DeepSeek V4 Flash: Strong Performance at a Fraction of the Cost

DeepSeek published DeepSeek-V4-Flash-0731 on July 31, 2026, moving the official V4-Flash API into public beta the same day. The timing was pointed: OpenAI had cut GPT-5.6 Luna's price 80% the previous day, and DeepSeek's response was to ship a meaningfully improved model at rates that still came in well below even the reduced OpenAI pricing. The model is a 284-billion-parameter Mixture-of-Experts architecture with 13 billion active parameters per token, a 1-million-token context window, and MIT-licensed weights on Hugging Face. It's not a frontier model in raw capability terms, but it occupies a genuinely useful position on the cost-intelligence curve.

Architecture

DeepSeek V4 Flash is a sparse MoE model. The total parameter count is 284 billion, but only 13 billion engage during any given forward pass. A router network examines each token and dispatches it to the most relevant expert sub-networks, leaving the rest dormant. This sparse activation is what makes the model fast and cheap to run relative to its knowledge capacity: you get inference costs closer to a 13B dense model while drawing on the knowledge encoded across 284B parameters.

The published checkpoint on Hugging Face reports 304 billion parameters, which includes the DSpark speculative decoding draft module on top of the base model. DSpark predicts blocks of tokens ahead, which the main model then verifies in parallel, improving generation throughput. The 0731 release is a re-post-training update from the April preview, not a new architecture. The improvements come from a new post-training pipeline focused on coding, agents, reasoning, and tool use.

Context length is 1 million tokens with a maximum output of 384,000 tokens. The model supports tools, reasoning, structured output, and temperature control.

Pricing

At launch, DeepSeek priced V4 Flash at $0.14 per million input tokens (cache miss), $0.0028 per million on a cache hit, and $0.28 per million output tokens. DeepSeek updated rates on August 16, 2026 to $0.22 input and $0.66 output, with cache hits dropping to $0.007 per million input tokens. Peak pricing (double rates) applies during 01:00–04:00 and 06:00–10:00 UTC.

The model is also available through OpenRouter and around 50 other third-party providers at varying rates.

Benchmark performance

On the Artificial Analysis Intelligence Index, which aggregates nine evaluations including GDPVal-AA v2, Terminal-Bench, and SciCode, DeepSeek V4 Flash scores 52.

This places it alongside GLM-5.2 (53) and Gemini 3.6 Flash (52). It's not at the top of the chart, trailing models like Claude Opus 5 and GPT-5.6, but it's a competitive score for a model at this price point. DeepSeek generates output at 107 tokens per second through the API, well above the median for models of similar size.

One note on evaluation costs: Artificial Analysis reported $323 to evaluate DeepSeek V4 Flash at max effort on the full Intelligence Index suite. The comparable figure for models like GLM-5.2 and Gemini 3.6 Flash is considerably higher, and the gap widens further for higher-tier models, which is where the cost advantage becomes most visible in practice.

Cost per task

Looking at the weighted average cost per individual task within the Intelligence Index:

DeepSeek V4 Flash comes in at approximately $0.03 per task, making it one of the cheapest options at its performance level. It's notably verbose in output: it generates about twice the median token count across the evaluation suite, which makes the low per-task cost more impressive rather than less.

Intelligence vs. cost

On the cost-versus-intelligence scatter plot, DeepSeek V4 Flash sits on the edge of the high-intelligence, low-cost quadrant. Models with slightly higher intelligence scores in that zone carry substantially higher per-task costs, which makes the marginal performance gap hard to justify for most workloads.

Building a personal finance dashboard

Eight models were given the same prompt in an empty folder: build a personal finance management dashboard with account tracking across savings, checking, investments, and debts; transaction history and spending analysis; financial goal monitoring; fund transfer functionality; and seeded backend data. No authentication required.

DeepSeek V4 Flash

DeepSeek V4 Flash produced a functional multi-page dashboard with overview, accounts, transactions, and goals views. The UI is dark-themed and includes a cash flow chart and spending donut chart. Responsiveness has some issues and some card sizing is off, giving it a slightly dated feel. The core functionality works: fund transfers and account management operate correctly.

The more notable result is the backend: DeepSeek V4 Flash chose Node SQLite for persistence, making it one of the few models in the test to use a real database rather than an in-memory store or a JSON file. The stack was React with Vite for the frontend and Node.js with Express for the server.

Cost: $0.08. Time: 23 minutes 48 seconds.

Comparison results

By cost:

DeepSeek V4 Flash at $0.08 came in second, just behind GPT-5.6 Luna at $0.06 (which was running at a temporary 50% discount). Kimi K3 was next at $0.58, then GPT-5.6 Sol at $0.90.

By speed:

DeepSeek V4 Flash's 23-minute runtime was near the bottom of the group. GPT-5.6 Sol finished in 3 minutes 42 seconds, GPT-5.6 Luna in 6 minutes 57 seconds. Speed is where DeepSeek V4 Flash falls short most clearly.

What the other models did

GPT-5.6 Luna produced arguably the best-looking UI in the test but stored data in memory with no persistence. GPT-5.6 Sol was the fastest and generated a polished professional dashboard, also without a persistent database. GLM-5.2 used Next.js with Prisma ORM, the most production-oriented stack in the test, but took 47 minutes. Claude Opus 5 used an in-memory database despite costing $9.65. Gemini 3.6 Flash used a JSON file for storage at a cost of $1.79. Kimi K3 used vanilla JavaScript and a JSON file.

The pattern is notable: DeepSeek V4 Flash was one of the few models to make a reasonable database choice, which matters for any application that needs to persist state across requests.

Where it fits

DeepSeek V4 Flash is well-suited for high-volume workloads where per-call cost accumulates and where the task doesn't demand frontier-level reasoning. The 1-million-token context window makes it viable for long-document processing, large codebase analysis, and extended agent sessions. The MIT license and open weights let you run it locally or through a provider of your choice without vendor lock-in.

The weaknesses are generation speed and UI quality. If latency matters, GPT-5.6 Sol or Luna will outperform it significantly. If frontend polish is the priority, the OpenAI models in this test consistently produced better-looking results. For backend logic and database architecture, DeepSeek V4 Flash performed comparably or better than models costing ten to a hundred times more per session.

A practical hybrid approach: use a fast model like GPT-5.6 Luna or Sol for frontend generation, use DeepSeek V4 Flash for backend logic, database schemas, and server-side code, and combine the outputs. The combined cost stays low while the final result benefits from each model's relative strengths.