For developers tired of watching OpenAI and Anthropic invoices climb each quarter, the calculation just changed overnight.
DeepSeek released its V4 series on 24 April, headlined by V4 Pro — a 1.6 trillion parameter behemoth that landed at number three on the independent Artificial Analysis Intelligence Index, ahead of Anthropic's Claude Opus 4.7 and level with Google's Gemini 3.1 Pro.
The eyebrow-raiser is the price tag. V4 Pro costs $1.74 per million input tokens. OpenAI's new GPT-5.5 charges $5. Claude Opus 4.7 also charges $5. That is roughly a 65 to 73 per cent discount on Western flagships for comparable intelligence, and the lighter V4 Flash variant runs at just $0.14 per million input tokens — territory that effectively drops frontier-level reasoning into hobbyist budgets.
What hybrid attention actually means
Strip away the buzzwords and the breakthrough is straightforward. Large language models normally pay a quadratic compute penalty as the input grows: double the context, quadruple the bill. That is why long context has historically meant expensive context.
DeepSeek's Hybrid Attention architecture combines two compression techniques — Compressed Sparse Attention and Heavily Compressed Attention — so the model pays full attention only to the bits that matter and a cheaper, summarised attention to everything else. The result, according to DeepSeek's own figures, is a one-million-token context window running on just 27 per cent of the compute previous architectures would need.
For developers, that translates into feeding entire codebases, lengthy legal bundles or multi-document research dossiers into a single prompt without watching costs explode.
Where it sits against the West
V4 Pro is not the smartest model on the planet this week — that crown belongs to GPT-5.5, which scored 60 on the Artificial Analysis Intelligence Index against V4 Pro's 58. But on the benchmarks startups actually care about, the gap closes fast.
V4 Pro scored a perfect 120/120 on the Putnam-2025 frontier mathematics benchmark and 93.5 per cent on LiveCodeBench, level with or ahead of Claude Opus 4.7 — the model that powers GitHub Copilot's coding agent.
Pair that with the open-weight licensing DeepSeek has stuck to throughout the V-series, and you have a frontier-class model that startups can self-host, fine-tune for niche domains, or run behind their own firewall. That is something neither OpenAI nor Anthropic currently offer at this performance tier.
The market signal
The V4 launch caps a frantic late-April release cycle that also produced GPT-5.5, xAI's Grok 4.3 and Moonshot's Kimi K2.6. But it is DeepSeek that has rewritten the economics.
Western providers have spent eighteen months arguing that frontier intelligence is intrinsically expensive. V4 Pro contradicts that thesis on day one. Expect aggressive price cuts from the incumbents within weeks, and expect founders building on OpenAI or Anthropic APIs to start running cost-comparison spreadsheets they had previously been too polite to open.
For UK and European startups operating on tight runways, the practical implication is the simplest one: the same product, built on V4 Pro instead of GPT-5.5, just got roughly two-thirds cheaper to run. That is not an incremental upgrade. It is the kind of step-change that reshapes which AI companies survive their next funding round.
Whether DeepSeek's pricing is sustainable, or a strategic loss-leader to capture market share before regulators tighten Chinese model exports, is the question that will define the rest of 2026.
