DeepSeek V4 Pro 0813: New Pricing, Benchmarks, Harness v0.1

DeepSeek V4 Pro's general-availability (GA) build, "V4-Pro-0813," was released on August 12, 2026. Compared with the preview build released on April 24, 2026, it shows large reported improvements on benchmarks such as Terminal Bench 2.1 and DeepSWE. At the same time, API pricing changes take effect on August 16, 2026, with the per-token cost of output roughly 2.3x higher. Around the same period, DeepSeek also released an agent execution framework, "DeepSeek Harness v0.1," under the MIT license. This article walks through the spec changes, the benchmark shifts, what the pricing update actually changes, the new agent stack, and how realistic self-hosting is. Note up front: the benchmark figures cited here are vendor-published or analyst-aggregated numbers, and independent third-party reproduction has not been confirmed as of this writing.

V4-Pro-0813 shows large reported score improvements over the preview build across several agentic and coding benchmarks.

On the Artificial Analysis Intelligence Index, the top-ranked model, Claude Opus 5, scores 63, while V4-Pro-0813 scores 53 — a large improvement from the preview build, but still a gap behind the top model. As noted above, all figures here are vendor-published or analyst-aggregated numbers, and independent third-party reproduction has not been confirmed as of this writing. Any adoption decision should be based on testing against your own use case rather than these figures alone.

The pricing update taking effect on August 16, 2026 raises input, output, and cache-hit input pricing. Figures below are per million tokens, off-peak.

During peak hours (UTC 1:00-4:00 and 6:00-10:00), the off-peak prices above are doubled. The cache-hit discount is also being reduced: previously, cache-hit input was priced at roughly 1/100 of standard input, and after the update that discount narrows to roughly 1/30. For agentic or long-context workloads that were designed around heavy prompt caching, this narrower discount translates directly into a real cost increase — worth checking closely.

- Bites hardest: agent workloads that repeatedly reuse the same long context (system prompts, large tool definitions, etc.). Because the cache discount narrows from 1/100 to 1/30, costs that used to be kept low via cache hits grow substantially
- Bites hardest: workloads with heavy output volume (long-form generation, detailed code generation, long reasoning traces). Output pricing is roughly 2.3x higher, so cost scales up with output volume
- Bites hardest: usage concentrated in peak hours (UTC 1:00-4:00, 6:00-10:00), where the same workload costs 2x the off-peak price
- Doesn't bite much: one-off, short queries or lightweight usage with little cache reliance — the input price increase (~1.5x) is comparatively smaller than the output and cache-discount changes
- Doesn't bite much: batch jobs where request timing can be shifted away from peak hours

DeepSeek released a Developer Preview (v0.1) of its agent execution framework, DeepSeek Harness, under the MIT license. Harness is built on top of "Cordis," a plugin system DeepSeek developed in-house, and its notable characteristic is that components such as tools, sandboxes, session management, and the UI are all provided as swappable plugins. This design, rather than locking users into a specific tool implementation or sandbox environment, appears intended to let organizations swap individual components to match their own execution environments and security requirements. For example, teams that want to reuse an existing tool-calling implementation, or integrate with their own sandbox and authentication infrastructure, can in principle swap out individual plugins rather than rebuilding the whole harness — a design that could lower the cost of building an agent stack in-house. Given its Developer Preview status, it's worth keeping in mind that the API surface and plugin interfaces may still change.

Weights for the V4 series (Pro and Flash) are released as open weights under the MIT license on Hugging Face. However, the weights for the 0813 build that just reached GA are not yet released as of this writing — the only version currently available on Hugging Face is the preview build from April 2026. So the 0813 build itself cannot be self-hosted right now.

It's still worth noting the hardware scale involved once weights are released. The following figures are community-reported estimates, not official confirmed values from DeepSeek.

- V4-Flash: community reports suggest roughly 170-175GB total VRAM as a rough guide
- V4-Pro: community reports suggest roughly 862GB total as a rough guide
- Both figures vary with quantization and configuration and are community-based estimates, not confirmed values

A VRAM footprint of roughly 862GB for V4-Pro is well beyond what an individual or small organization can realistically self-host on their own. Outside of organizations with dedicated infrastructure, full-precision self-hosting isn't a practical option for most, and using the API remains the realistic path for most users. The lighter V4-Flash (roughly 170-175GB total VRAM) starts to look feasible on a multi-GPU setup, but that scale still sits well above typical personal use and is more realistically a small-organization option.

Neither the benchmark gains nor the pricing update is, on its own, a clean verdict of "good" or "bad." How much the change actually affects your costs depends heavily on how much your existing workflow relies on caching and how much output volume it generates. The following points are worth checking against your own usage before deciding.

- Check your current reliance on prompt caching: workloads with high cache-hit rates are hit hardest by the narrower discount (1/100 to 1/30). Review your monthly token breakdown by input, output, and cache-hit before the update lands
- Re-run cost estimates for output-heavy workflows: output pricing is roughly 2.3x higher, so long-form generation and code-generation-heavy use cases see a comparatively larger cost impact
- Consider whether peak-hour concentration can be avoided: batch jobs and retry-heavy processes that can be timed can shift to off-peak hours to reduce cost
- Don't take the benchmark numbers at face value: the figures cited here are vendor-published or analyst-aggregated, without confirmed independent reproduction — validate adoption decisions against your own use case
- Treat self-hosting as off the table for now: the 0813 build's weights aren't released, and even once available, V4-Pro's ~862GB VRAM estimate puts it out of reach for most. Plan around API usage for now
- Avoid simplistic model-vs-model rankings: the right choice depends on your use case, latency requirements, and existing operations, so weigh both the pricing change and the performance gains against your own usage pattern