Claude Jupiter Leaks and a Gemini Flash Upgrade

The AI industry is doing that thing again where everyone pretends not to be racing while obviously racing. This week brought leaks, quiet upgrades, and benchmark results that suggest we're both closer and further from AGI than the hype cycles would have you believe.

Anthropic's Planetary Naming Scheme Returns

Testing Catalog reports that Anthropic has begun internally red-teaming a new model codenamed "Claude Jupiter v1." If that sounds familiar, it's because Anthropic used "Neptune" as an internal codename before unveiling the Claude 4 family last year. Planetary names apparently = imminent launch in Anthropic's playbook.

The timing is conspicuous: Anthropic's "Code with Claude" developer conference is scheduled for May 6th—next week. The current Claude lineup features Opus 4.7 prominently, but updated Sonnet and Haiku variants remain conspicuously absent. For developers who've integrated Claude into coding workflows (and there are many), this gap matters. Sonnet models have become workhorses for AI agents and development tools, so an upgrade would fill an actual need rather than just increment version numbers.

According to Testing Catalog's report, Jupiter is currently undergoing "safety evaluations, jailbreak testing, and constitutional classifier stress tests" as part of Anthropic's responsible scaling policies. That's the bureaucratic way of saying: they're making sure it won't immediately do something embarrassing when released to the public.

Prediction markets on Polymarket show heavy betting against a Claude 5 launch before May 31st, which tracks with reality. Most speculation centers on this being Sonnet 4.8 or possibly Haiku 4.7—meaningful upgrades to existing model tiers rather than a full generational leap. As the video creator notes, "I think Claude 5 is still unlikely right now, but a meaningful Sonnet upgrade feels very realistic."

What makes Jupiter interesting isn't just the model itself but what it suggests about Anthropic's release cadence. They're not sitting still while OpenAI and Google make moves.

Cardinal: Your AI Activity, Visualized

While we're on Anthropic: internal strings also reference a feature called "Cardinal," which appears to be a visual memory analytics dashboard. Instead of Claude's memory system operating silently in the background, Cardinal would reportedly let users see monthly breakdowns of their activity—conversation clusters, working patterns, areas of focus.

It's the kind of feature that sounds minor until you think about what it enables. Right now, most AI memory systems are black boxes. You trust they're learning about you, but you can't really audit what they've learned or how they're using it. Cardinal flips that: your interaction patterns become visible data you can inspect and potentially adjust.

The feature is expected to appear in settings alongside a redesigned layout for mobile, web, and desktop apps. Whether it ships alongside Jupiter or separately remains unclear.

Google's Stealth Gemini Upgrade

With Google I/O less than three weeks away, Google appears to be stress-testing new Gemini models in public. The Gemini 3 Flash model in LM Arena (now rebranded as just "Arena") quietly updated this week, keeping the same model slug but delivering dramatically different output quality.

Early testers report the performance jump feels "about two tiers above the current version," with reasoning and response quality closer to Gemini 3.1 Pro than the standard Flash model. AI Battle ran comparison tests and noted getting the new Flash variant in six out of seven Arena battles—a remarkably high hit rate for what's supposedly still in testing.

One tester requested a Minecraft clone and got back working code with block placement, breaking mechanics, and infinite terrain generation. For a Flash model (read: the lightweight, cheap-to-run version), that's genuinely surprising output quality.

Whether this is early Gemini 3.1 Flash, 3.2 Flash, or the rumored 3.5 Flash remains unclear. Google isn't saying, probably because they're saving the formal announcement for I/O. But the fact they're testing publicly suggests confidence in the upgrade's readiness.

Meanwhile, Vertex AI customers received emails about Gemini 3.1 Flash Lite becoming generally available soon, indicating Google is preparing rollouts across multiple tiers simultaneously.

OpenAI Adds Pets to Codex (No, Really)

OpenAI's latest Codex update includes an animated pet system that's more useful than it sounds. Using /pet summons a companion that floats on your screen, and /hatch lets you customize its appearance. Cute, sure—but these pets actually surface useful information.

The overlay displays live Codex activity: which thread is active, whether the agent is running or waiting for input, progress updates on background tasks. For long-running AI agent workflows, this means you can monitor what's happening without constantly reopening the app. As the video notes, "It's such a smart way to make long-running AI agent workflows feel more alive and interactive."

Codex also got a migration system that imports settings, plugins, agents, and project configurations in a few clicks. The friction reduction is intentional—OpenAI wants switching to Codex to feel inevitable rather than effortful.

ARC-AGI-3: The Humility Benchmark

And then there's ARC-AGI-3, the benchmark that exists primarily to humble everyone. New scores dropped this week for GPT-5.5 and Opus 4.7, and they're... not inspiring.

OpenAI's GPT-5.4 scored 0.02% in high mode. GPT-5.5 improved to 0.4%. Anthropic's Opus 4.6 Max hit 0.5%. Opus 4.7 is at 0.2%. These are rounding errors, not capabilities.

The benchmark tests abstract reasoning and pattern recognition in ways that require genuine generalization rather than pattern matching on training data. Despite massive improvements in coding, reasoning, and agentic workflows, frontier models still barely scratch these tasks. As the video creator puts it: "Despite all the massive improvements we have seen in coding, reasoning, and agentic workflows, benchmarks like this continue reminding everyone that true generalized intelligence is still very hard."

This matters because it's easy to mistake specialized competence for general intelligence. GPT-5.5 can write production code and argue philosophy, but it fails basic abstract reasoning tests that humans find trivial. AGI isn't just "better at everything we currently test"—it requires capabilities we're still figuring out how to build, let alone benchmark.

The Pricing Wars Continue

GitHub is preparing to launch Copilot Max, a $99/month tier targeting power users and advanced workflows. Right now it's mostly a placeholder, but the pricing signal is clear: GitHub believes there's a market for premium AI coding tools beyond the standard Copilot subscription.

Meanwhile, xAI launched Grok 4.3 through their API alongside "Imagine Agent Mode," which combines brainstorming, content generation, image creation/editing, and video generation in one continuous workspace. It's positioning Grok as a creative canvas rather than just a chatbot—text, images, and video generation all accessible without switching tools.

Whether this approach gains traction depends heavily on execution quality, but the ambition is notable. xAI is explicitly trying to build workflows, not just model endpoints.

What Next Week Holds

May 6th isn't just a date on Anthropic's calendar—it's become a focal point for speculation about where Claude fits in the current AI landscape. If Jupiter turns out to be Sonnet 4.8, that's significant for developers who've built on Claude. If it's something else, that's a different kind of significant.

Google I/O follows shortly after, likely bringing formal announcements of whatever they're currently stealth-testing in Arena. And somewhere in the background, OpenAI continues making Codex stickier while benchmark creators keep designing tests that humble everyone.

The race is heating up, but ARC-AGI-3 suggests the finish line is further than the version numbers imply.

—Yuki Okonkwo