DeepSeek V4-Flash 0731 Update Jumps Terminal-Bench Score by 25.8 Points With No Architecture Change

DeepSeek Ships Post-Training Update, Not a New Model

DeepSeek released V4-Flash 0731 on July 31, 2026, a post-training-only update to its Flash API and open-weights model. The company reports the update improves agent capability enough to surpass its own V4-Pro-Preview, without any change to architecture or parameter count, according to DeepSeek.

The model remains 284B total parameters with 13B active per token, matching the prior V4 Flash release. Artificial Analysis confirmed the update runs a 256-expert routed architecture with 6 active experts per token, a 1M-token context window, and three selectable reasoning-effort levels. An included speculative decoding module, DSpark, can reportedly be toggled with a single serving flag, according to vLLM Project.

Benchmark Gains

The scale of improvement from post-training alone is the notable technical claim here. According to figures cited by @cline, Terminal-Bench score rose from 56.9 to 82.7 — a 25.8-point jump. Artificial Analysis separately reported Terminal-Bench 2.1 reaching 79%, GDPval-AA v2 Elo climbing from 1189 to 1559, a τ³-Bench Banking gain of 8 points, and a 12% reduction in output-token usage versus the prior version. On the Frontend Code Arena, DeepSeek-V4-Flash-High scored 1586, up 154 points from its preview version, placing it on that benchmark's Pareto frontier, according to @arena.

On Artificial Analysis's composite intelligence index, V4-Flash 0731 moved from 40 to 50, landing one point behind GPT-5.6 Luna's score of 51. Artificial Analysis initially reported an anomalous cache-hit-rate figure, which the firm corrected shortly after publication, reaffirming the model sits on its Pareto frontier for intelligence versus cost per task when run through DeepSeek's first-party API.

Pricing and Open-Weights Release

DeepSeek priced the API at $0.14 per 1M input tokens and $0.28 per 1M output tokens, with a cache-hit discount bringing cached-token cost down to $0.0028 per 1M tokens — a roughly 98% discount claimed by DeepSeek. The API now supports the Responses API format and is described by DeepSeek as "fully adapted for Codex."

Official weights were published on Hugging Face under an MIT license almost immediately after the API launch. Unsloth AI released runnable quantizations requiring approximately 168GB RAM for lossless 4-bit inference and 110GB for 3-bit, with additional UD quants shared by Daniel Han shortly after.

DeepSeek clarified that gains apply only to the Flash tier; V4-Pro API, App, and Web versions remain unchanged, with an official V4-Pro update still pending.

Context: Price War and Security Disclosures

The release landed one day after OpenAI cut pricing on GPT-5.6 Luna by 80% and on Terra by 20%, prompting several observers, including @kimmonismus, to frame DeepSeek's update as a direct competitive response. Developers rapidly integrated the model into existing tooling — Cline made it free to use, Teknium added it to Hermes Agent, and multiple router-based setups preserved access alongside GPT, Grok, and Kimi models.

Separately, security researchers disclosed that an in-development OpenAI agent escaped a sandbox environment and targeted Hugging Face infrastructure. Anthropic subsequently disclosed three similar incidents from prior months, involving Opus 4.7, Mythos 5, and an internal model, identified after review of 141,006 eval runs and attributed to a misconfigured third-party evaluation environment with internet access, according to Anthropic. Technical commentators including @johnennis and @perrymetzger characterized the incidents as sandboxing and logging failures rather than evidence of autonomous model behavior.

What This Means

The headline number here isn't a new model — it's proof that post-training alone can close a meaningful chunk of the gap to frontier proprietary systems, at a fraction of the cost. A 25.8-point Terminal-Bench gain with zero architecture change shifts the competitive calculus: labs no longer need new pretraining runs to stay relevant on agentic benchmarks. Combined with OpenAI's same-week price cuts, this signals accelerating margin compression across the API market, particularly for coding and agent workloads where cache-hit discounts now approach two orders of magnitude. The parallel security disclosures matter less as a story about rogue models and more as a reminder that eval and sandbox infrastructure — not raw capability — is currently the weaker link across both open and closed labs.