DeepSeek’s August 13, 2026 announcement combined three moves: the general release of DeepSeek-V4-Pro-0813, a sharp redesign of V4 API pricing, and the open-sourcing of DeepSeek Harness (dsh) v0.1. The model release strengthens DeepSeek’s position in coding and agentic workloads, while the pricing change makes clear that its highest-capability inference is no longer being offered through a simple ultra-low flat-rate strategy.
The pricing change took effect on August 17
DeepSeek replaced flat pricing for V4 Pro and V4 Flash with peak and off-peak billing at 00:00 Beijing time on August 17, 2026, equivalent to 16:00 UTC on August 16. Off-peak prices are half of peak prices. The documented peak windows are 09:00–12:00 and 14:00–18:00 Beijing time on weekdays; other hours are off-peak. 
914
For V4 Pro, the new off-peak prices per million tokens are:
| Billing category |
Previous price |
New off-peak price |
Change |
| Cached input |
¥0.025 |
¥0.15 |
+500% |
| Uncached input |
¥3 |
¥4.50 |
+50% |
| Output |
¥6 |
¥13.50 |
+125% |
The figures above are reported in the announcement coverage and DeepSeek pricing summaries. 
11
16 Peak rates are twice the off-peak rates, putting V4 Pro at ¥0.30 per million cached-input tokens, ¥9 per million uncached-input tokens, and ¥27 per million output tokens during peak periods. 11
13
The impact is uneven. Cached input sees the largest percentage increase in the off-peak comparison, which matters for agents that repeatedly send the same project files, instructions, or conversation context. Output is also substantially more expensive: the dollar-denominated rate rises from $0.87 per million tokens before the change to $1.98 off-peak and $3.96 at peak. 
69
Across the V4 family, DeepSeek described increases ranging from 50% to 1,100%, depending on the model, token category, and time of use. 
2 That makes the new schedule more than a conventional price increase: developers must now account for both model choice and execution time when estimating API costs.
What V4-Pro-0813 adds
V4-Pro-0813 is the general-availability version of DeepSeek’s V4 Pro flagship model. DeepSeek’s published comparison table shows substantial gains over the V4 Pro preview on several agentic and coding benchmarks. For example, Terminal-Bench 2.1 increased from 72.1 to 87.9, while DeepSWE increased from 12.8 to 62.7. 
4
40
The 7.3 figure sometimes cited in comparisons is not the V4 Pro preview score: the model card lists 7.3 for V4 Flash Preview and 12.8 for V4 Pro Preview on DeepSWE. 40 Keeping those baselines separate is important because otherwise the size of the V4 Pro improvement is overstated.
DeepSeek’s model card presents V4-Pro-0813 as competitive with models including Claude Opus 4.8, Claude Fable 5, Kimi K3, and GLM-5.2 on selected tests. The table also shows trade-offs rather than a universal win: V4-Pro-0813 leads some comparisons but trails competitors on others, including certain knowledge, repository-generation, and agent evaluations. 
35
3640
These results should be treated as vendor-reported benchmark claims, not as independent proof that V4 Pro universally surpasses competing systems. Independent analysis included in the available reporting notes that third-party validation remains limited and that results can vary with the harness, prompting, tools, and evaluation setup. 
41
DeepSeek Harness makes the launch about agents, not only models
Alongside V4-Pro-0813, DeepSeek released DeepSeek Harness v0.1, an open-source agent framework under the MIT license. It was positioned as an alternative to agent products such as Anthropic’s Claude Code. 
18
22
The harness itself is free software. Users still pay for whichever model or API they connect to it, so the cost of running an agent depends on inference usage rather than on a separate Harness subscription. 
17
23
That pairing is strategically significant. DeepSeek is giving developers a workflow layer for building coding and automation agents while charging for the model calls that power those workflows. In practical terms, the relevant business metric becomes less about the price of one isolated prompt and more about the cost of completing a task across multiple steps.
What the announcement signals
The changes do not prove that low-cost AI tokens have ended across the industry. They do show that DeepSeek’s earlier flat-rate pricing for a higher-capability production model was not a permanent commitment.
Peak/off-peak billing also has an operational function. DeepSeek says the system is intended to allocate resources more reasonably and encourage users to schedule work according to usage patterns. 14 For developers, that creates an incentive to move batch jobs, evaluations, indexing, and other deferrable workloads into off-peak periods. Interactive workloads that run during peak hours will face the highest rates.
The broader shift is from token economics to task economics. A model that completes a coding or research task with fewer failed attempts, less orchestration, or less human review may still be economical even if its token price is higher. Conversely, a low token rate may not translate into a low total cost when an agent needs many retries or produces unreliable outputs.
DeepSeek’s August 13 package therefore looks less like a retreat from cost competition than a repositioning of it. V4-Pro-0813 supplies a stronger premium model, DeepSeek Harness supplies an open agent layer, and peak pricing monetizes scarce or time-sensitive inference. The evidence supports a move toward complete agentic workflows—but not a claim that DeepSeek has already replaced the model endpoint with a fully integrated task-delivery business.
