DeepSeek Ships V4-Pro-0813, Open-Sources Agent Harness, Raises API Prices Up to 52%
DeepSeek released build V4-Pro-0813 with major agent benchmark gains, open-sourced its Deepseek Harness agent framework under MIT license, and announced API price increases of up to 52% effective August 16.
DeepSeek has moved its flagship model out of preview with build V4-Pro-0813, simultaneously open-sourcing its agent software and raising API prices by as much as 52%.
The deepseek-v4-pro endpoint now runs the new build. Model name, parameter count, and the 1-million-token context window are unchanged, and DeepSeek says existing integrations require no changes. The model is available in DeepSeek's app and web interface under "Expert Mode," and now natively supports OpenAI's Responses API with Codex integration. Reasoning effort can be set to "low," "high," or "max," with DeepSeek recommending "high" for everyday agent use.
According to DeepSeek's own benchmark comparison, Terminal Bench 2.1 scores rose from 72.1 to 87.9, and DeepSWE scores climbed from 12.8 to 62.7. DeepSeek claims the new build beats Claude Opus 4.8 on several agent benchmarks, though it only outperforms Kimi K3 and Fable 5 in select categories.
Third-party benchmarking firm Artificial Analysis confirms the improvement but places it in context: V4-Pro rises from 45 to 53 on the Intelligence Index, tying GLM-5.2. That score still trails Muse Spark (57), Qwen 3.8 Max (58), Kimi K3 (60), and Claude Opus 5, which leads at 63. DeepSeek has not yet published weights for the new build; the April preview remains the only version on Hugging Face.
The update also responds to competitive pressure from within DeepSeek's own lineup. The smaller V4 Flash model, updated in late July with build 0731, nearly matched the Pro Preview's Intelligence Index score at a fraction of the cost.
Agent harness goes open source
DeepSeek Harness v0.1 ships as a Developer Preview under the MIT license, positioned as an open alternative to OpenAI's Codex and Claude Code. It runs on DeepSeek's new Cordis plugin system, where tools, sandboxes, sessions, and the UI are all swappable components. A continuous session log records every prompt, tool call, and result, allowing runs to be resumed, branched, and replayed. The project is led by Cui Tianyi, who joined DeepSeek from quantitative trading firm Jane Street in March 2026. A beta tester call in early August drew 712 project sign-ups within three days.
Prices rise, cache hits hit hardest
New rates take effect August 16 at 4:00 p.m. UTC. DeepSeek introduced peak/off-peak pricing tied to Chinese business hours (1–4 a.m. and 6–10 a.m. UTC), with off-peak rates costing half as much.
Off-peak, V4-Pro input rises from $0.435 to $0.66 per million tokens, and output climbs from $0.87 to $1.98. During peak hours, both rates double to $1.32 and $3.96. Cache hits see the steepest jump — from $0.003625 to $0.022 off-peak and $0.044 at peak — shrinking the cache discount from roughly 1/120th to 1/30th of the standard input rate. This change hits agents that repeatedly re-read the same files hardest, and reverses much of the price cut DeepSeek issued in May.
The price increase comes as DeepSeek raises new capital and prepares for an initial public offering.
What this means
DeepSeek is charging more for a model that, by independent measurement, still ranks behind four competitors on general intelligence. The bigger story may be the harness: giving away MIT-licensed agent infrastructure lowers the barrier for developers to build on DeepSeek's ecosystem even as usage costs rise, a bet that infrastructure lock-in matters more than raw benchmark leadership heading into an IPO.
Related Articles
DeepSeek Releases V4-Flash-Vision-Exp, First Multimodal Model in V4 Family
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family, adding visual understanding to the V4-Flash architecture. The 305B-parameter model shows substantial gains on multimodal agent benchmarks while holding steady on text-only tasks.
Hugging Face Transformers v5.16.1 Adds Support for GLM-5.3-Flash, a 320B-Parameter Multimodal MoE Model
Hugging Face's transformers v5.16.1 release adds support for GLM-5.3-Flash, a 320B total-parameter (18B active) multimodal mixture-of-experts model. Zhipu AI claims it outperforms GLM-5.2 while approaching Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the cost.
DeepSeek Releases Experimental V4-Flash-Vision-Exp, Claims Near-Parity With Opus 4.8 on Agent Benchmarks
DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal extension of V4-Flash that adds image understanding while preserving text reasoning capabilities. The company claims the model approaches or beats Anthropic's Opus 4.8 on its internal multimodal agent benchmarks.
DeepSeek Releases V4 Flash Vision Exp, an Experimental Multimodal MoE Model with 1M Context
DeepSeek has released V4 Flash Vision Exp, an experimental vision-enabled variant of DeepSeek V4 Flash 0731 that adds image understanding while matching the base model's text performance. The sparse mixture-of-experts model uses 13B active parameters out of 284B total and supports a 1M token context window.
Comments
Loading...