On July 31, 2026, DeepSeek promoted V4 Flash from preview to official release. The new checkpoint, DeepSeek-V4-Flash-0731, costs $0.14 per million input tokens and $0.28 per million output tokens. On every agentic benchmark DeepSeek publishes, it outperforms the V4-Pro preview — a model with nearly 4x the activated parameters. MIT-licensed open weights dropped the same day.

If you build with AI models, this is the week your cost-per-query math changed.

What V4 Flash actually is

The architecture is unchanged from the preview: 284 billion total parameters with roughly 13B activated per forward pass, using DeepSeek's CSA+HCA (Compressed Sparse Attention + Hybrid Chunked Attention) design. What changed in the 0731 build was the post-training — specifically re-tuned for agentic tasks — plus native support for DeepSeek's Responses API and Codex integration. The DSpark speculative-decoding module ships attached by default.

The Artificial Analysis Intelligence Index scored V4 Flash 0731 at 50, a 10-point jump over the April preview. That puts it ahead of models from OpenAI and Anthropic that charge 20 to 50 times more per token.

The economics are real. DeepSeek's pricing page lists V4 Flash at $0.14/$0.28 per million tokens (input/output). Their cache-hit discount — which applies when your prompt shares context with previous requests — drops cached input tokens to $0.014/M. For agent loops that reuse system prompts and tool schemas across turns, that 98% discount becomes the dominant cost factor.

The Sonnet 5 problem nobody's talking about

While DeepSeek is racing to the bottom on price, Anthropic is quietly raising the effective cost of Claude Sonnet 5. The intro pricing — $2/M input, $10/M output — ends August 31, 2026. After that, Sonnet 5 reverts to $3/M input, $15/M output. That part is public.

What is less public: Anthropic's new tokenizer, which shipped with Sonnet 5, Opus 4.8, and Fable 5, produces significantly more tokens from the same text than the previous one. Testing by PlayCode across 16 real-world fixtures — English prose, JavaScript, Python, TypeScript, Rust, HTML, JSON tool schemas, and agent system prompts — found the new tokenizer generates roughly 30% more tokens from identical content.

The worst case was TypeScript, where the same file produced 73% more tokens on Claude than on GPT's o200k tokenizer. English prose came in at 34% more. An agent system prompt of about 42,000 characters tokenized to 14,953 tokens under the new tokenizer versus 10,761 under the old one — a 39% increase.

During the intro period, the lower list price roughly compensates for the extra tokens. After September 1, you are paying the old rate for 30% more tokens per request. In effective terms, Sonnet 5 post-intro behaves like $4.50/M input and $22.50/M output when measured against GPT's tokenization. That is not a price cut. That is a price increase wearing a disguise.

What this means for anyone building with AI

The math is getting harder to ignore. DeepSeek V4 Flash at $0.14/M with a 98% cache-hit discount on cached context versus Sonnet 5 at an effective $4.50/M after its intro window closes — that is a 32x difference on the input side alone.

This does not mean Sonnet 5 is the wrong choice. Anthropic's models still lead on certain safety-critical and instruction-following benchmarks. Claude's tool-use reliability in production agent loops remains excellent. If your workload depends on those qualities, the premium may be justified.

But for a growing class of tasks — code generation, document summarization, data extraction, multi-step agent orchestration — the performance gap between a $0.14/M model and a $4.50/M model has narrowed to the point where the cost difference dominates the decision. DeepSeek's open weights also mean you can self-host V4 Flash on your own infrastructure. Someone already benchmarked a 4-node DGX Spark cluster running V4 Flash 0731 at roughly 2,500 tokens/second prefill and 90 tokens/second decode, with a total hardware cost around $20,500. If you are processing serious volume, that breaks even against API costs fast.

The pricing war is the real story

Neither DeepSeek nor Anthropic is operating in a vacuum. OpenAI's GPT-5 line, Google's Gemini models, and a half-dozen other players are all competing on the same axis: capability per dollar. The trend line is clear — frontier-level performance is becoming a commodity, and the margin is shifting to infrastructure efficiency, caching, and developer experience.

What changed on July 31 is not just that a cheaper model exists. It is that the cheaper model is provably better at the tasks that matter most for the next generation of AI applications — agentic workflows, tool use, and long-context reasoning — while simultaneously being open-weight and self-hostable.

If you have not re-evaluated your model costs in the last month, now is the time. The floor just dropped.

Sources