0:00 / 3:44
Chapters
Sources
DAILY ROUNDUP
Ox Alpha: A Free Stealth Model Beating GPT And Fable At Coding
calendar_today Date:
schedule Duration: 3:44
visibility 2 Views
A mysterious free stealth model, Ox Alpha, beats GPT-5.6 and Claude Fable 5 at coding, Claude Security now runs on Mythos 5, AI agents burn 5x more tokens than humans, and graduate jobs drop nearly 50% in a year.
- 01. Ox Alpha's technical fingerprint points to Zhipu AI's GLM-5.X family, with a 1M-token context window and free access during its stealth preview
- 02. Claude Security's Mythos 5 scans trace data flows across files and classify findings by severity, confidence and suggested fixes, no separate model access needed
- 03. AI agents now burn nearly 5x the tokens humans do, up 14x since February, as agentic workflows loop through many more steps per task
- 04. Graduate vacancies fell from over 15,000 to just over 8,000 in a year, with some companies cutting entry-level hiring by up to 80% in a quarter
Today's AI news: a mysterious free model called Ox Alpha has shown up on OpenRouter and inside coding tools like Cline, beating GPT-5.6 and Claude Fable 5 on several coding benchmarks - the technical fingerprint points to Zhipu AI's GLM family, though nobody has claimed it officially. Claude Security scans now run on Claude Mythos 5, in public beta for every Claude Enterprise customer. A new a16z chart shows AI agents burning nearly 5x as many tokens as human users. And graduate job vacancies have dropped by almost half in a year as employers replace entry-level roles with AI.
Chapters:
0:00 Today's AI News
0:25 Ox Alpha Stealth Model
1:24 Claude Security Mythos 5
2:06 Agents Outpace Humans
2:58 Graduate Jobs Drop
Blend Roundup 2026-08-24
https://x.com/elder_plinius/status/2090910167345049661
https://x.com/cline/status/2090854216399220985
https://x.com/claudeai/status/2090852314319880425
https://x.com/a16z/status/2091200032162857328
https://x.com/PolitlcsUK/status/2091844930662465876
[quick] A free stealth model called Ox Alpha is beating GPT and Fable at coding, Claude Security now scans codebases with Mythos 5, AI agents now burn five times more tokens than humans, and graduate jobs have dropped nearly 50 percent in a year. [excited] Here's today's AI news.
[quick] A mysterious free model called Ox, or maybe O-X Alpha has shown up on OpenRouter and inside coding tools like Cline, and early testing suggests it's genuinely strong, beating GPT-5.6 and Claude Fable 5 on several coding benchmarks.
Nobody has claimed it officially, but the technical fingerprint points heavily towards Zhipu AI, the company behind the GLM family, and Pliny the Liberator says his own agent confirmed it as an unreleased model from the GLM-5.X line.
It ships with a million-token context window, up to around 130,000 tokens of output, and support for text, image and video input, and it's completely free during this stealth preview, which is expected to run about a week.
Results are mixed depending on which benchmark you trust, so treat the strong scores as promising rather than settled until an official release confirms what it actually is.
[quick] Claude Security scans now run on Claude Mythos 5, Anthropic's most capable cybersecurity model, and it's now available in public beta for every Claude Enterprise customer.
It scans a chosen repository, traces data flows across files, and returns findings classified by severity, confidence and suggested fixes, catching complex multi-component vulnerability patterns that traditional scanners typically miss.
There's no separate model access required and no extra add-on cost - Mythos 5 runs quietly in the background during a scan, billed through your existing Enterprise plan's normal token usage.
[quick] Humans are now officially the minority user of AI - a new chart from a16z shows AI agents burning through nearly five times as many tokens as human users, based on data from OpenRouter.
Agentic usage first overtook human usage back in February this year, and it's grown roughly fourteen times since then, while human-driven usage has only grown about two-and-a-half times over the same period.
The reason is structural - an agent plans, calls a tool, reads the result, calls another tool, drafts, checks itself and revises, looping through many steps for one human prompt.
It's a genuinely useful reminder that the AI economy isn't really about chat anymore - it's increasingly agents talking to tools, with humans further back in the loop than the usage numbers used to suggest.
[quick] Graduate job vacancies have dropped by almost half in the space of a single year, with employers replacing entry-level roles with AI rather than hiring junior staff.
One jobs site found just over eight thousand graduate vacancies listed in July, down from over fifteen thousand at the same point last year, and big tech firms globally have cut fresh-graduate hiring by more than half over the last three years.
Some companies that adopted generative AI cut entry-level hiring by as much as eighty percent in a single quarter, and recent graduate unemployment has climbed alongside it.
Employers are also restructuring the junior roles that remain, loading them with senior-level skills like judgement and stakeholder management - what researchers are now calling seniorisation.
Meta Data
LLM:
Model: