Alibaba shipped Qwen3.8-Max at 2.4 trillion parameters on August 3, just under Moonshot's Kimi K3 at 2.8 trillion. The US answer this week is a voluntary safety testing framework presented to OpenAI, Anthropic and Google.
The Mixture-of-Experts model landed Monday, August 3 with 2.4 trillion total parameters and roughly 95 billion active per token. It scored 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Fable 5 at 84.6 but behind GPT-5.6.
A research firm cited by Reuters puts V4-Flash more than 100 times below Anthropic's Claude on price to run. It landed the same day Alibaba claimed near frontier coding scores, so the cheap tier and the top tier converged in one news cycle.
The President said the United States postponed what he called the biggest attack since World War II in favour of renewed negotiations. Energy and shipping desks will price the reprieve before builders feel it.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →Alibaba unveiled Qwen3.8-Max on Monday, August 3, calling it its most powerful model to date: a Mixture-of-Experts system with 2.4 trillion total parameters, roughly 95 billion active per token, and a 1 million-token context window. Alibaba says it ran a software engineering project autonomously for 16 days in...
Continue reading arrow_outwardAlibaba's Qwen3.8-Max landed Monday with 2.4 trillion total parameters, roughly 95 billion active per token, a 1 million-token context window, and a claimed 16-day autonomous software engineering run.
Trump says the United States shelved what he called the "biggest attack since World War II" on Iran in favour of talks he claims begin Monday, after a second consecutive night of strikes and Iranian retaliation.
Sentiment sits at Extreme Fear, an unforgiving backdrop for capability announcements priced in trillions of parameters and gigawatts of compute.
OpenAI's continuous-voice GPT Live keeps a channel open instead of trading turns, which only pencils out economically at the inference prices DeepSeek's V4-Flash is now setting more than 100 times below Claude.
Qwen3.8-Max at 2.4 trillion parameters approaches but does not exceed Moonshot AI's Kimi K3 at 2.8 trillion, and both target the same workload: long-horizon coding and knowledge work behind a 1 million-token context window. Two Chinese labs setting the size frontier within days of each other changes what counts as a reference point for frontier scale.
A research firm cited by Reuters puts DeepSeek's V4-Flash more than 100 times below Anthropic's Claude on price to run, landing the same day Alibaba claimed near-frontier benchmark scores including 86.6 on a coding evaluation. The argument is no longer that Chinese models match US quality; it is that they compete on unit economics while getting close enough on capability.
The White House is convening OpenAI, Anthropic and Google this week to present a US framework for voluntary safety testing, an opt-in approach stemming from a June executive order on AI cybersecurity. Opt-in review binds only the labs in the room, which excludes the developers currently resetting both the size and price frontiers.
3 Cents Versus 3 Dollars
Read full analysis arrow_outward
Deep Dive
7.0/ 10
Your agent stalls on a yes-or-no question while you are making coffee; Pushary moves that question to your lock screen and gets the run moving again.
This is for people running two or more coding agents who keep losing minutes to approval prompts nobody is at the desk to answer. If that is you, the $9.99/mo is cheap against the dead time, and the MCP-plus-hooks install is genuinely a one-command affair with Claude Code. If you run a single agent on Claude Code with Claude Max, Anthropic's own Remote Control already covers that case for free and Pushary is redundant.
“Supporters like that it collapses an agent pause into a simple yes button on your phone, so the work keeps moving while you are away from the desk.”Product Hunt
Raycast's mobile app now exposes dictation, AI commands, Snippets and Quicklinks directly from the keyboard, which means the launcher shortcuts you built on desktop are usable mid-conversation on a phone. The practical win is Snippets: move your three most-retyped blocks of text there and test whether the keyboard surface is fast enough to replace the copy-paste habit. Check the site for current pricing tiers before committing a team.
Try it arrow_outwardThere's An AI For That relaunched with an extensive tool database, personalized suggestions and an AI Job Impact Index. The database is the least useful part; the Impact Index is worth twenty minutes because it forces you to name which of your own recurring tasks are already automatable rather than browsing tools with no problem in hand. Access is free, with paid options available.
Try it arrow_outwardBeatMV appears in directory listings at v3.2.6, which suggests a shipping product rather than a landing page, but the public description is thin on what the pipeline actually does with your audio. Give it one track you already own the rights to and judge the cut timing against the beat before you plan any client work around it. Treat the output as a rough assembly, not a deliverable.
Try it arrow_outwardMake3DAI is running as a featured listing in the tool directories this week, aimed at generating 3D models from prompts. If you need placeholder geometry for a prototype scene or a product mock, that is exactly the low-stakes job to test it on, since topology quality is where most text-to-3D tools fail. Export early and open the file in your own editor before you decide it fits the pipeline.
Try it arrow_outwardBenchmark your coding agent against Qwen3.8-Max. Take one long-horizon task you already run on Claude Opus 4.8 or GPT-5.6 and rerun it on Qwen3.8-Max. The 86.6 Terminal-Bench 2.1 score and the 16 day autonomous run claim only matter if they hold on your repo.
Build a two-tier routing layer this week. Route cheap, high-volume calls to DeepSeek V4-Flash and reserve frontier models for the hard 10 percent. At more than 100 times below Claude on price to run, the arbitrage pays for the routing work in days.
Put agent permissions where you will actually see them. Install Pushary so agent permission prompts hit your lock screen instead of a buried terminal, and pair it with the Raycast for iOS update for on-the-go triage. Voluntary safety frameworks will not audit your own agents for you.
Chinese labs set both ends of the range this week: Kimi K3 holds the 2.8 trillion parameter ceiling, Qwen3.8-Max claims near frontier coding at 2.4 trillion, and V4-Flash sits more than 100 times under Claude on cost. Washington's contribution is a voluntary testing framework presented to OpenAI, Anthropic and Google, which means procurement and liability, not policy, will do the filtering. TrustHouse.AI's August 3 continuity product and widening obligations from August 2 are the tell that vendors are already selling insurance against this. Build for a world where you route across tiers, verify autonomy claims yourself, and treat safety instrumentation as your own engineering line item.
AI analysis, world news, markets, and tools. One briefing, delivered free.