In partnership with
Two of the strongest coding models you can call today were released on the same day this week, and neither came from a lab you hold a contract with. One of them shipped under MIT, which means the weights are yours to run on your own hardware.
The model layer was always going to commoditise faster than the people building on top of it expected. This is the week it stopped being an argument and turned into a price list. So the rest of this issue is about what happens next: a Ukrainian pre-seed that just sold the one thing cheap models still cannot do, a Reuters investigation into what the same cheap capability does in the hands of a ransomware crew, and a Chinese lab trying to build a toll booth before the road is finished.
Zhipu priced the frontier at fifteen cents
A model called Ox Alpha sat anonymously on OpenRouter and OpenCode for a week, serving what its maker says was around 100 trillion tokens a day. On 26 August Zhipu claimed it as GLM-5.3-Flash: 320 billion parameters with 18 billion active, a one-million-token context window, 1,773 on GDPVal-AA v2, and $0.15 per million input tokens against $0.50 output. The weights went up on Hugging Face under an MIT licence. The entire trial ran on Chinese silicon.
Alibaba shipped the same day. Qwen3.8-Flash-Next activates 6 billion parameters per token out of a 125 billion backbone and scores 62.5 on SWE-bench Pro where Claude Opus 4.6 scores 53.4, at $0.16 per million input. Read both licences before you plan around either. Zhipu's MIT lets you run those weights on a box in Frankfurt or Prishtina and answer to nobody; Alibaba ships under qwen-community-1.0, which carries conditions. Pressure is arriving from the hardware side too: SemiAnalysis's teardown of OpenAI's in-house Jalapeño chip shows better tokens per megawatt than NVIDIA's Blackwell before it has reached volume.
Take the one feature you have been rationing because the token bill frightened you, and run it against GLM-5.3-Flash this week. Two years of "we cannot afford to do that on every request" is now an untested assumption, and re-testing it changes what you can put in front of a customer in Q4. The teams that reprice in September will ship what their competitors have quietly priced out.
Holiday Creator Calendars Are Filling Up. Q4 Panic Is Optional.
Creators lock in their holiday content calendars 90 days out, before most ecommerce brands finalize their commission strategy and way before Black Friday and October deal events.
Get ahead of the seasonal rush with The 90-Day Holiday Sprint, a practical guide for brands that want creators driving holiday demand while competitors are still recruiting:
Structure commissions by lifetime value, not just first-order margin
Lead with the right products so creators promote with confidence
Recruit and onboard creators with a day-by-day plan for the first 30 days
Read performance early and pull program levers by Day 60
Brief creators with a holiday checklist before calendars fill up
Your 90-day countdown starts now.
A Ukrainian team of under ten sold the thing cheap models cannot do
On 25 August Embedd closed a €2.31M pre-seed led by Seedcamp, with Roosh Ventures, Vesna Capital, U.ventures, Cocoa, Connect Ventures, 2100 Ventures, Underline Ventures and Common Magic alongside. Mykhailo Lazarenko, Maksym Horinov and Valentyn Hololobov founded the company in 2023, after their previous hardware venture died between COVID supply chains and the full-scale invasion. The product builds digital models of physical chips so an agent has enough context to write the integration code itself. Microchip Technology is already working with them, and Embedd claims six times faster delivery of chip software.
Hold that next to fifteen-cent tokens. A frontier model has read most of the public internet and still cannot tell you which pin on a particular microcontroller does what under which errata, because that knowledge lives in PDFs nobody scraped and in the heads of people who have shipped firmware. When reasoning gets cheap, the margin moves to the substrate you feed it. Embedd is selling the substrate, not the reasoning.
Seedcamp leading a Ukrainian pre-seed at this size is also a price mark for the whole non-defence deep-tech cohort in the region, and it arrives while most regional capital is still chasing drones. For founders in Lviv, Cluj, Skopje or Novi Sad, the pitch that clears a partner meeting right now is not another wrapper on a cheap model. It is a body of knowledge the cheap model cannot scrape, packaged so an agent can use it.
A ransomware crew ran seven break-ins through a coding agent
Reuters reported on 27 August that Gambit Security, a Tel Aviv firm, recovered 28 chat sessions between a Russian-speaking crew calling itself Aur0ra and Cursor's AI agent. Between 8 April and 21 May the group breached at least seven companies. The named victims are Christeyns, a Belgian hygiene manufacturer, Teckentrup, a German garage-door maker, Scotland's Helideck Certification Agency, Bayou Title in Louisiana, an Argentine pharmaceutical distributor and an Italian manufacturer. The guardrail bypass was one sentence: the operators told the agent the work was a simulation, and the reasoning trace shows it accepting the story.
Read that victim list again. Four of the six identified are mid-market European industrials, which is the profile of half the manufacturing clients billing hours in Belgrade, Zagreb and Bucharest right now. Gambit's threat intelligence director Eyal Sela puts the operator speed-up at 30 to 50 percent. The same price collapse that lets you run an agent on every support ticket lets someone else run one on your network, and their unit economics improved this week too.
Neither Cursor nor Anthropic commented. That leaves the work with you, and it is not a tooling purchase. Pull up whatever your incident response plan says about detection windows and ask whether it survives an attacker moving 40 percent faster than the one it was written for. Then ask your dev leads which agents hold production credentials, because that inventory is a security document now, not a procurement one.
Moonshot wants thirty percent before the weights are free
Reuters reported on 26 August that Moonshot AI is asking Microsoft, Amazon and Google for up to 30 percent of the revenue those clouds earn from services built on Kimi K3, its 2.8-trillion-parameter model. The talks cover how revenue gets split, what data access looks like, and how token usage is audited. Nothing is signed and the sources are explicit that no agreement is certain. Moonshot raised more than $2 billion in May. A deal would be the first large revenue-share between a Chinese lab and a US hyperscaler.
Most cost models in this region quietly assume Chinese open weights stay free forever, because so far they have. Moonshot is the first serious attempt to convert that distribution into rent, and the mechanism it is negotiating, auditing token usage inside someone else's cloud, is the part to watch rather than the percentage. Zhipu's MIT licence is the generous end of a spectrum, not the direction the whole market is travelling.
Nothing here needs action this month, which is exactly why it gets forgotten. The watch item: when a licence or pricing change on your primary open-weight model would break your margin, you do not have a cost advantage, you have a subsidy. Find out which one you are running on before your next board update, because the answer takes an afternoon now and a quarter later.
Short Signals
Five things to install or check this week, tagged by the seat they help.
Dev: IBM shipped Granite 4.2 under Apache 2.0. The 3B, 8B and 30B family landed on 25 August with context to 512,000 tokens, toggleable thinking modes and OpenAI-format tool calling. The 8B and 30B were trained with agentic reinforcement learning for tool use, code execution and sandboxed search. Free, on Hugging Face and Ollama. The smallest genuinely permissive option for anything that has to run inside a client's firewall.
Sales: Salesforce moved its CRM inside Claude. Claudeforce shipped to pilot customers on 26 August with 37 prebuilt sales skills covering deal review, meeting prep and pipeline updates, with open beta in September. Claude also becomes the default model in Slack. If your team lives in Salesforce, put a September calendar block on testing whether the chat window replaces three of your tabs.
Productivity: Gemini 3.5 Transcribe went to public preview. Google's new speech model covers 85+ languages with auto-detection, posts 2.6% word error rate non-streaming, and cuts latency about 70 percent against Chirp 3. It strips filler words and handles self-corrections. Roughly half a cent a minute, with a free tier. Worth pointing at your last five client calls before you pay a transcription vendor again.
Research: Accel funded a web index built for agents. Keenable came out of stealth on 25 August with a $26M seed and an index of over 100 billion documents, built because Google and Microsoft closed their search APIs to everyone but enterprise. It ships an MCP server at up to 1,000 requests an hour. Founded by Andrey Styskin, previously running search at Yandex.
Founders: 550 engineers on what agents did to their week. Temporal's State of Development 2026 is free and published 26 August. Daily agent use hit 80.8 percent, up from 47.3 percent. Median five agents per person. But 41.1 percent hit agent failures daily, and 92.3 percent have tried rebuilding software they already pay for. Read the failure numbers, not the adoption ones.
Next edition soon,
Çelik