V4 Pro 0813 pushes DeepSWE from 12.8% to 62.7% and Terminal-Bench 2.1 to 87.9, while API prices rise in the same breath. Agentic coding is now the scoreboard that matters.
V4 Pro 0813 lifts DeepSWE from 12.8% to 62.7% and pushes Terminal-Bench 2.1 to 87.9, with gains concentrated in agentic and coding work rather than chat quality. API pricing moves up alongside it.
Unsloth Desktop shipped in beta on August 11 as a free open-source app for macOS, Windows and Linux, covering GGUF, MLX, diffusion image and video, and audio models. It claims 2x faster training with 70% less VRAM plus sandboxed code execution.
On August 12 the president said the United States controls the waterway and will "keep it," touting a "wall of steel" naval posture while claiming Iran has "no money" and 300% inflation.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →DeepSeek formally released V4 Pro 0813 with reported gains concentrated in agentic and coding work: DeepSWE climbing from 12.8% to 62.7% and Terminal-Bench 2.1 reaching 87.9. The jump lands days after DeepSeek warned customers that its API prices are expected to rise, making the model a test of whether buyers...
Continue reading arrow_outwardDeepSeek's V4 Pro 0813 lifts DeepSWE from 12.8% to 62.7% and hits 87.9 on Terminal-Bench 2.1, days after the lab told customers API prices are expected to rise.
Trump declared on August 12 that the United States controls the Strait of Hormuz and will "keep it," describing a "wall of steel" naval posture while the chokepoint stays effectively closed to normal traffic.
Mood is Fear: a closed Hormuz and rising frontier API rates both point the same direction on input costs.
Three of today's five AI stories are about cheapness rather than intelligence: NVIDIA's Nemotron 3.5 Lightning activates 3B of 30B parameters, Unsloth Desktop claims 70% less VRAM, and DeepSeek is testing whether buyers will pay more anyway.
DeepSeek's headline gains are concentrated in DeepSWE and Terminal-Bench 2.1, not chat quality, and NVIDIA aimed Nemotron 3.5 Lightning at code review, tool use and security alert monitoring. The competitive axis has moved from what a model knows to what it can finish unattended. Price per completed task, not per token, is the number to watch.
Unsloth Desktop landed on August 11 as a free open-source app for macOS, Windows and Linux, covering GGUF, MLX, diffusion image and video, and audio models with claimed 2x faster training at 70% less VRAM. Paired with a 30B model that activates 3B parameters per token, the on-premises floor is rising just as hosted API prices signal an increase.
Koray Kavukcuoglu sits at the center of a leadership reshuffle, with fresh concerns that DeepMind's autonomy is being absorbed into Google proper, and the company emphasized Flash demand and its Cyber model rather than a frontier launch. Two days later at Made by Google '26 it pushed Gemini into Pixel 11, Pixel Watch 5 and a new Pixel Tag. Distribution to hundreds of millions of handsets is the hedge no rival lab can copy.
Deep Dive
7.4/ 10
Most CRMs record what your reps did; Close ships an AI agent named Chloe that picks up the phone and does it.
Close is for small and mid-size sales teams who live on the phone and are tired of paying for a CRM plus a dialer plus a sequencer plus an AI SDR bolt-on. If inbound speed-to-lead is your bottleneck, the built-in dialer and Chloe make it worth a trial; if you need deep enterprise customization or a quoted price before you commit, this is a harder sell.
Re-benchmark your coding agent this week. Point your harness at DeepSeek V4 Pro 0813 and score it on real repository tasks, then compare the delta against your current model at the new API price before you migrate anything.
Test Nemotron 3.5 Lightning on code review. The 30B mixture-of-experts activates only 3B parameters per token, so route pull request summaries and tool-calling steps to it and measure latency and cost against your general-purpose model.
Install Unsloth Desktop and fine-tune locally. Grab the free beta on macOS, Windows or Linux, load a GGUF or MLX checkpoint, and run one small fine-tune to see whether the claimed 70% VRAM reduction holds on your own hardware.
DeepSeek's quintupled DeepSWE score and NVIDIA's code-review-focused Nemotron 3.5 Lightning point the same direction: the models that matter now are judged on whether they can finish a task in a terminal. That shift makes your evaluation harness the most valuable thing you own, because vendor-reported jumps mean nothing until they survive your repository. Meanwhile Unsloth Desktop pushes capable training down to a laptop, so the gap between hosted frontier models and what you can run yourself is narrowing at both ends. Price your agents carefully, since DeepSeek raised API rates in the same announcement that raised the bar.
AI analysis, world news, markets, and tools. One briefing, delivered free.