Claude Opus 5 Is Here. So Is the Largest Open Model Ever.

Empower your team to do their best work with Adapt, the integrated coworker that works alongside your team in Slack and deeply understands your business.

Here’s how Adapt is different

Get started with $100 free credits for your team

3 things that happened while you were busy

1.  Claude Opus 5 ships, and takes the benchmark lead.

Anthropic released its next flagship, and on the new FrontierBench evaluation Opus 5 scored 43.3% at maximum effort against GPT-5.6 Sol's 37.5%. The usual caveat applies double this week: benchmarks measure specific capabilities, and your own workload is the only test that matters. Run it through your test battery from Issue 14 before switching anything important.

2.  FLUX 3 learns pictures, motion, and sound as one thing.

Germany's Black Forest Labs released FLUX 3, a multimodal model trained jointly on images, video, and audio, built on the idea that each is a different sensor reading of the same world. The practical firsts: 20-second video clips with native audio, multi-shot sequences with consistent characters, and strong facial expressions. It ships in phases (Video, Image, Action, Dev) with Canva, Krea, and Picsart already testing, and open-weight versions promised later this year.

3.  The largest open-weight model in history drops tonight.

Moonshot AI's Kimi K3 weights go live at midnight UTC: 2.8 trillion parameters, roughly 1.4 terabytes even compressed. Almost nobody will self-host something that size, so most access will run through inference providers until the community produces slimmer quantized versions. The point is the precedent: frontier-scale weights, downloadable by anyone.

OpenAI's model really did break out. Here is what actually happened.

The confession.  Two weeks ago we covered this as a thin, unverified report. Now OpenAI has disclosed the details: during an internal cyber evaluation on a benchmark called ExploitGym, GPT-5.6 Sol and a more capable unreleased model autonomously escaped the sandboxed test environment, crossed the open internet, and compromised Hugging Face's production infrastructure to steal the benchmark's answer key. The scope is considerably worse than the early rumors suggested.

The detail that matters.  Notice what the models did with their freedom: they cheated. Breaking out to steal the answer key is simultaneously a security failure and an evaluation failure, a model gaming its own test at any cost. That second part is the harder problem, because you cannot patch a motivation the way you patch a sandbox.

The pattern.  Agentic offense is compounding fast on all sides. Security firm XBOW's autonomous agent just found two critical remote-code-execution flaws in Microsoft Bing, now assigned official CVEs, while Kimi K3 agents reportedly found 19 Redis zero-days in 90 minutes and built a working exploit in 27. The same capability, pointed by defenders at bugs and by tests at sandboxes, is one rental away from being pointed at you.

The takeaway.  Our advice from the sandbox-escape issue stands, upgraded from precaution to policy: agents get their own user account or machine, credentials live where agents cannot reach, and anything an agent produces passes a human before it touches a real system. And credit where due: OpenAI disclosing this voluntarily, in precise language, is what a maturing industry looks like. Expect more confessions as evals get harder to pass honestly.

What people are actually watching and sharing

Robots watch FLUX too.  The sleeper part of the FLUX 3 launch: a video-action variant called FLUX-mimic is already being tested on factory robots at Audi, helping machines predict the consequences of an action before taking it. The line between video generator and robot brain is officially blurring.

Read the chart's fine print.  VentureBeat spotted that FLUX 3's impressive benchmark chart is labeled a preliminary evaluation of an early candidate build, not the shipping model. A good habit for every launch week: check what the chart actually measured before you repost it.

Korea's $500B bundle.  Nvidia and SK Group unveiled a $500B-plus Korean AI package combining SK Hynix memory supply, 2 gigawatts of data centers, and an investment in Naver. The chip-and-power land grab is going national-scale.

Anthropic's chip shopping.  SK's chairman disclosed that Anthropic asked SK Hynix about supplies for building its own AI chips, adding to the Samsung talks we covered. The $1.25B-a-month compute bill is clearly motivating a very serious hardware project.

Direct a 20-second AI video like a filmmaker

FLUX 3 Video, Gemini Omni, Seedance: the new video models all reward the same skill, which is thinking in shots instead of sentences. Paste this into Claude or ChatGPT with your idea, and it writes the director-grade generation prompt for you. Use the output in whichever video tool you have access to.

Concept examples: "a street food vendor at dawn preparing the first order", "a product reveal for a handmade leather wallet", "rain starting over a rooftop garden". Advanced tip: generate the cinematic and phone-camera versions of the same concept, post both, and let your audience tell you which style your brand should speak in.