👀 GPT-5.6 is rumoured to be dropping in 10 DAYS! And here's all we know about 5.6:
5.6 briefly appeared in OpenAI's own Codex logs before disappearing, it's usually a sign that internal testing is ramping up. 🔥
Rumors so far:
• Major improvements in token efficiency (better performance at lower cost)
• Stronger reasoning and agent workflows
• Cleaner code generation and frontend output
• GPT-5.6 and a Pro version launching together
• Possible context windows up to 1.5M tokens under codenames like iris-alpha
Alpha: Polymarket degens are heavily betting on June 9–11 window 👀
And mind you, it has been just 1 month since GPT-5.5 dropped! 🤯
Proving rumours right, Anthropic releases Claude Opus 4.8 📈
The new flagship model claims to improve judgment, honesty about its limits, and long-running autonomy with minimal human input.
→ Priced the same as 4.7.
→ New Fast Mode delivers 2.5× speed at one-third the cost.
Btw, Elon Musk responded: “Nice work.”
Gemma 4 meets retro gaming… but it's actually an agentic masterclass. 😳
Google just dropped AIventure, an open-source dungeon crawler built to show devs how to actually wire up agentic workflows.
So the thing is, it’s basically vibe-coding as a game mechanic. You can literally prompt an in-game NPC to build a functioning web app for you on the fly. 🥸
So instead of grinding through terminal, we be clearing pixels to deploy software? yeah.
It's a clever trojan horse though. Pack a nostalgic dev tutorial inside an 8-bit aesthetic, but use it to sneakily map out how multi-agent ecosystems should actually communicate. 👀
If you're trying to figure out how to transition from static prompts to dynamic, multi-agent infrastructure, this is the architecture to look at.
Anthropic’s “restricted” Mythos model might not stay restricted for long. 👀🚫
New leaks suggest Anthropic is preparing Mythos 1 (“claude-mythos-1-preview”) for Claude Code and Claude Security, and some users briefly spotted it live inside Claude’s UI. 😳
That’s interesting because Anthropic previously positioned Mythos-class models as locked down until stronger safeguards were ready. But now, the language seems to be shifting.
Mythos is already helping protect a wider range of organizations, including open-source projects, through project glasswing and Anthropic has hinted a future general release.
But this still doesn’t look like a normal public Claude launch.
It looks like Anthropic is taking the enterprise-first route, putting Mythos into coding + security products first, alongside a revamped Claude Security dashboard with vulnerability discovery, historical scan charts, and deeper triage.
Basically…their most powerful models may show up as products before they show up as chat models.
Oh, and Claude Opus 4.8 is reportedly being tested with partners too.
Anthropic’s shipping a lot quietly right now. 🤫
Latest: DeepMind CEO Demis Hassabis says current AI systems are still far from AGI. 🤖
He noted that even if AI can solve large numbers of Erdős-style math problems, it remains nowhere close to the creative invention and genius of minds like Ramanujan.
Just In: Anthropic just released their project glasswing's first update. 👀
Project Glasswing is Anthropic’s collaborative cyberdefense initiative. It uses their Claude Mythos to scan critical open-source and enterprise software, catching and fixing vulnerabilities before anyone can exploit them.
✋ Here are the key takeaways so that you don't have to read it fully:
• Massive Scale: Found 10,000+ critical software bugs globally in just one month.
• 10x Speedup: Partners like Cloudflare and Mozilla are finding bugs ten times faster than before.
• High Accuracy: Achieved a verified 90.6% true-positive precision rate.
• Human Bottleneck: AI is finding vulnerabilities much faster than human developers can patch them.
• Major Exploit Found: Uncovered a critical flaw in wolfSSL that could forge bank certificates.
• Financial Defenses: Successfully stopped a live $1.5 million fraudulent wire transfer at a partner bank.
• Public Release Delayed: Mythos-class models remain restricted because they are still too dangerous if weaponized.
Latest: Perplexity open sourced Bumblebee... yeah, like the transformer, but for security. 🐝
They’re giving away the internal tool they use to stop hackers from hijacking developer laptops. So tldr: less "we got breached through a bad npm package" and more actual peace of mind.
And not so surprisingly, the core idea is still very grounded: bad code → broken systems → Perplexity down. Bumblebee stops that.
But the interesting part isn’t the open-source announcement, it’s how it works. Normal security scanners actually run code to check it, which can accidentally trigger the exact virus they're looking for.
Bumblebee is 100% read-only. It just looks at the files, never runs them, and checks for:
• Dodgy code packages (npm, Python)
• Sketchy browser/code extensions
• Rogue AI agent configs
No code execution, no network monitoring, no accidental hacks.
and tbh Perplexity releasing a cybersecurity tool for Mac and Linux was not on our cards.