontier models. Scale’s SEAL leaderboard does not yet carry a verified DeepSeek V4 entry, so the number warrants scrutiny, but developer reactions on X have been positive enough that the figures do not feel invented.
Execution speed matters for agentic pipelines, where a model might make dozens of tool calls in a single task. Artificial Analysis puts V4-Flash at roughly 165 seconds per task on average, faster than all three Claude Opus models.
The Price War’s New Chapter
This release lands in a more combustible environment than past DeepSeek drops. The V3 and R1 announcements in early 2025 sent AI infrastructure stocks down sharply because they demonstrated frontier-level performance without frontier-level compute budgets. This time the scope is narrower — a public API beta, not a full model release — but the timing is pointed.
OpenAI cut GPT-5.6 Luna to $0.20 input and $1.20 output on July 30, an 80% reduction from its previous rate. Anthropic’s Haiku 4.5 remains at $1/$5. DeepSeek’s pricing table now reads $0.14/$0.28, with cache hits dropping to $0.0028 per million input tokens — a 98% discount on repeated prefixes.
The practical implication for enterprises sitting on six-figure OpenAI contracts: the question is no longer whether DeepSeek is a credible alternative. It is how much switching cost they are willing to absorb, and how long they plan to wait.
Open Weights Change the Calculation
DeepSeek’s V4-Flash is not just cheaper to call. The weights are downloadable. A company that wants to run the model on its own infrastructure can do so without paying per token at all. The public API is the convenience layer, not the only option.
That is a different competitive posture than OpenAI or Anthropic can offer. It is also the reason the price comparison, while real, only tells part of the story. When Together AI closed $800 million at an $8.3 billion valuation in July, the company disclosed that open-weight model usage on its platform had tripled over the prior twelve months and annual bookings had crossed $1.15 billion. Enterprises are voting with their invoices.
This builds on a pattern we have tracked all year. DeepSeek V4 matching GPT-5.5 at 86% less cost in April, DeepSeek making its 75% discount permanent in May, and now the V4-Flash public beta landing within 24 hours of OpenAI’s own price cut. The cadence is deliberate.
What Comes Next
The legacy deepseek-chat and deepseek-reasoner model names were deprecated on July 24, signalling consolidation around the V4 line. The V4-Pro, which carries 1.6 trillion total parameters with roughly 49 billion active per token, is still unreleased. DeepSeek declined to give a timeline.
If V4-Pro clears V4-Flash’s benchmarks by a meaningful margin, the labs that benchmarked against V4 preview in their enterprise evaluations will need to run the numbers again. The cheaper, faster model already beating the older flagship across several task categories is the kind of result that makes enterprises reconsider which tier they actually need.
NZ Angle
New Zealand’s AI-dependent businesses — and there are more than the local tech press acknowledges — face the same switching-cost calculation as everyone else. The difference is that NZ companies typically run leaner teams with smaller AI budgets, which makes a 4x output price gap harder to ignore. A startup in Auckland paying $2,000 a month for GPT-5.6 API calls could cut that to $500 by pointing at DeepSeek’s endpoint. The open-weights option also matters for NZ’s sovereign AI conversation: if the model runs on your own hardware, your data never leaves the country.
❓ FAQ
Is DeepSeek V4-Flash available right now? Yes. The public beta is live at platform.deepseek.com. Developers can generate API keys and start calling the OpenAI-compatible endpoint immediately.
How does V4-Flash compare to V4-Pro? V4-Flash is the faster, cheaper model with 284 billion parameters. V4-Pro, with 1.6 trillion parameters, has not been released yet. DeepSeek has not given a timeline for the Pro launch.
Can I run DeepSeek V4-Flash on my own hardware? The model weights are downloadable. You need sufficient GPU memory to run a 284-billion-parameter MoE model, which typically means multiple high-end GPUs or a cloud instance with adequate VRAM.
Does switching from OpenAI require rewriting my code? No. The endpoint is OpenAI-compatible. Most teams can switch by changing the base URL and API key in their existing configuration.
Is DeepSeek’s benchmark of 79% on SWE-bench Verified reliable? The number is self-reported and Scale’s SEAL leaderboard has not yet verified it. Developer reactions on X have been positive, but independent verification is still pending.
🔍 THE BOTTOM LINE
DeepSeek just made the API war a three-front conflict: OpenAI on quality, Anthropic on safety, DeepSeek on price and openness. The V4-Flash public beta does not need to beat GPT-5.6 on every benchmark to win. It needs to be good enough that the 4x price gap — and the ability to self-host — tips enterprise decisions. This week, it cleared that bar.
📰 Sources
- TechNode
- Startup Fortune
- DeepSeek API Docs
- Bloomberg Technology
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
