The model behind every deepseek-v4-flash API call changed today. DeepSeek released the official public beta of its V4-Flash API on July 31, 2026 under the build designation V4-Flash-0731, and the headline isn't a bigger model or a new architecture — it's a retrained one. The 0731 build scores higher than DeepSeek's own V4-Pro-Preview on all nine agent and coding benchmarks the company published, raising a pointed question for any developer currently choosing or pricing an agentic AI stack: when a budget model can beat its own flagship through retraining alone, what does that tell us about where AI capability improvements are actually coming from?

For developers already calling deepseek-v4-flash, the migration cost is zero — same endpoint, same API key, same model name. The upgrade is silent. What they get is a model DeepSeek says scores 82.7 on Terminal Bench 2.1 (compared to 72.1 for the V4-Pro-Preview and 61.8 for the Flash Preview) and 54.4 on DeepSWE, up from 7.3 for the preview — a 645% jump on a benchmark designed specifically to test real-world GitHub issue resolution. That gap between 7.3 and 54.4 was produced without changing a single parameter in the base model. Only the post-training was rerun.

What Actually Changed: Re-Post-Training, Not Architecture

V4-Flash-0731 keeps the same structure as the preview: 284 billion total parameters with 13 billion active per token, a one-million-token context window, and the MIT license that allows self-hosting. What DeepSeek ran was a new round of post-training — the phase of model development that shapes how a model uses its pre-trained knowledge, not what knowledge it holds.

Post-training for agent capability works differently from the post-training that taught the model to be polite or answer questions clearly. For coding agents, the primary mechanism is reinforcement learning with verifiable rewards: the model attempts multi-step coding tasks, code gets executed, tests pass or fail, and the model is updated to favor the sequences that led to passing tests. Code execution provides a binary ground-truth reward signal that's uniquely reliable for this kind of training. A model can learn to write working code agents in a way it can't learn to write better poetry, because the reward signal is unambiguous.

The 645% jump on DeepSWE is plausible for exactly this reason. DeepSWE (from Datacurve) tests 113 real-world software engineering tasks across 91 GitHub repositories — long-horizon agentic work where the model must plan across many steps, call tools, handle errors, and complete a task verifiably. The preview build likely used general-purpose post-training. The 0731 re-post-training targeted this domain specifically. When a verifiable reward signal is available and the model has the base knowledge to succeed at the task, RL-based post-training can produce gains that look qualitatively large because the capability was latent in the pre-training and post-training is what activates it.

The Benchmark Numbers: What DeepSeek Published

The performance figures DeepSeek published show a model that competes within a few points of Anthropic's Opus-4.8 — broadly considered one of the leading frontier models on agent tasks — on several benchmarks, at a fraction of the price, according to DeepSeek's official API changelog:

Terminal Bench 2.1 (complex command-line agentic work): 82.7 for V4-Flash-0731, compared to 72.1 for V4-Pro-Preview, 61.8 for Flash Preview, and 85.0 for Opus-4.8. Agents' Last Exam: 25.2 for V4-Flash-0731 versus 25.7 for Opus-4.8 — a gap of half a point. Toolathlon (verified): 70.3. Cybergym: 76.7. NL2Repo: 54.2. On DeepSWE and DSBench-FullStack (an internal full-stack development test), the 0731 build scored 54.4 and 68.7 respectively, compared to 7.3 and 37.0 for the preview.

Two caveats apply and neither is minor. First, the DeepSWE score of 54.4 is a vendor-reported figure using DeepSeek's own Harness framework, which the company acknowledged in the same changelog will be released "soon" — meaning third parties cannot yet replicate the test. That caveat matters because V4-Pro's independently measured DeepSWE performance tells a humbling story: on a contamination-free, independently run DeepSWE test from yage.ai, V4-Pro scored 8% pass@1, compared to 70% for GPT-5.5 and 54% for Claude Opus 4.7, despite posting 80.6% on SWE-bench Verified under a vendor-controlled harness. The verifier that SWE-bench uses accepts roughly 8.5% of incorrect solutions; DeepSWE's verifier runs at 0.3% false positives. V4-Flash-0731's 54.4 on DeepSWE is on the more credible benchmark — but it's still self-reported under an unreleased harness, and independent re-runs will be the real verdict. Second, DSBench-FullStack and DSBench-Hard are explicitly internal test sets: useful as directional signals, not as externally verifiable performance data.

Developer Tooling: Responses API and Codex Integration

The capability gains arrive alongside practical additions for developers building production agents. V4-Flash-0731 natively supports the Responses API format — the interface standard, popularized by OpenAI, that structures how AI models call tools and manage multi-step task sequences, per DeepSeek's official API changelog. Prior versions of V4-Flash required workarounds for Responses API workflows; the 0731 build supports it directly, with specific training to use that interface well.

The build is also specifically adapted for Codex, the coding agent workflow tool. For development teams running agent loops through Claude Code, OpenCode, or similar tools that talk to DeepSeek as a backend, the 0731 update means the model has been trained on the exact interface patterns those tools use.

One detail to flag: thinking mode is on by default in V4-Flash, and thinking tokens — the internal reasoning steps the model generates — are billed at the output token rate even though they don't appear in the final response. For agent tasks that trigger extended reasoning chains, this can materially increase effective cost per call above the headline $0.14 per million input tokens. Teams migrating from the preview should test effective cost per completed task, not just per token, before finalizing infrastructure decisions.

One operational note for teams using third-party API providers: as of the release announcement, only DeepSeek's own provider was serving the 0731 build. Teams routing through OpenRouter or other aggregators may still receive the April 2026 preview build until those providers update their routing.

Re-Post-Training as an Industry Signal

The V4-Flash-0731 result is not just a product announcement. It's a data point in a larger argument about where the real innovation in AI development is now happening.

For most of 2024 and 2025, the primary competitive axis in large language models was pre-training scale: how many parameters, how many training tokens, how much compute. That axis has become harder to differentiate on, in part because architectural efficiencies like DeepSeek's own Mixture-of-Experts design (which lets V4-Flash activate only 13 billion of its 284 billion parameters per token, making inference cost equivalent to a ~13B dense model) have decoupled parameter count from inference cost. The frontier labs now treat post-training as the primary lever for capability improvement on deployed models — Claude Code, OpenAI's Codex CLI, and agent-focused fine-tuning workflows are all post-training plays, not pre-training ones.

What V4-Flash-0731 demonstrates is that re-post-training — applying a new post-training cycle to an already-released model — can produce qualitatively different behavior in specific domains without any change to the model's size, architecture, or knowledge base. The pre-training fixed what V4-Flash knows. The re-post-training changed how it acts on that knowledge in agentic contexts. For developers choosing AI infrastructure, this means the model at any given API endpoint is not a fixed object — it can change substantially on a vendor's schedule, as it did today.

What the V4-Pro Official Release Means for This

DeepSeek signaled in the same announcement that the official release of V4-Pro will also follow "as soon as possible." If V4-Pro undergoes the same re-post-training refinement that V4-Flash did, the gap that currently exists between V4-Flash-0731 (which beat V4-Pro-Preview on all nine agent benchmarks) and a post-retrained V4-Pro-Official could be significant. V4-Pro has 1.6 trillion total parameters with 49 billion active per token — substantially more capacity than Flash's 13 billion active. If re-post-training activates a similar proportion of latent agentic capability in a model with more total capacity, V4-Pro's official release could represent a meaningful step forward from the preview that its current benchmark profile (including the 8% independent DeepSWE result) does not predict.

That's speculative — but it's the right question to hold while the community verifies the V4-Flash-0731 claims.

Price and the China Data Law

V4-Flash-0731 is priced at $0.14 per million input tokens and $0.28 per million output tokens, unchanged from the preview, per DeepSeek's published API pricing. V4-Pro remains at $0.435 per million input tokens and $0.87 per million output tokens — an order of magnitude more expensive per output token than Flash. For high-volume coding agent pipelines, the gap is significant: a pipeline consuming 100 million output tokens monthly costs roughly $28 on V4-Flash versus roughly $87 on V4-Pro. DeepSeek also announced upcoming peak/off-peak time-of-use pricing, with peak hours set at 9:00–12:00 and 14:00–18:00 Beijing Time at twice the standard rate, though the effective date for that change is still to be confirmed.

Before treating the price gap as the primary decision factor, one condition requires direct acknowledgment. DeepSeek is operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., a Chinese company. China's National Intelligence Law (2017), Article 7, requires all organizations to support, assist, and cooperate with national intelligence work — an obligation that applies regardless of DeepSeek's stated privacy policy, the physical location of its servers, or any contractual arrangements with users. China's Cybersecurity Law (2017) and Data Security Law (2021) add data localization and government-access requirements on top of that baseline.

When using DeepSeek's hosted API, prompts and query metadata transit to servers subject to these laws. Developers who self-host V4-Flash's MIT-licensed open weights on infrastructure outside China avoid the data-routing concern, though doing so requires meaningful engineering investment and sacrifices DeepSeek's operational support. For regulated industries — healthcare, finance, legal, government — the hosted API is not an appropriate tool under current regulatory guidance in most jurisdictions. For unregulated workloads on non-sensitive data where cost-per-task economics dominate, the decision calculus is different, and V4-Flash-0731's verified agent benchmark performance will be the relevant variable once independent re-runs confirm today's self-reported figures.

Frequently Asked Questions

What does it mean that the model behind deepseek-v4-flash changed today?

Any developer calling the deepseek-v4-flash endpoint had their production model updated to the V4-Flash-0731 build without any required action on their part. The model identifier, API base URL, pricing, and architecture are unchanged. What changed is the post-training — the behavioral training that shapes how the model acts, not what it knows. DeepSeek reports the new build scores substantially higher on agent and coding benchmarks than the preview did. Developers should test their specific workflows on the updated model, since behavioral post-training changes can improve some tasks and regress others in ways benchmarks don't fully predict.

What is re-post-training, and why can it change a model's capabilities so dramatically without changing its size?

Post-training is the phase of model development that happens after pre-training: it shapes how a model behaves by training it on examples and rewards, without changing the underlying knowledge stored in its weights. Re-post-training applies a new post-training cycle to an already-deployed model. For agent tasks specifically, the mechanism is reinforcement learning with verifiable rewards — the model attempts tasks like fixing GitHub issues, code gets executed, tests pass or fail, and the model learns from the binary outcome. Because code execution provides a clear right-or-wrong signal, RL on this domain can produce large capability gains in a single training run. The 645% jump from 7.3 to 54.4 on DeepSWE suggests the preview build's agent capability was not limited by the model's knowledge or architecture, but by its procedural training on how to act in multi-step agentic contexts.

How reliable are the V4-Flash-0731 benchmark scores, and should developers trust the DeepSWE figure of 54.4?

With specific caveats. DeepSeek's own benchmark harness has not yet been released, so the scores cannot currently be independently replicated. The DeepSWE benchmark itself is contamination-free and uses a tight verifier (0.3% false positive rate), making it more credible than vendor-controlled evaluations — but the 54.4 figure is still self-reported under an unreleased framework. Context for calibration: V4-Pro's self-reported SWE-bench Verified score was 80.6%, while its independently measured DeepSWE score from yage.ai was 8% pass@1, a gap attributable largely to the difference in verifier quality. V4-Flash-0731's 54.4 on DeepSWE uses the tighter benchmark, which is genuinely more meaningful — but independent confirmation from yage.ai or a comparable lab running the same tasks under their own harness should be the deciding signal for production adoption on high-stakes workloads.

Is the DeepSeek API safe to use for enterprise or sensitive data workloads?

For regulated or sensitive data, significant limitations apply. When using DeepSeek's hosted API, data transits to servers in China and becomes subject to Chinese national law. China's National Intelligence Law (2017), Article 7, legally requires all organizations to cooperate with government intelligence work on request — with no exceptions and no court order required. DeepSeek's privacy policy confirms that personal data is stored in the People's Republic of China. These are fixed legal conditions, not variables to weigh against price. Enterprises handling protected health information, financial data, client communications, or government-adjacent work should treat the hosted API as incompatible with their data governance obligations. Self-hosting the MIT-licensed V4-Flash weights on infrastructure outside China removes the data-routing concern but requires meaningful engineering investment and still leaves open questions about behavioral constraints in the model's weights.