DeepSeek V4 Pro 0813: Benchmarks & Verdict

• 0813 is the third DeepSeek V4 Pro release in four months: an April preview, a July GA with a full benchmark table, and this August production build, which DeepSeek itself frames as a serving-efficiency update rather than a new architecture.
• Same core spec as GA: 1.6 trillion total parameters, 49 billion active per token (Mixture-of-Experts), 1-million-token context, 384K max output, MIT licence, open weights on Hugging Face.
• A large vendor-vs-neutral benchmark gap: DeepSeek's own harness reports 87.9 on Terminal-Bench 2.1; CoderSera's neutral-harness re-test found 54.68%, a 33-point drop, with V4 Flash outscoring Pro (67.04%) under the same neutral conditions.
• Honesty scores near the floor: Artificial Analysis's AA-Omniscience factual-reliability metric puts the 0813 build at 0.83 against Claude Opus 5's 37.07, continuing a pattern flagged in DeepSeek V4's earlier GA evaluation.
• NIST CAISI's safety findings still apply, with a caveat: a 94% jailbreak-compliance rate on the most restrictive configuration versus 8% for US reference models, and an eight-month capability gap versus the frontier, both measured on the April preview, not the 0813 build specifically.
• New peak/off-peak pricing lands the same week: off-peak rates are unchanged from GA ($0.435/$0.87 per million input/output tokens), but from 17 August 2026 peak-hour multipliers reach up to 12x on cached input.
• A genuinely useful side comparison: Alibaba's newly released, far smaller Qwen3.8-27B (27.78B parameters, Apache 2.0) held its own against DeepSeek V4 Flash in one creator's local, self-hosted coding test.

DeepSeek V4 first appeared as an open preview on 24 April 2026, with both a Pro and Flash tier released as MIT-licensed open weights alongside API access from day one. On 19 July 2026, the Pro tier reached General Availability with a full published benchmark table, and DeepSeek explicitly said the flagship's official release "will follow soon" after the Flash tier's own GA graduation on 31 July. That promise came through on 13 August 2026 with the "0813" build, the version now shipping as deepseek-v4-pro across the API, no model-name change required for existing integrations.

What 0813 actually changed is narrower than a full version bump might suggest. According to CoderSera's technical breakdown, the update added "four new DSpark keys and about 51.7 billion additional parameters," described as focused on speculative decoding for serving efficiency rather than a fundamental change to the model's reasoning architecture. The core spec, 1.6 trillion total parameters, 49 billion active, a 1,048,576-token context window and 384K maximum output, carried over unchanged from July's GA release. What did change was performance under DeepSeek's own harness: the vendor reports DeepSWE jumping from 12.8 to 62.7 versus the April preview, CyberGym from 52.7 to 83.3, and Terminal-Bench 2.1 from 72.1 to 87.9.

The release also arrived into a busier field than July's GA did. Alibaba shipped the small, self-hostable Qwen3.8-27B on 14 August, a day after 0813, and Z.AI's GLM 5.3 landed the same week, prompting several tracked creators to frame the period as a fresh wave of "is this the new #1 open source model" content rather than a story about any single release. See our GA review from July for the fuller architecture and lineage background this article builds on.

See our DeepSeek V4 Pro tool listing for a quick-reference summary alongside other tracked models. DeepSeek V4 Pro is a sparse Mixture-of-Experts model: 1.6 trillion total parameters, with roughly 49 billion active on any given token. The architecture combines two mechanisms DeepSeek calls Manifold-Constrained Hyper-Connections (mHC), which the company says constrains signal amplification to under 2x to keep training of a model this large stable at only 6.7% computational overhead, and DeepSeek Sparse Attention (DSA), a hybrid of compressed and hierarchical attention designed to make the 1-million-token context window computationally viable by pushing cost growth closer to linear than quadratic. DeepSeek has previously said this combination cuts single-token inference compute to roughly 27% and KV cache to around 10% of what its earlier V3.2 generation needed at the million-token setting.

The 0813 update sits on top of that same base architecture rather than replacing it. CoderSera's technical notes describe the change as adding four new "DSpark" keys, roughly 51.7 billion additional parameters oriented toward speculative decoding, a technique that predicts several tokens ahead and verifies them in parallel to speed up generation, rather than any change to how the model reasons. The full weight set now runs to approximately 892.7GB at FP8 precision, which rules out realistic single-workstation self-hosting; DeepSeek V4 Flash, at 166.9GB (284 billion total parameters, 13 billion active), is the tier anyone outside a multi-GPU server setup would actually consider running locally.

Concurrency limits differ meaningfully between tiers too: the default account limit is 500 concurrent requests for V4 Pro versus 2,500 for V4 Flash, a detail that matters more for production deployments than most benchmark charts. Modality remains text-only for both tiers; DeepSeek has not added image, audio or video input in this release, a gap several rival Chinese labs (Alibaba's Qwen 3.8 Max among them) have already closed.

Reasoning-effort control and agent tooling

0813 ships with three explicit reasoning-effort settings, low, high and max, letting a caller trade latency for depth on a per-request basis, alongside native support for the OpenAI Responses API and an Anthropic-compatible API surface, so existing coding-agent tooling built for either ecosystem can generally point at DeepSeek with a base-URL and model-ID change. DeepSeek also released DeepSeek Harness on 13 August, the same day as the 0813 build, an MIT-licensed, plugin-first agent framework positioned as an open alternative to Claude Code, and several of this week's creator videos test V4 Pro specifically through that harness rather than the bare API.

Coding and frontend generation

Hands-on testing from MindStudio and WorldofAI both single out frontend and one-shot UI generation as the model's clearest strength, describing output as "visually competitive with frontier-tier models" on tasks like Three.js scenes, physics demos and one-shot app clones. WorldofAI's video specifically runs V4 Pro through a MacOS clone build, an engine simulator, a catapult physics demo, a flight simulator and a Call of Duty Zombies-style clone as informal stress tests of agentic, multi-step coding rather than single-function benchmarks.

Task planning and clarifying questions

MindStudio's review notes the model asks clarifying questions when a prompt is ambiguous rather than guessing, and is comparatively strong at decomposing a larger job into sequenced steps, both genuinely useful agentic traits. The same review flags the flip side directly: a tendency to overthink straightforward problems and to overengineer solutions, restructuring files or adding abstractions a simpler prompt did not call for, sometimes producing a worse result than a more direct approach would have.

This is where the 0813 release needs the most scrutiny. DeepSeek's own agent-harness numbers show large jumps over the April preview: Terminal-Bench 2.1 from 72.1 to 87.9, CyberGym from 52.7 to 83.3 (a table lead, ahead of Claude Fable 5), DeepSWE from 12.8 to 62.7, and Automation Bench at 31.8, also a table-topping score. Artificial Analysis's own Intelligence Index, which aggregates nine separate benchmarks including GPQA Diamond, SciCode and Humanity's Last Exam, scores the 0813 build at 53, ranked third among the roughly 106 models it tracks and well above the field median of 27.

The Terminal-Bench gap is the single most important number in this table. CoderSera's independent re-test, run on a neutral harness rather than DeepSeek's own agent tooling, found V4 Pro scoring 54.68%, a 33-point drop from the vendor's 87.9. More strikingly, V4 Flash, the smaller and cheaper tier, scored 67.04% on the identical neutral harness, actually beating Pro. That inversion is a strong signal that at least some of Pro's vendor-reported gain comes from harness-specific tuning rather than a genuine capability improvement over its own smaller sibling. SWE-bench Verified tells a more encouraging story in CoderSera's own testing (96.40%, second only to Claude Opus 5's 97.00%, at a fraction of the per-test cost), so the picture is not uniformly negative, but it is uneven enough that a single headline number should not be trusted without checking which harness produced it.

None of DeepSeek's own 0813 harness numbers had been independently reproduced by LMArena at the time of writing, and NIST CAISI's government evaluation, discussed in full below, covers the April preview build rather than 0813, so whether the gaps it found have closed is genuinely unknown rather than resolved either way.

DeepSeek has not published a Responsible Scaling Policy, ASL-style capability tiering, or a dedicated pre-deployment safety evaluation comparable to what Anthropic, OpenAI or Google DeepMind publish alongside frontier releases. Its public model-card language is comparatively generic: it states that users are responsible for ensuring safe integration, including implementing guardrails and other safety mechanisms, prior to deployment, and that foundation and fine-tuned models require additional use-case-specific testing before production use. That is a real disclosure, but it is a materially thinner safety document than the frontier US labs now publish, and it has not changed meaningfully between the April preview, July GA and August 0813 releases.

The most concrete independent safety data on this model line comes from NIST's Center for AI Standards and Innovation (CAISI), which ran a nine-benchmark, five-domain government evaluation of DeepSeek V4 Pro in April 2026. CAISI found the model's most restrictive configuration complied with 94% of malicious jailbreak requests in its test set, against 8% for the US reference frontier models it compared against, a large gap for any enterprise deployment where content policy actually matters. On capability, CAISI found V4 Pro performed closest to GPT-5, a model released roughly eight months earlier, rather than to the frontier models DeepSeek's own self-reported tables reference, with the largest specific gaps in cybersecurity (32% on CAISI's CTF-Archive-Diamond benchmark, versus 71% for GPT-5.5) and abstract reasoning; maths performance was close to parity with frontier models.

The important caveat, stated plainly: CAISI's evaluation covers the April preview build, not the August 0813 release this article is about. DeepSeek reports benchmark gains of up to roughly 50 percentage points on some agentic tasks between preview and 0813, and it is plausible some of that also reflects improved robustness against jailbreak-style prompts, but no independent evaluator has published an updated CAISI-equivalent test of 0813 at the time of writing. Readers should treat the jailbreak-compliance and capability-gap figures as the most recent independently-measured data point on this model family, not as a confirmed description of the current release.

DeepSeek has not published dedicated autonomous-replication, self-exfiltration or agentic-misuse evaluations for V4 Pro of the kind Anthropic and OpenAI now run and publish for their own frontier agent models, and none of the third-party sources reviewed for this article report running such tests either. That is a genuine information gap rather than a clean bill of health: the model now ships with 0813's reasoning-effort controls, native Responses API and Anthropic-compatible tooling, and the same-day launch of the open-source DeepSeek Harness agent framework, all of which make it materially easier to wire V4 Pro into autonomous, multi-step agent loops than the April preview was. CAISI's own evaluation domains did include an autonomy-adjacent category as part of its five-domain test, alongside cybersecurity, but its detailed autonomy findings were not broken out with the same specificity as the cyber and reasoning gaps; readers running V4 Pro inside an agentic harness with real-world write access (file systems, payments, outbound network calls) should apply the same sandboxing and human-approval gates they would for any model without a published agentic-safety evaluation, rather than assuming the absence of a bad headline means the absence of risk.

This remains the release's clearest weak point, and it has not visibly improved since July's GA. Artificial Analysis's AA-Omniscience metric, which specifically measures whether a model declines or hedges on questions it should not confidently answer versus confabulating a plausible-sounding response, scores the 0813 build at 0.83, near the bottom of the scale Artificial Analysis tracks, against Claude Opus 5's 37.07 on the same metric. That is not a small gap; it is close to a floor score on a test specifically designed to catch models that answer instead of admitting uncertainty. The earlier GA-era evaluation of this model line found V4 Pro and V4 Flash answering rather than declining or hedging in 94% and 96% of cases respectively, among the highest rates Artificial Analysis had recorded across its comparison set at the time, and the 0813 AA-Omniscience score suggests that underlying pattern has persisted into the current build rather than being specifically addressed by it.

Practically, this means confident-sounding answers from V4 Pro on open-ended factual questions, especially ones outside its strongest domains of coding and maths, deserve more scepticism than the same answer from a model with a stronger calibration record. For factual research or anything where a wrong-but-confident answer carries real cost, pairing V4 Pro with a verification step or a second model is the safer pattern several reviewers, including this site's own earlier GA coverage, have converged on.

MindStudio's independent eight-question hands-on test scored V4 Pro at 76.25% (61/80), including a perfect 10/10 on a hard maths problem and a long-horizon agentic task, and the highest score any model in its comparison achieved (7/10) on a 3D wristwatch modelling challenge, genuinely strong results that sit between the vendor's 87.9 and the neutral-harness 54.68. The same review's clearest practical finding is that V4 Pro is not a strict upgrade over its own smaller sibling for everyday work: "for simple, everyday tasks, V4 Flash often performs better because it doesn't overthink and rarely makes mistakes from excessive reasoning," a direct echo of the overengineering tendency noted in the capabilities section above.

Creator commentary points the same direction. Julian Goldie's coverage frames the release around independent testing catching up with self-reported claims, specifically contrasting DeepSeek's 87.9 Terminal-Bench figure against an "87.9 vs 88.0" framing versus Claude, and flags cache-hit pricing, a lack of vision input, and self-hosting cost as the three practical catches worth knowing before switching a real workflow over, his own stated view rather than an independently verified claim, but a reasonable summary of where this review lands too: genuinely competitive on cost and several agentic benchmarks, not a settled "beats the frontier" story once anyone outside DeepSeek runs the numbers.

Those off-peak rates carried over unchanged from July's GA release. What changed on 17 August 2026 is the peak/off-peak structure layered on top. Rates now roughly double during Beijing business hours, 09:00-12:00 and 14:00-18:00 daily, but the increase is uneven across token types rather than a flat multiplier: cache-hit input rises the most steeply, 6x off-peak and up to 12x at peak versus the previous rate (DeepSeek's own RMB pricing shows peak cache-hit climbing from ¥0.025 to ¥0.30 per million tokens), cache-miss input rises 1.5x-3x, and output rises 2.25x-4.5x. V4 Flash sees smaller multipliers throughout, 2.5x-5x on cache-hit input. In blended terms across a typical workload, that works out to roughly a 1.96x input and 2.94x output increase at peak versus the pre-17-August baseline.

Even after the increase, V4 Pro remains meaningfully cheaper per token than Claude Opus 5 or GPT-5.6 Sol at off-peak hours, and CoderSera's SWE-bench testing put its cost per test at roughly 1/36th of Grok 4.6's and 1/59th of Claude Opus 5's. That said, the specific "148x cheaper" figure WorldofAI cites in its coverage is a creator-reported comparison from a specific benchmark scenario on WorldofAI's own benchmarking tool, not an official DeepSeek claim or an independently verified blended rate, and readers should model their own workload's cache-hit ratio and peak/off-peak mix before assuming any single multiple applies.

• A large vendor-vs-neutral benchmark gap on Terminal-Bench 2.1: 87.9 on DeepSeek's own harness versus 54.68% independently, with V4 Flash outscoring Pro under neutral conditions.
• Weak honesty and calibration: AA-Omniscience score of 0.83, near the bottom of the scale, meaning the model very rarely declines or hedges on questions it should.
• Safety data lags the current release: NIST CAISI's 94% jailbreak-compliance and eight-month capability-gap findings are the most recent independent safety numbers available, but they cover the April preview, not 0813.
• No published agentic-safety evaluation: unlike frontier US labs, DeepSeek has not released autonomous-replication or agent-misuse testing for this model, despite shipping a same-day open-source agent harness that makes autonomous deployment easier.
• Text-only: no image, audio or video input in either V4 tier, a gap rival releases like Qwen 3.8 Max have already closed.
• 892.7GB flagship weights rule out realistic single-workstation self-hosting for the Pro tier; only V4 Flash is a practical local option, and even then requires substantial hardware.
• New peak-hour pricing adds planning overhead: cache-hit input can cost up to 12x more during Beijing business hours, so cost forecasting now depends on when, not just how much, an application calls the API.

Against DeepSeek V4 Flash, the more interesting comparison this week is not V4 Pro at all, it's Flash against Alibaba's newly released Qwen3.8-27B, a genuinely different class of model. Qwen3.8-27B is a 27.78-billion-parameter dense model (64 transformer blocks, a hybrid layout of 48 Gated DeltaNet layers and 16 full-attention layers) that Alibaba released on Hugging Face under Apache 2.0 on 14 August 2026, small enough to run on a single high-end consumer or prosumer GPU at 4-bit precision, a dramatically different deployment profile from V4 Flash's 284 billion total parameters, let alone V4 Pro's 1.6 trillion.

In that video, creator Bart Slodyczka ran both models locally on the same Mac Studio against three coding tasks of increasing difficulty: a weather dashboard pulling live data, a tower-defence game, and an Excel-style spreadsheet with working formulas. His result, published alongside a GitHub repo of the full test setup and outputs: the far smaller Qwen3.8-27B held its own throughout, and on the hardest task, the spreadsheet with live formulas, it was the only one of the two models that produced something that actually worked. That is one creator's own local test on his own hardware, not a controlled independent lab benchmark, and it should be read as exactly that, but it is a striking result given the roughly tenfold parameter gap between the two models.

Qwen's own published benchmark deltas support the direction of that result, even if the absolute numbers come from a different comparison. Qwen3.8-27B posts a +28.9 percentage-point gain on DeepSWE 1.1 and +20.4 on OSWorld-Verified over its own 3.6 predecessor, genuinely large agentic-coding improvements for a model this size, alongside a published SWE-bench Pro score of 61.7% and a Terminal-Bench 2.1 score of 73%, both figures that would have looked implausible for a 27-billion-parameter dense model a year earlier. As with every vendor-published chart in this article, these are Qwen's own numbers, not independently reproduced ones, and the chart itself carries that caveat directly on the image.

Against Western closed frontier models, V4 Pro's own numbers put it competitive on SWE-bench Verified but clearly behind on the neutral Terminal-Bench re-test, and its 0.83 AA-Omniscience score is well behind Claude Opus 5's 37.07 on honesty specifically. See our DeepSeek V4 Pro GA review for the fuller July-era comparison against Claude Fable 5, and our Qwen 3.8 Max review and Kimi K3 review for how the rest of the Chinese open-weight field is positioned this month.

Worth trying now: teams already comfortable weighing vendor benchmarks against independent numbers, who want a genuinely cheap, MIT-licensed hosted API for coding and agentic workloads and can schedule heavier usage into off-peak hours to avoid the new peak multipliers. V4 Pro's frontend-generation and multi-step task-planning strength, backed by MindStudio's and WorldofAI's hands-on tests, makes it a reasonable second model to route specific coding tasks to even if it is not your primary agent.

Better to wait, or use something else: anyone who needs the model to reliably say "I don't know" rather than confidently guess, given the 0.83 AA-Omniscience score; anyone deploying an autonomous agent with real-world write access, given the absence of a published agentic-safety evaluation for this specific release; anyone whose workload leans on Terminal-Bench-style agentic coding specifically, where the neutral-harness score (54.68%) is well below both the vendor figure and V4 Flash's own neutral score; and anyone who wants to self-host the full Pro tier rather than the far smaller, genuinely local-friendly Qwen3.8-27B or DeepSeek V4 Flash.

DeepSeek V4 Pro's 0813 build is a real, if incremental, production update, not the ground-up leap some of this week's video titles imply. The core architecture, pricing tier structure and open-weight licence are unchanged from July's GA; what changed is a serving-efficiency addition, several large vendor-reported benchmark gains, and a new peak/off-peak pricing split that partially offsets DeepSeek's famous cost advantage during business hours. The single most important fact for anyone evaluating this release is the size of the gap between DeepSeek's own Terminal-Bench figure (87.9) and CoderSera's neutral-harness result (54.68%), a 33-point difference that should temper how much weight any of DeepSeek's other self-reported 0813 numbers are given until someone outside the company reproduces them.

The more durable story this week may again not be about DeepSeek Pro specifically. Alibaba's Qwen3.8-27B, released a day after 0813 at roughly 1.7% of V4 Pro's parameter count, is already holding its own against DeepSeek's own mid-tier model in at least one independent local test, a reminder that the race to shrink capable models for local deployment is moving just as fast as the race to scale up flagship ones.