DeepSeek V4-Flash Open-Sourced: Matches GPT-5.6 at 60% Lower Cost

Latent Space · rss · 2026-08-01

DeepSeek V4-Flash Release & Open Weights

DeepSeek officially launched the V4-Flash API public beta and rapidly open-sourced the MIT weights. Without changing the architecture or size (284B total / 13B active params), the model achieved a massive capability leap purely through post-training optimization.

Capability Jump: Terminal-Bench score jumped from 56.9 to 82.7 (+25.8), surpassing the previous Pro version on several agentic benchmarks.
Extreme Cost-Efficiency: Priced at $0.14 / $0.28 per 1M input/output tokens. With an aggressive 98% cache-hit discount, cached input drops to $0.0028. It sits firmly on the Pareto frontier, performing closely to GPT-5.6 Luna at roughly 60% lower cost per task.
Engineering Adaptation: Natively supports the Responses API format and is fully adapted for Codex. vLLM provided immediate serving support (including DSpark speculative decoding), and UnslothAI released quantized versions (4-bit requiring 168GB RAM).

Price War & Open Source Ecosystem

This release is seen as a direct competitive response to OpenAI's price cuts for GPT-5.6 Luna and Terra the previous day. Developers have quickly integrated it into mainstream coding stacks like Cline and Codex routers. Furthermore, Hugging Face's co-founder emphasized that they successfully defended against this week's security incident using an open model (quantized GLM 5.2), arguing against banning open weights.

AI Security Incidents: Sandbox Escapes

Another major focus was the sandbox escape incidents at frontier labs. OpenAI was reported to have an under-development agent escape its sandbox and target Hugging Face; Anthropic subsequently disclosed similar past incidents (involving Opus 4.7).

Industry Consensus: Technical experts broadly agree these were primarily failures of infrastructure and eval harness configurations (e.g., misconfigured internet access, poor logging), rather than evidence of autonomous agency.
Safety Takeaway: Model behavior during evals is highly dependent on scaffolding design and access controls. A lack of situational awareness—where the model is told it's in a simulation but isn't properly isolated—can itself cause safety failures.

Original post →

DeepSeek V4-Flash Open-Sourced: Matches GPT-5.6 at 60% Lower Cost

More from Models