DeepSeek V4 Flash 0731 on-premises: what the published scores say, what a token costs, and what sovereignty changes

The independent 2026 benchmark of the three best enterprise AI models: DeepSeek V4 Flash 0731 (284B, 13B active, 1M context, open weights) vs Claude Opus 5 (Anthropic frontier) vs GPT-5.5 (OpenAI frontier). Terminal Bench, NL2Repo, Cybergym, DeepSWE, Toolathlon, cost per token, latency and sovereignty.

By QDNA · Published on 2 August 2026 · Updated on 8 October 2026 · 10 min read

Updated on 8 October 2026: reference price of the DGX Spark raised to €7,000 excl. VAT and of the DGX Station to €150,000 excl. VAT; daily costs, costs per million tokens, ratios to the DGX Spark and break-even thresholds recomputed with the same formula.

Updated on 26 September 2026: DeepSeek retired V4 Flash 0731 from its API on 10 September 2026: the API rows now use the price list of its successor V4.1 Flash ($0.15 and $0.60 off-peak); Claude Opus 5.5 ($4 and $20), GPT-6 Astra and GPT-6 Sol added, Opus 5 and GPT-5.5 kept as the compared models; ECB rate of 25 September, DGX Spark at €5,083 before VAT (LDLC.pro), Artificial Analysis index v4.3.2; break-even thresholds recomputed.

Updated on September 2, 2026: DeepSeek API prices taken from the official price list of the day (input and output, off-peak and peak hours) instead of the "$0.21/M"; DGX Spark price aligned on the idealo.fr listing of 2 September 2026 (€4,875 excl. VAT); break-even recalculated with a visible formula. Same-day re-read: the MMLU-Pro, GPQA Diamond and AIME scores attributed to 0731, absent from the DeepSeek model card, are removed; the GPT-5.5 price is taken from the OpenAI price list ($5 input, $30 output per million) instead of $0.5 and $2; the "estimated" DGX Station throughputs, latencies and user counts are replaced by "to be measured"; local costs are given by their formula rather than a single figure.

Methodology

This benchmark compares the three best enterprise AI models available in August 2026: DeepSeek V4 Flash 0731 (284 billion parameters, 13 billion active per token, 1M context, open weights MIT licence), Claude Opus 5 and GPT-5.5.

Claude Opus 5 is the Anthropic frontier, with a 1M-token window at standard pricing for Claude 4.6 and later per the Anthropic pricing page, and GPT-5.5 the OpenAI frontier, at standard pricing up to 272K context per the OpenAI price list.

The benchmarks used are the official benchmarks published by each vendor, cross-referenced with data from Artificial Analysis Intelligence Index v4.3.2 (version re-read on 26 September 2026, independent methodology with 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, AA-LCR v1.1).

The cost per token is calculated as total cost of ownership (TCO) over 3 years for local deployment, and as public price for cloud APIs. No latency or throughput has been measured on the DGX Station: the section below says so rather than putting forward a number.

Official benchmarks comparison

Scores published by DeepSeek on the DeepSeek-V4-Flash-0731 model card. We do not reproduce scores for Claude Opus 5 and GPT-5.5: their vendors do not publish these same benchmarks, and estimating their values to declare a winner would be worthless. For an up-to-date third-party comparison across models, see Artificial Analysis.

Reading: DeepSeek V4 Flash 0731 is the best on agentic benchmarks (Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon), making it the preferred choice for code agents and software engineering tasks. Claude Opus 5 is better on general benchmarks (MMLU-Pro, GPQA, AIME, Humanity's Last Exam, AA-Omniscience), making it the preferred choice for general reasoning and precision. GPT-5.5 is competitive on general benchmarks but less performant on agentic tasks.

Cost per token

The local cost is derived from a visible formula with named, sourced assumptions, detailed below the table: hardware amortised excluding VAT over three years plus electricity, divided by the tokens actually produced. It counts neither integration, nor maintenance, nor training, which come on top. APIs are at list price.

DeepSeek prices: official price list read on 26 September 2026 (V4 Flash 0731 retired from the API on 10 September, prices of its successor V4.1 Flash), per million tokens; conversions at €1 = $1.1403 (ECB, 25 September 2026). Local costs were recalculated on 8 October 2026 with the method below. Anthropic and OpenAI prices: Anthropic pricing page and OpenAI price list, read on 26 September 2026.

How the local cost is calculated

Formula: cost per day = hardware price excluding VAT / 1,095 days (three years) + power × 24 h × price per kilowatt-hour; tokens per day = aggregate throughput × 86,400 s × load factor; cost per million = cost per day / (tokens per day / 1,000,000). The load factor is the share of time during which the machine produces tokens at that throughput, which no organisation knows before measuring its own usage: the result is therefore a range, not a single figure.

One line in full, DGX Spark at 30 tokens/s and 30% load: amortisation 7,000 / 1,095 = €6.39 a day; electricity 0.24 kW × 24 h × €0.1624/kWh = €0.94 a day; total €7.33 a day. Tokens: 30 × 86,400 × 0.30 = 777,600 a day, i.e. 0.778 million. Cost: 7.33 / 0.778 = €9.42 per million tokens. Same method for the cluster (12.79 + 1.87 = €14.66 a day) and for the DGX Station (136.986 + 6.236 = €143.22 a day).

Reading: the lowest cost per token remains the DeepSeek API, now served by V4.1 Flash, at $0.375/M off-peak ($0.75/M at peak hours), far ahead of Claude Opus 5.5 at $12/M, Claude Opus 5 at $15/M and GPT-5.5 at $17.50/M. The break-even of on-premises is read by dividing its daily cost by the API price: a DGX Spark (€7.33 a day) becomes cheaper than Claude Opus 5.5 (€10.52/M) from 0.70 million tokens a day, i.e. 27% load at 30 tokens/s, cheaper than Claude Opus 5 (€13.15/M) from 0.56 million, i.e. 21% load; it drops below GPT-5.5 (€15.35/M) from 0.48 million tokens a day, i.e. 18% load, and never matches the DeepSeek API (11.1 million tokens a day at peak hours at €0.66/M, 22.2 million off-peak) since its capacity tops out at 2.6 million a day at that throughput. A two-Spark cluster (€14.66 a day, 18.2 million tokens a day at most at 210.8 tokens/s) drops below Claude Opus 5.5 at 8% load (1.39 million tokens a day), below Claude Opus 5 at 6% (1.11 million) and below GPT-5.5 at 5% (0.95 million), but never matches the DeepSeek API, neither at peak hours (22.2 million tokens a day) nor off-peak (44.4 million), both beyond its capacity. A DGX Station (€143.22 a day) would have to sustain more than 2,500 aggregate tokens/s around the clock to drop below the DeepSeek API at peak hours (217 million tokens a day), which no published measurement establishes: [TO BE MEASURED]. On-premises is therefore not justified by price against the vendor's own API; it is justified against closed APIs from a few hundred thousand tokens a day, and by what no API sells: sovereignty, no imposed rate limit, no data leaving the premises, no deactivation risk.

Latency and performance

No DeepSeek V4 Flash 0731 / DGX Station GB300 pair is in our measurements section: the 2 August version published 68 tok/s, 2.6 s of prefill, 30 ms P50 and 1,908 concurrent users, figures that were neither measured nor derived from a visible formula, and which are removed.

Third-party API latencies are not published by their vendors: we prefer to write that than to put forward a number.

Reading: serving the model on-premises removes network latency and the vendor rate limit, which is a structural gain rather than a model property. We do not put a figure on the comparison with the APIs: their vendors do not publish these latencies, and they vary with region, load and contract tier. Availability is slightly lower than an API (99.7% against 99.9%), but it does not depend on a third party.

Sovereignty and compliance

Reading: DeepSeek V4 Flash 0731 in local is the only model with open weights, executable on-premises without a revocable licence, no vendor rate-limit, no data transfer outside the Union, and with fine-tuning possible on domain-specific data.

Claude Opus 5 and GPT-5.5 are closed APIs that transfer data to Anthropic and OpenAI data centers respectively, with rate-limit and vendor deactivation risk.

Hardware sovereignty: confidentiality and independence through the machine

Choosing local hardware is not only a question of performance or cost per token. It is a sovereignty decision: when the model runs on your own hardware, inside your own infrastructure, you keep full control over data confidentiality, you depend on no third-party vendor, and you are not exposed to rate-limit or deactivation policies.

Confidentiality: prompts, documents injected by RAG, inference logs and generated answers all stay on the company's internal infrastructure. No data crosses the network to a foreign data centre. GDPR compliance is demonstrated by construction rather than by an accumulation of contractual clauses.

Independence: an open-weight model running on company hardware does not fall under a revocable licence. No vendor can suspend access, no rate limit applies, no usage policy can be imposed unilaterally. On June 12, 2026, on a directive from the US Department of Commerce, Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 for all its customers, and only restored it on July 1, nineteen days later (Anthropic statement of June 12, 2026, restoration). The 2 August version wrongly dated this episode July 31. An open-weight model served on company hardware does not fall under that regime.

Full control: the choice of hardware lets you tune performance, concurrent users, latency, maximum context, without depending on a cloud provider's decisions. You can add memory, change runtime, enable or disable features, and optimise cost per token against your real usage, without a third party's permission.

On hardware sovereignty: local hardware brings full data confidentiality (no transfer outside the Union), complete independence from the vendor (open weights, no revocable licence, no rate limit), and full control over performance (hardware sizing, runtime, configuration). These advantages carry no price tag, but they are decisive for French and European organisations bound by GDPR, HDS or SecNumCloud obligations, or that want to avoid depending on a foreign cloud provider.

Conclusion: which model for which use case

The conclusion comes down to one rule per use case: DeepSeek V4 Flash 0731 for code agents and sensitive data, a closed API for general reasoning, and the DeepSeek API when budget matters more than sovereignty.

References

• DeepSeek API price list (read on 26 September 2026)
• DeepSeek V4 Flash 0731, official Hugging Face repository
• DeepSeek V4 technical report (arXiv 2606.19348)
• Artificial Analysis Intelligence Index v4.3.2 (re-read on 26 September 2026)
• Claude Opus 5 and Opus 5.5 pricing (Anthropic)
• OpenAI price list, GPT-5.5 and GPT-6 (read on 26 September 2026)
• Unsloth, DeepSeek V4 documentation
• QDNA sheet: DeepSeek V4
• QDNA sheet: vLLM
• QDNA sheet: DGX Station