Chosun reporting puts 1.4 trillion tokens through five apps in 72 hours for the anonymous ox-alpha model, served free at hundreds of millions of won. The trail points to China.
Chosun's numbers reframe ox-alpha from curiosity to capital commitment: 1.4 trillion tokens across the top five developer apps on OpenRouter within three days of launch. Free serving at an estimated cost in the hundreds of millions of won, with the trail now pointing to China.
Chris Lehane told the Guardian to prepare for ongoing, persistent cyber-attacks from models that can plan and launch offensives themselves. Defense budgets built around discrete incidents do not survive that shift.
Abbas Araghchi dismissed fresh US sanctions threats on 23 August 2026 as a sign of desperation. The same day, S. Jaishankar arrived in Russia for the 27th India-Russia Inter-Governmental Commission ahead of the BRICS summit.
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 →New reporting from Chosun puts numbers behind the anonymous frontier model on OpenRouter: over 1.4 trillion tokens flowed through just the top five developer apps within three days of launch, with free serving costs estimated in the hundreds of millions of won. The model carries a 1,048,576-token context window...
Continue reading arrow_outwardChosun reporting traces the anonymous OpenRouter frontier model ox-alpha to a Chinese developer, after over 1.4 trillion tokens moved through its top five apps in three days at an estimated free-serving cost in the hundreds of millions of won.
Iranian Foreign Minister Abbas Araghchi dismissed fresh US sanctions threats on 23 August 2026 as a sign of 'desperation', keeping a floor under oil and shipping risk premiums.
Sentiment sits at Greed, undeterred by an unattributed frontier model, an OpenAI warning about persistent AI-driven cyber-attacks, and sanctions escalation over Tehran.
Three stories, one missing signature: nobody has claimed ox-alpha, nobody signs the record when an AI agent runs regulated analysis, and OpenAI says the attackers now plan their own offensives.
Chosun's numbers reframe ox-alpha from curiosity to capital commitment: 1.4 trillion tokens across five apps in 72 hours, served free at an estimated cost in the hundreds of millions of won, with a 1,048,576-token context window and image plus video input. Only an actor willing to eat that bill for reach would run this playbook. The reporting points to China, which makes the giveaway a distribution strategy rather than a demo.
OpenAI's Chris Lehane told the Guardian to prepare for 'ongoing, persistent' cyber-attacks from models that can now plan and launch offensives, framing it as 'a different chapter'. The timing matters: it lands the same week OpenAI said it paused training of some frontier work. A lab warning about capability while restraining its own pipeline is an argument aimed at everyone else's pipeline.
The August 22 syndicated analysis puts a procedural question on the table: audit trails, sign-off authority and liability were all built around a named human analyst, and agents do not have names in that sense. Medtech and finance hit this first because regulators there demand attribution. Pair it with an unattributed frontier model and autonomous offensive capability, and the common thread is systems operating with nobody's name on them.
Your team is already running an agent with a real shell and real credentials; OneCLI's whole pitch is putting walls around that before someone rm -rf's production.
OneCLI is for platform and security leads who have already lost the argument about whether staff will use AI agents and now need those agents boxed in: per-person sandboxes, human-in-the-loop approvals, and a credential layer where the agent never sees the keys. If you are self-hosting anyway and want an open-source harness rather than a black-box SaaS, it is worth a pilot this quarter. If you want the most capable general software agent, this is not that; it is the governance layer around one.
Epho hosts the agent instead of your laptop, so long refactors and test runs keep going after you close the terminal, and each task runs sandboxed against the repository rather than your local filesystem. Point it at one contained job first, a dependency bump or a test backfill, and compare the diff quality against the same agent running locally before you route real work through it.
Try it arrow_outwardRather than a single generator, MiniMax Design is a desktop creative agent that chains image, video, voice, and editing models into a finished asset, which suits marketing batches more than one hero image. Test it on a campaign you already shipped by hand and time the two runs; the orchestration only pays off if you avoid re-editing every output.
Try it arrow_outwardAdam targets hardware teams and connects to the CAD, PLM, supplier, and tooling stack already in place, including Onshape and Autodesk Fusion, so nobody has to migrate a parts library to try it. Give it a real bill-of-materials question or a revision lookup, the tedious lookups that eat mechanical engineers' afternoons, before you judge it on generative geometry.
Try it arrow_outwardKIVA pulls Search Console data, layers user intent customization on top, and offers one-click content creation, which makes it an agency workflow tool rather than another keyword list. Start with the free tier on one client property and check whether its intent groupings match what your analyst would have written; that comparison tells you if it saves hours or just generates more drafts to review.
Try it arrow_outwardPoint Epho at one real repo today. Run Claude Code, Codex, or OpenCode against a live branch in the cloud and compare all three on the same open issue. Log wall-clock time and token spend per agent so you have your own cost curve, not a vendor's.
Sandbox the agent you already trust too much. Move whatever your team runs unsupervised into OneCLI and watch what it actually reaches for. Given Lehane's warning about persistent, model-driven offense, treat an unsandboxed shell agent as an open port.
Add a named signer to every generated artifact. Following the August 22 analysis on audit trails and sign-off authority, stamp each agent output with a human owner, a timestamp, and the model version used. If nobody will sign it, it does not ship to a regulated workflow.
Someone is spending hundreds of millions of won to serve 1.4 trillion tokens for free and declining to sign their name to it. That is a deliberate strategy, and it lands the same week OpenAI's policy chief warns that model-driven attacks are becoming continuous rather than episodic. Builders now face two unsigned problems at once: unknown provenance on the models they call, and unknown accountability on the outputs those models produce. Pick your sandbox and pick your signer this week, because both gaps close slower than the capability curve rises.
AI analysis, world news, markets, and tools. One briefing, delivered free.