AI Digest August 1: DeepSeek V4-Flash Upends Agent Economics
DeepSeek shipped V4-Flash 0731 with open weights: +25.8 on Terminal-Bench with no architecture change and prices from $0.14 per million tokens. ARC-AGI-3 showed that evals measure the whole agent system, not the model; in the markets — a $30B AI fund implosion and one billion ChatGPT users.
Today's top stories
DeepSeek shipped V4-Flash 0731 — and open-sourced the weights immediately. The public API beta launched with a claim that the upgraded Flash now surpasses V4-Pro-Preview on agentic work, supports the Responses API format and is "fully adapted for Codex" (DeepSeek announcement). The jump was visible at once: Terminal-Bench 82.7 — up 25.8 points over the April preview. Most striking: the architecture did not change — still 284B total parameters (13B active), 1M context, priced at $0.14/$0.28 per million tokens with a 98% cache-hit discount (Artificial Analysis breakdown). The weights landed under MIT: vLLM documented 256 routed experts, 6 active per token and a bundled speculative-decoding module, while Unsloth published runnable quants starting around 110GB of memory.
ARC-AGI-3 made it official: we are evaluating systems, not models. François Chollet clarified the benchmark rules: bespoke benchmark-specific harnesses are banned, but general-purpose API features are fine if settings and cost are disclosed. On the official semi-private set, Opus 5 scores 30.2% while GPT-5.6 Sol gets 7.8% under the standard harness — yet with retained reasoning and context compaction Sol reaches 38.3% on the public set. The takeaway echoed across the field: long-horizon evals measure truncation policy, memory and tool orchestration as much as raw weights.
Money of the week: a $30B hedge fund implosion — and one billion ChatGPT users. The AI Daily Brief unpacks what the collapse of a heavily leveraged AI fund really means: lab revenues keep climbing, demand still exceeds capacity, and market turmoil is not the same thing as weak fundamentals. Meanwhile Ben's Bites marks the 1-billion-users milestone for ChatGPT, and The Rundown reports that OpenAI's models now cut their own inference costs.
Numbers and facts
- DeepSeek V4-Flash 0731: 284B parameters (13B active), 1M context, $0.14/$0.28 per million tokens, 98% cache discount (Artificial Analysis).
- Terminal-Bench for Flash: 82.7 versus 56.9 for the April preview (Cline).
- ARC-AGI-3: Opus 5 — 30.2%; GPT-5.6 Sol — 7.8% under the standard harness and 38.3% with retained reasoning (summary).
- Running Flash locally: ~168GB of memory for 4-bit and ~110GB for 3-bit quants (Unsloth).
Different perspectives
The debate of the day: open versus closed models in the wake of the recent security incidents. Pro-open: the head of Hugging Face argues that the platform defended itself using an open model, and banning open weights would hurt defenders, startups and researchers most. The other side: "a safe world with closed models" is a position with its own arguments, and the discussion is far from settled.
Tools and techniques
- DeepSeek inside familiar stacks. A router lets you run V4-Flash inside Codex while keeping GPT, Grok and Kimi in the same picker; Cline made the updated model free, and enthusiasts spun up a free public endpoint.
- Skipping the Kimi K3 waitlist. Matt Wolfe shows a working path via the developer platform — an account plus a minimal deposit unlocks the single-agent version without the queue.
In brief
- Superhuman: two startups take on content and invoicing.
- The AI Daily Brief: six questions every enterprise has to answer about AI.
- Matt Wolfe's weekly recap: Opus 5, the Slack killer and Google Earth AI.