Deepseek ships improved V4 Pro, open-sources its agent software, and raises API prices
Deepseek released an updated V4-Pro model (buildV4-Pro-0813) with significant benchmark improvements, including Terminal Bench 2.1 scores jumping from 72.1 to 87.9 and DeepSWE scores rising from 12.8 to 62.7 Deepseek Harness v0.1 is now available as open-source agent software under the MIT license, built on the Cordis plugin system with swappable components for tools, sandboxes, sessions, and UI API pricing is increasing with a new peak/off-peak structure aligned to Chinese business hours, with
Analysis
TL;DR
- Deepseek released an updated V4-Pro model (buildV4-Pro-0813) with significant benchmark improvements, including Terminal Bench 2.1 scores jumping from 72.1 to 87.9 and DeepSWE scores rising from 12.8 to 62.7
- Deepseek Harness v0.1 is now available as open-source agent software under the MIT license, built on the Cordis plugin system with swappable components for tools, sandboxes, sessions, and UI
- API pricing is increasing with a new peak/off-peak structure aligned to Chinese business hours, with cache hit costs rising dramatically from $0.003625 to $0.022-$0.044 per million tokens
- The updated V4-Pro now supports the OpenAI Responses API with Codex integration and offers three reasoning effort levels (low, high, max)
- Despite improvements, V4-Pro still trails leading models like Claude Opus 5 (63 vs 53 on Artificial Analysis Intelligence Index) and Kimi K3 (60)
Why It Matters
Deepseek's moves signal the intensifying competition in the open-weight and agent infrastructure space, as the company responds to its own V4 Flash model eroding the flagship's value proposition. The pricing strategy—particularly the steep cache hit increases—reflects Deepseek's shift from aggressive market penetration toward profitability ahead of its anticipated IPO, directly impacting cost structures for AI practitioners building agent systems.
Technical Details
- V4-Pro Update (buildV4-Pro-0813): Maintains the same parameter count and one-million-token context window as the previous version. Adds native OpenAI Responses API support with Codex integration, three-tier reasoning effort settings (low/high/max), and is accessible via "Expert Mode" in the app and web interface. Existing integrations remain compatible without modifications.
- Deepseek Harness v0.1: Open-source agent framework released under MIT license as a Developer Preview. Built on the Cordis plugin system where all components—tools, sandboxes, sessions, and UI—are swappable plugins. Features continuous session logging tracking every prompt, tool call, and result, with support for resuming, branching, and replaying runs. Launches via npx through a local web interface. Project led by Cui Tianyi, formerly of Jane Street.
- Benchmark Performance: Terminal Bench 2.1 improved from 72.1 to 87.9; DeepSWE from 12.8 to 62.7. Beats Claude Opus 4.8 on several agent benchmarks but ranks below Muse Spark (57), Qwen 3.8 Max (58), Kimi K3 (60), and Claude Opus 5 (63) on the Artificial Analysis Intelligence Index. Weights for the new build have not been published.
- Pricing Structure: New rates effective August 16 at 4:00 p.m. UTC. Off-peak hours (outside Chinese business hours: 1-4 a.m. and 6-10 a.m. UTC) cost half of peak rates. V4-Pro input rises from $0.435 to $0.66 (off-peak) and $1.32 (peak) per million tokens; output from $0.87 to $1.98 and $3.96 respectively. Cache hits increase from $0.003625 to $0.022 (off-peak) and $0.044 (peak), reducing the cache discount from ~1/120 to ~1/30 of regular input price.
Industry Insight
- Agent infrastructure is becoming a key differentiator: Deepseek's open-sourcing of Harness positions it to capture developer mindshare in the agentic AI space, competing directly with OpenAI's Codex and Claude's agent capabilities. The modular Cordis plugin architecture could accelerate adoption by allowing teams to customize agent behavior without fork maintenance.
- Cache pricing changes will reshape agent economics: The dramatic increase in cache hit costs disproportionately impacts agents that repeatedly read the same files or documents. Teams building retrieval-augmented or codebase-aware agents should reassess their token budgets and consider alternative caching strategies or off-peak scheduling to mitigate costs.
- Deepseek's pricing pivot signals market maturation: The move from aggressive discounting to peak/off-peak pricing with substantial increases—partially undoing the May price cuts—coincides with IPO preparations and capital raising. This suggests the "burn cash for market share" phase is ending for frontier model providers, and practitioners should expect continued price normalization across the industry.
Disclaimer: The above content is generated by AI and is for reference only.