Gemini 4 Pro Leaks, GPT-6 Soul in Testing: From Arena to Google Cloud, the Week's AI Shockwave

This is the host's first long briefing after a month away from his home setup, and it compresses eight separate leaks into one sweep: Gemini 4 Pro, GPT-6 Soul, Grok 4.7, Union Alpha, Typesafe's Jev, Google's Dream RSI framework, Projects in Claude Code, and Xiaomi's Mimo 2.6 training run. None of them arrived as a formal launch; each surfaced through ghost tests on Arena, quota listings on Google Cloud, free harness trials, or public research logs. For me that is the real story of the week: how models are tested has become as newsworthy as the models themselves.

Gemini 4 Pro Ghost-Testing on Arena

The clearest trace for Google is a checkpoint circulating on Arena under the name Gemini 3.8 Flash. When a user ran an infinite JSON stress test asking for a thousand planets with dense detail, the real Flash summarized and quit while this checkpoint kept producing non-repetitive, detailed records until it hit the browser's character limit and crashed the page. That stamina lines up with a leaked 256k output token limit attributed to Gemini 4 Pro, which strengthens the guess that this is the flagship checkpoint. The video also notes how to try it — sign in to Arena, open Battle mode and pick Code — and stresses that even at Arena's lower reasoning setting the outputs look surprisingly strong, implying launch-time capacity could be even higher.

The checkpoint's creative outputs form the core of the briefing. A smooth Mario Kart-style game with world generation, a Formula 1 car rendered in 3D with varied lighting and camera angles, a voxel-art temple complete with falling cherry-blossom leaves, fish in the river and a full day cycle, and a pure SVG PlayStation 5 — all are presented as unusually polished for Arena. The most talked-about example is an animated SVG of a Nintendo Switch where the controllers snap into the body and the screen boots with the logo; the host says Gemini 4 Pro clearly bested Opus 5 on that prompt. A ten-minute SVG generation that shows depth, multiple viewpoints, the mute-button glow and even a skeleton layer, plus a side-by-side pelican-riding-a-bicycle prompt where GPT-6 Astra still leads on visual depth, together make the case that the base reasoning level of the ghost checkpoint is already competitive with top models — fueling the Google comeback narrative while still leaving independent verification open.

GPT-6 Soul: Cutoff, Tiers, and Early Feel

On the OpenAI side, the briefing frames GPT-6 Soul as a near-term efficiency tier ahead of Dev Day. Its knowledge cutoff is described as April 30, the same date attributed to GPT-6 Astra, hinting at a shared lineage, with lower tiers possibly called Luna and Terra. Soul is said to have entered gray testing inside the ChatGPT web app, where early testers report richer formatting and noticeably better memory and personalization. A PS5 controller SVG generated with high reasoning effort in about three minutes and a 3D house interior coded in a 3GS-like environment are offered as signs that even the cheaper tier approaches Astra-like quality. The host also relays claims that Soul beat the unreleased Claude Opus 5.2 on a couple of prompts while using fewer tokens — but cautions that these are unverified early impressions, not benchmarks, and that pricing versus Astra and real speed will only be clear at launch.

Grok 4.7 Surfaces in Cloud Quotas

For Grok, the signal is its appearance inside Google Cloud quotas and system limits. Combined with Elon's earlier 'coming soon' note, the listing is read as a release being days away, even if the original ten-day window has already slipped. Early numbers cited in the video put Grok 4.7 at 6.4 on Colony Bench Mega, a clear jump over Gemini 3.8 Flash at 2.7 and GPT-5.6 Soul at 2.6, yet still well behind GLM 5.3, DeepSeek 4.1 Flash, GPT-6 Astra Pro and Claude Fable 5. On the CUDA side the picture is tighter: 9.5 on the GLM 5.2 fused test, just above Grok 4.6's 9.4, with small gains on other CUDA workloads. The host's take is balanced — a real step over 4.6, but often marginal — and suggests waiting for independent reruns before extrapolating.

Union Alpha Unmasked: Pareto 269 and the Harness Truth

The week's biggest mystery, Union Alpha, is resolved right after the main edit. Initially presented as a stealth model available for free via OpenCode, OpenRouter and Klein, with a 256k context window, multimodal support and agentic coding focus, it was credited with near-Astra and near-Opus 5 performance at roughly one-eighteenth of the expected cost. The follow-up reveal identifies it as Pareto 269 from Unbiased — not a single model but a system that routes work across several open and frontier models, in the spirit of the earlier Sukuna setup. Listed pricing is $2.50 per million input tokens and $0.25 for cached input, $7.50 for output, claimed to be less than a quarter of Astra's list price. That routing explains why it looked so close to frontier models on some tasks: the harness forwards work to those models when needed. My read is that Union Alpha should be judged as an orchestration product rather than a weight; the critical question is how transparent the routing logic is, not just the headline price.

Typesafe's Jev, developed with Dio, sits on a different axis entirely. Unlike classic large language models that generate text token by token, Jev takes unstructured input and a predefined set of possible outputs and returns structured decisions with probabilities, computing everything in parallel. Its training approach is called RLCD, reinforcement learning for calibrated decisions. The company's own benchmarks claim 20 to 200 times quicker while costing 40 to 400 times less across the decision workflows they tested, with pricing at four cents per million input tokens and free outputs — claims that will need independent validation. The key limitation is explicit: Jev does not generate normal text, so it is not a replacement for GPT-6 Astra or any Claude model; instead it is positioned as a companion that handles thousands of cheap, fast decisions around a frontier model — routing, classification, judging outputs, or picking the next agent action.

Dream RSI, Claude Projects, and Xiaomi's Public RL Run

Google DeepMind's Dream RSI is framed as a loop that is close to, but not yet, closed. Model weights do not directly improve; the agent that explores the policy learns from its own history of discovery attempts and gets better at searching for solutions over time. The video says the loop was demonstrated across algorithms, GPU kernel engineering and model development, suggesting the policy improvement generalizes beyond a single task. The caveat matters: this is recursive improvement of the search policy, not recursive self-improvement of weights, so while the label RSI is exciting, the result is a controlled step toward it rather than full autonomous weight-level evolution.

Claude's new Projects inside Claude Code is presented as a practical upgrade for on-the-go work: give Claude several tasks in one conversation and it routes them into parallel threads that keep running after you close your laptop, with shared memory and files staying in sync across threads. It is currently in limited beta for Pro and Max subscribers, with a wider rollout anticipated — potentially together with Opus 5.2. Xiaomi's Mimo 2.6, after nearly six months of silence, is described as mid-way through a massive reinforcement learning run that scales RL compute, multitask agentic environments and overall compute, generating roughly two billion tokens per training step across thousands of parallel rollouts. The team is said to be streaming the run publicly and to open-source more training details in the coming weeks, though the model itself is still one to two months away.

The briefing closes with a short robot demo where GPT-6 Astra powers dance steps, a symbolic reminder that large models now stretch from SVG craftsmanship to physical choreography. Taken together with the host's plugs for his newsletter, benchmark and vibe-coding platform, the picture is clear: Google is quietly auditioning a comeback via Arena, OpenAI is layering a family to lower cost, xAI is warming a release through cloud quotas, and Xiaomi is earning trust by opening the training log. My lesson from the week is to read it not as a single winner but as competing transparency strategies — whose ghost test is most honest, whose quota leaked earliest, whose log teaches most — that is where the real contest is playing out.