DeepSeek V4 Flash 0731 Hits 50 on AI Intelligence Index — One Point Behind GPT-5.6 Luna, 60% Cheaper

One point behind Luna, at a fraction of the cost

The Artificial Analysis Intelligence Index now puts DeepSeek V4 Flash 0731 at 50 points — a 10-point jump over the V4 Flash preview from April 2026, and six points ahead of DeepSeek V4 Pro. GPT-5.6 Luna (max) sits at 51. The kicker: DeepSeek's cost per task is roughly 60 % lower, even after OpenAI cut Luna's prices by 80 % on July 30.

Let that sink in. OpenAI's most aggressive price reduction in its history — slashing Luna from $1.00/$6.00 to $0.20/$1.20 per million input/output tokens — and DeepSeek still undercuts them on real cost per task. The reason is twofold: a radically aggressive cache discount and more efficient token usage.

The pricing gap, explained in numbers

Here is the current pricing comparison between DeepSeek V4 Flash 0731 and GPT-5.6 Luna, converted to EUR as well:

Sources: DeepSeek API docs, The Decoder on OpenAI pricing. EUR conversions at ~0.93 USD/EUR.

The 98 % cache hit discount is the real killer feature. Every API provider offers some form of context caching — input tokens that haven't changed between requests get billed at a steep discount. The industry standard is 90 %. DeepSeek pushes that to 98 %, meaning a cache-hit input token costs just $0.0028 per million. For applications that send the same system prompt and document context repeatedly — think chatbots, code assistants, document Q&A — the effective cost drops into the noise floor.

Token efficiency: 12 % less output for the same work

DeepSeek didn't just cut prices. The new model uses fewer tokens to do the same job. According to Artificial Analysis, the 0731 version consumed roughly 206 million output tokens to complete the full Intelligence Index evaluation suite, compared to ~234 million for the previous V4 Flash — a 12 % reduction. That means you pay for fewer output tokens to get equal or better results. In a model where output tokens cost twice what input tokens cost ($0.28 vs $0.14), a 12 % token efficiency gain is real money at scale.

Biggest gains in agentic tasks

The benchmark breakdown shows where the upgrade really matters. On GDPval-AA v2, Artificial Analysis's evaluation for agentic real-world office work, V4 Flash 0731 jumps from 1,189 to 1,559 Elo — a 370-point gain. That makes it the second-highest open-weights model on that benchmark, behind only Kimi K3 (max, 1,687) and ahead of GLM-5.2 (max, 1,510).

Other agentic benchmarks show similar leaps: Terminal Bench 2.1 rises 17 points to 79 %, DeepSWE (software engineering) jumps from 7.3 % to 54.4 % — a sevenfold improvement — and Cybergym climbs from 38.7 % to 76.7 %. On coding benchmarks specifically, DeepSeek reports 68.7 % on their internal DSBench-FullStack test, up from 37 %.

These are not incremental tweaks. A model going from 7 % to 54 % on a software engineering benchmark suggests substantial architectural improvements to the agentic reasoning pipeline, not just more training data. The Hugging Face model card notes that this release includes a speculative decoding module (DSpark) and three reasoning effort levels — low, high, and max — that control how much deliberation the model spends before answering.

Fewer hallucinations, same accuracy

One of the more interesting findings from Artificial Analysis: the model's AA-Omniscience Hallucination Rate fell from 96 % to 84 % — a 12-point improvement — while overall accuracy stayed flat at 37 %. That is consistent with a model that is the same size as its predecessor (284B total / 13B active parameters) but has been fine-tuned to be more cautious about what it claims to know. The hallucination rate is now comparable to GPT-5.6 Terra (85 %) and Mistral Medium 3.5 (82 %).

For production use cases — customer support, legal document review, medical triage — reduced hallucinations matter more than a point or two of raw accuracy. A model that knows when to say "I don't know" is worth more than one that confidently invents facts 96 % of the time.

Open weights, MIT license: the European advantage

The model weights are available on Hugging Face under an MIT license — one of the most permissive open-source licenses. For European companies and developers, this matters in concrete ways:

• GDPR compliance: You can self-host the model on your own infrastructure (EU-based servers, your own GPU cluster), keeping all data within European borders. No need to send user queries to US or Chinese API endpoints.
• No vendor lock-in: Running the model locally via vLLM, SGLang, or Ollama means you control the inference stack. If DeepSeek changes pricing or terms tomorrow, your deployment keeps running.
• Batched deployment at fixed cost: For high-volume European services — e-government portals, university research platforms, enterprise document processing — the cost per query on self-hosted hardware approaches zero at scale, limited only by your electricity bill and GPU amortization.

The model retains the same architecture as the original V4 Flash: 284 billion total parameters, 13 billion active at inference time, and a 1-million-token context window. That 1M context is large enough to ingest an entire codebase, a full legal contract repository, or hundreds of pages of regulatory text in a single prompt — which is precisely the kind of workload European enterprises deal with under GDPR documentation requirements and AI Act compliance assessments.

The price war is real, and it's accelerating

Context matters here. OpenAI's July 30 price cut was itself a response to growing pressure from Chinese providers. Zhipu AI's GLM-5.2 had already closed in on closed-source coding scores, and DeepSeek's original V4 Flash launched at prices that made GPT-4o look like luxury goods. One day after OpenAI's dramatic cut, DeepSeek ships a model that still wins on price-to-performance. The cycle is compressing.

DeepSeek also announced on their pricing page that they will soon introduce peak/off-peak pricing — double rates during Beijing business hours (9:00–12:00 and 14:00–18:00 UTC+8). For European users, this is actually good news: peak hours in Beijing (UTC+8) correspond to 03:00–06:00 and 08:00–12:00 Central European Summer Time. Most European working hours fall outside Beijing's peak window.

What this means in practice

If you are building an AI product in mid-2026, the floor just dropped again. A model scoring 50 on the Intelligence Index — competitive with GPT-5.6 Luna and Gemini 3.6 Flash — now costs less than €0.50 per million tokens all-in when you factor in cache hits. That is cheap enough to power every chatbot, every code review agent, every document summarizer in your stack without users noticing the bill.

For European startups building on tight budgets, the math is simple: self-host DeepSeek V4 Flash 0731 on a couple of rented GPU nodes in Frankfurt or Amsterdam, and you have API-grade intelligence with zero per-token marginal cost. That is the kind of economics that makes AI-first products viable in markets where US venture capital isn't flowing freely.