AI Weekly News: Mystery Models and Screen Watching

The strangest AI release this week has no public name, no confirmed maker, and a 1 million token context window. At the same time, Google is removing visible AI watermarks, while ChatGPT is testing an opt-in history of everything users do on their computers. The common thread is control. Models are getting harder to identify, generated media is getting harder to spot, and assistants are asking for wider access to private work.

Ox Alpha appeared with frontier-level results and no owner

Ox Alpha is available free through OpenRouter. Its listing claims multimodal input, zero data retention, a 1 million token context window, and capacity for 100 trillion tokens per day.

That mix is unusual. A million-token window can hold large codebases, long document collections, or months of structured records in one request. Zero data retention also makes the model more interesting for controlled testing, although teams still need to verify what that policy actually covers before sending sensitive data.

Early testers produced a 3D black hole simulation with light bending. Other outputs included a playable Minecraft-style clone with infinite terrain, mobs, and animated experience drops, reportedly at a low reasoning setting. On matched coding prompts, some testers placed its output near Claude Opus 5. These are informal comparisons, not controlled benchmarks, so they show capability rather than a reliable ranking.

Nobody has publicly claimed Ox Alpha. The main guesses point toward Z.ai, which develops the GLM model family, or Xiaomi. Those guesses come from tokenizer fingerprints and prior release patterns. Xiaomi previously tested MiMo 2 through OpenRouter under the Hunter Alpha and Healer Alpha names. Shared tokenizer behavior across Xiaomi, Qwen, and GLM models makes attribution harder, not easier.

The key insight: tokenizer similarities can reveal shared infrastructure, training ancestry, or compatible tooling, but they don't prove ownership. Until a lab claims the model, Z.ai and Xiaomi remain informed guesses.

Why it matters: Free access makes Ox Alpha worth testing now, but anonymous models require stricter data, licensing, and continuity checks.

GLM 5.3 Flash and Tencent HY4 point to a crowded China model cycle

A separate checkpoint called GLM 5.3 Flash was reportedly found in internal testing. Based on the available fingerprints, it isn't the same model as Ox Alpha. The signs include a video encoder, GLM 5 tokenizer behavior, and audio handling. If accurate, the GLM line is moving from text-focused systems toward native vision, audio, and other multimodal inputs.

The reported development path is interesting too. Z.ai appears to have expanded post-training and reinforcement learning on an existing base model instead of training an entirely new foundation model. Post-training can change how a model follows instructions, reasons through tasks, and uses tools. It typically costs less than rebuilding a base model, although it can't fully repair missing knowledge or weak representations from the original training run.

Tencent is also testing HY4 inside its own application. The interface reportedly labels it as an expert-level model above HY3 and DeepSeek. Tencent previously confirmed that HY4 was training with a larger parameter count. Claims that it matches Opus 5 remain unverified until Tencent publishes model details or independent evaluators get stable access.

The table shows why model names alone are becoming less useful. Access conditions, identity confidence, and architecture evidence now matter as much as benchmark claims.

Why it matters: Chinese labs are testing several high-capability models at once, but buyers need to separate accessible products from leaked checkpoints and marketing comparisons.

GPT-6 Astra and Fable 5.1 may arrive close together

OpenAI reportedly delayed GPT-6 Astra by another couple of weeks. The updated checkpoint is said to be in internal use through Codex, with work focused on reducing reward hacking.

Reward hacking happens when a model satisfies the scoring system without completing the intended goal. In coding, that might mean changing a test instead of fixing the implementation, or hiding an error instead of resolving it.

A September release is plausible based on the reported delay, but it isn't confirmed. Launch dates can move when internal evaluations expose safety, reliability, or tool-use problems.

Anthropic is reportedly holding a finished Fable 5.1 for a similar launch window. Anthropic has not confirmed the model name or schedule through the supplied official source, so both details remain rumors.

Back-to-back launches would make comparison difficult. Early benchmarks often mix different tool settings, reasoning budgets, system instructions, and context sizes. Enterprise teams get better evidence from internal tasks that reflect their own code and approval process.

There's a procurement angle here as well. A team that commits immediately after the first release may face a materially different option days later. Short evaluation contracts and provider-neutral application layers can reduce that risk.

Why it matters: The next model contest may be decided by agent reliability and reward-hacking resistance, not raw benchmark scores.

The Codex quota dispute exposes a trust problem

Some Codex Pro users report losing 15 to 20 percent of their weekly quota during a single session, sometimes in under an hour. OpenAI reportedly investigated and denied making a hidden quota reduction.

Many affected accounts had the Subto API in common, according to the research briefing. That correlation doesn't establish whether the API, account routing, usage display, or another dependency caused the depletion.

Users have a reason to remain skeptical. An earlier confirmed cache-hit optimization caused limits to drain faster before OpenAI rolled it back. A cache hit normally reuses previously processed content and should reduce compute. If the quota system charges cached tokens incorrectly, efficient workloads can appear more expensive than uncached ones.

Teams can protect themselves by recording three values for each agent run: wall-clock duration, reported token use, and remaining quota. A screenshot or export before and after a controlled test is more useful than a general complaint.

The same measurement approach applies across hosted models, including options reached through OpenRouter. Provider dashboards should be treated as one source of evidence, not the complete usage record.

Why it matters: Unclear quota accounting turns an engineering limit into a trust issue, especially when autonomous agents can consume capacity without obvious user actions.

Gemini is removing visible watermarks but keeping SynthID

Google is removing visible watermarks from generated images, video, and music. Google describes the visible mark as a toggle, although many users reportedly see it disabled without taking action.

The invisible SynthID watermark remains embedded. When supported media is uploaded again, Gemini can inspect it for a machine-readable signal that identifies AI generation.

That creates a split between human visibility and machine detection. Clean output is easier to publish and edit, but viewers can no longer rely on a visible badge.

Organizations publishing generated media can keep the original file, prompt record, generation date, model name, and approval status together. That internal provenance record remains useful when a platform can't detect the watermark.

The change also shifts moderation power toward the model provider. Google controls the detector, while publishers and viewers may not have an independent way to inspect the same signal.

Why it matters: Cleaner exports help creators, but invisible provenance concentrates verification inside the platform that created the content.

Google is using study tools and free storage to build long-term habits

Eligible US students reportedly receive one year of Google AI Pro at no cost. The plan normally costs $20 per month and includes 5 TB of storage, while students in other regions may receive AI Plus instead.

Google's new study and learn notebooks begin with a diagnostic quiz. They generate a study guide from the results, then test the student again to find topics that remain weak. That's more useful than requesting a generic summary. The first quiz creates a baseline, while later quizzes show whether the student can retrieve information without seeing the source material.

Gemini 3.7 Flash is also reported to run in the Gemini app for Pro and Ultra users, AI Mode in Google Search, and Sheets. According to the briefing, AI Mode and Gemini notebooks do not count against normal Gemini usage limits.

That creates a practical routing choice. General research can go through AI Mode, study material can stay in notebooks, and higher-cost work can remain in the main Gemini interface.

Gemini notebooks are also expected in the Search sidebar. Gemini Live can start Deep Research that continues while a phone is locked, which moves longer research tasks away from continuous foreground use.

Students still need to check eligibility, regional terms, renewal dates, and storage consequences before moving files. A free year can create migration work later if the account returns to a smaller storage allowance.

Why it matters: Google's strongest student offer combines tutoring, storage, and unmetered surfaces, which can shape where users keep both their files and learning history.

ChatGPT can open Drive files inside the conversation

ChatGPT's web application can connect to Google Drive and display Docs, Sheets, and PDFs in a Library section. Typing @ followed by a document name reportedly opens that file beside the conversation.

The side-by-side design reduces context switching. A user can discuss a document while keeping the source visible, rather than copying its contents into a new chat.

Editing is less predictable. Testing described in the research briefing produced one stalled session and another case where ChatGPT exported a Word copy instead of editing the original Google Doc. That difference can create duplicate files and conflicting versions.

Teams can test the connector with a disposable document before allowing it to modify shared operational records. Access scope matters too. Connecting an entire Drive may expose more content than a task requires, especially when folders mix personal, customer, legal, and project files.

A safer rollout uses a dedicated folder with a few non-sensitive files. The team can then verify read access, write behavior, version history, export format, and revocation before expanding the scope.

For broader coverage of agent permissions and code safety, see AI Developer Trends: Agents, Local Models and Code Safety.

Why it matters: Document-aware chat can save time, but unreliable write behavior and broad file access require a controlled test before regular use.

ChatGPT computer history trades forgotten context for total activity visibility

ChatGPT's desktop application is testing an opt-in computer history setting. It records a timeline of applications, work, and timestamps so the assistant can answer questions about earlier activity.

A user might ask what was open before the last break. The system could also observe a repeated workflow and turn it into a reusable skill.

The feature is off by default, according to the research briefing. That default makes sense because an activity timeline can expose customer names, private messages, health information, internal systems, and personal browsing.

The real concern is not a single screenshot. A continuous activity record can reveal relationships between events: which document followed a meeting, which account opened after an alert, or which project consumed most of a day.

A useful memory system needs selective capture. Application exclusions, pause controls, local processing, short retention, and searchable deletion would make the tradeoff easier to assess.

The privacy race is moving elsewhere too. Anthropic is reportedly building a system that lets zero-data-retention enterprise customers keep safety-monitoring data themselves, rather than having Anthropic store prompts for up to 30 days. Anthropic has not published the detailed system through the supplied source.

OpenAI is reportedly working on a similar private safety-processing model. Other unconfirmed tests include an image system called Mona Lisa 1 and an Apple Messages plugin for ChatGPT on macOS.

Why it matters: Computer history could become a powerful work memory, but it also creates one of the most complete behavioral records an assistant can collect.

Claude Academy and private safety processing broaden the enterprise contest

Anthropic launched Claude Academy as a free learning hub, according to the research briefing. Its courses range from introductory material to daily-work guidance.

Training portals matter because model capability does not guarantee consistent use. A shared course can give teams common terms for prompting, document handling, verification, and safe tool access.

Anthropic's reported customer-held safety data system tackles a harder enterprise problem. Zero-data-retention customers may reject a service if safety monitoring still requires the provider to retain prompts temporarily.

Customer-controlled storage offers one possible compromise. The provider can request safety evidence when required, while the customer keeps direct custody of the underlying records.

The tradeoff is operational responsibility. Customers may need to secure logs, manage retention, answer legal requests, and prove that records have not been altered.

Why it matters: Enterprise model adoption increasingly depends on training, custody, and audit controls rather than benchmark leadership alone.

Worth watching

• Whether Ox Alpha's developer claims the model or changes its free access terms.
• Independent confirmation of GLM 5.3 Flash's video, audio, and tokenizer fingerprints.
• Official dates and model cards for GPT-6 Astra and Anthropic Fable 5.1.
• Clear documentation for computer-history retention, exclusions, and deletion.
• Evidence that invisible SynthID survives common publishing and editing workflows.

Hands-On Steps

Start here

Run five existing work prompts through Ox Alpha on OpenRouter, then score accuracy, latency, formatting, and instruction compliance from 1 to 5.

Quick wins

• Create a test Google Drive folder with three disposable files, then verify ChatGPT's read, edit, export, and revocation behavior.
• Record Codex quota before and after one 30-minute session, including token counts, retries, and tool calls.
• Check whether Gemini exports visible watermarks on three media types, while keeping the originals for later SynthID testing.

Deep dive

• Build a one-page AI data-access register covering connected folders, retention periods, model providers, account owners, and deletion controls.
• Compare Ox Alpha with one current production model using ten real tasks and two blind reviewers, rather than public benchmark scores alone.

Useful Resources

• OpenRouter - Access point for Ox Alpha and other hosted models.
• Z.ai - Official home of the GLM model family.
• Anthropic - Official source for Claude products, enterprise controls, and company announcements.
• AI Developer Trends: Agents, Local Models and Code Safety - Related coverage of agent permissions and local-model choices.
• Qwen3.8-27B: Local AI Performance, Context and Quants - A practical look at local model context and quantization tradeoffs.

The Bottom Line

Ox Alpha is the week's most interesting model because it combines free access, a huge context window, and unknown ownership. Its origin remains speculation, so results should be tested without assuming stable access, licensing, or provider accountability.

The quieter product changes may have longer consequences. Invisible watermarks reduce human visibility, Drive connectors widen document access, and computer history asks users to exchange a detailed activity record for better memory.

Next week, the useful signals will be ownership disclosures, retention documentation, reproducible evaluations, and clear launch dates. Model intelligence matters, but control over files, logs, provenance, and billing is becoming the more practical dividing line.