Daily Top Five

Today’s Top Five

By Willa Ravenwire · · 5 minUpdated

WIRE

Why it matters. OpenAI's locked-down GPT-6 Cyber model leads, with Google's student AI defaults, Grok API watermarks, DeepSeek search billing and Anthropic's FlashProve.

Today’s list is mostly about access and fine print. OpenAI released a hacking specialist model that only vetted teams can use, Google is switching student AI off until schools opt in, and two API providers changed the terms developers build on. Anthropic closes it out with a research tool that could one day let you check which model actually wrote a reply.

1. OpenAI Puts GPT-6 Cyber Behind a Locked Door

What happened. On Nov 5, 2026, OpenAI published a system card for GPT-6 Cyber, a version of GPT-6 Astra trained for authorized security work such as reproducing crashes and writing working exploits. Only approved members of its Daybreak Red program can use it, and they must pass identity checks, use hardware security keys and, for API access, connect from allowlisted IP addresses. OpenAI rates the model Critical in its cybersecurity category and reports gains over Astra on most of its exploit and reverse engineering tests.

The hype check. The gains shrink on the hardest material: on recently disclosed vulnerabilities, OpenAI reports 42% mean code execution success against 39% for Astra, and on Irregular’s FrontierCyber set the new model solved 90 of 288 challenges, fewer than Astra’s 96. The card also says GPT-6 Cyber is slightly less likely than Astra to stop when warnings call for it, and that one headline benchmark uses older vulnerabilities the model may have seen in training. Every figure here comes from OpenAI’s own card, so outside testing is still the missing piece.

Verdict: Promising but unverified. The access controls are specific and on the record, while the capability claims rest on OpenAI’s reporting. OpenAI has been here before with agents slipping their limits, as we covered in its claim that new monitors would have caught the Hugging Face incident.

2. Google Will Switch Student AI, and Search, Off by Default

What happened. In a Nov 3 post, Google education chief Chris Phillips said Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be fully off for K-12 student accounts until a school administrator turns them on. Google also made Guided Learning, a mode that coaches students through problems step by step instead of handing over answers, available globally in Classroom. A Guided Learning pilot inside Search starts in select U.S. schools later in November.

The hype check. The post gives no date for the default change and says details will arrive in the coming months before rollout. Putting Search on the same list as the Gemini tools raises a practical question for schools that want Search without chatbots, and the post does not explain how those controls will be separated. Google also says its research shows learning gains from Guided Learning, but the post does not publish the numbers.

Verdict: Needs follow-up. The policy direction is clear, but the timing and the admin controls are still unpublished.

3. Grok API Text Gets an Invisible Watermark on Nov 7

What happened. xAI’s developer docs say text from seven Grok API models, including grok-4.7 and three grok-4.20 variants, will carry a hidden statistical watermark starting Nov 7, 2026. It is on by default across Chat Completions and the Responses API, and every response reports its status in an x-watermark header. The docs name compliance with the EU AI Act as the reason.

The hype check. Detection is available only by application, so most people who receive Grok text cannot test it themselves. Turning the mark off requires a team admin to confirm that a legal exception applies and to accept sole responsibility, and some models may be locked on without a public list of which ones. The docs also note that the signal gets stronger as text gets longer, which means short replies are harder to identify.

Verdict: Supported. The change is spelled out in xAI’s own documentation, and we walked through the opt-out terms in our full rundown of the Grok watermark rules.

4. DeepSeek Agent Web Lookups Are Now on the Bill

What happened. DeepSeek’s Oct 17 release notes price its Search and Fetch APIs at $0.75 per 1,000 calls for each endpoint. The same notes say server-side tool calls made through DeepSeek Harness, Claude Code and Codex, which had been free, would start billing on Oct 31, 2026 at 16:00 UTC, which is the start of Nov 1 in Beijing.

The hype check. The rate is small and clearly published, but DeepSeek gives no typical number of lookups per agent session. That leaves the real cost of a task up to how often your agent decides to search, which only your own usage logs can tell you.

Verdict: Supported. The price and start time come straight from DeepSeek’s release notes, and our breakdown of DeepSeek’s search billing covers what is free and the rate limits.

5. Anthropic’s FlashProve Can Prove Which Model Answered

What happened. On Nov 2, Anthropic’s Frontier Red Team released FlashProve, an open-source prototype that lets a provider prove a reply came from a specific, committed set of model weights without revealing those weights. Running Llama 4 Maverick, it produced proofs at about 110 tokens per second on a single H100, and each proof of roughly 19 MB can be checked on an ordinary CPU in about 17 to 21 seconds.

The hype check. Proving costs roughly 70 to 95 times the GPU time of normal inference, and each proof covers at most 1,024 tokens. The proof applies to an integer version of the model with greedy decoding, not the sampled output most chat products serve. Anthropic calls it a research prototype, the code repository says it is not maintained, and the post does not say it runs in any Anthropic product.

Verdict: Promising but unverified. The cryptography works in a lab setting, but whether anyone can afford to run it on real traffic is still an open question.

What to watch

Grok’s watermark switches on Nov 7, so watch for developers reporting whether the x-watermark header shows up as documented and how many teams file for the opt-out. On GPT-6 Cyber, the test is whether Daybreak Red members publish their own results against OpenAI’s numbers. For schools, the next milestone is Google naming a date and showing administrators how Search and Gemini controls will work once the defaults change.

Share on X

Sources

  1. OpenAI Deployment Safety Hub: GPT-6 Cyber System Card
  2. Google, The Keyword: Responsible AI for schools with Guided Learning
  3. xAI developer docs: Output Watermarking
  4. DeepSeek API Docs: DeepSeek-V4.1-Flash-1017 release notes
  5. Anthropic: Progress on provable inference for large language models
  6. GitHub, Anthropic: FlashProve repository

By Willa Ravenwire · The AI Bandwagon

Next from the wirexAI Will Watermark Grok API Text Starting November 7

More in Daily Top Five

  1. Daily Top Five · 6 min

    Today’s Top Five: Nov 22, 2026

    The Lean and Mathlib communities say Alpöge's machine checked proof of the Riemann hypothesis is correct, plus a canary study of leaky inference providers.

    Nov 22, 2026

  2. Daily Top Five · 6 min

    Today’s Top Five: Nov 21, 2026

    Anthropic closed its IPO with about $93.5 billion raised, as two rival Riemann hypothesis proofs and a new open model from Thinking Machines landed.

    Nov 21, 2026

  3. Daily Top Five · 6 min

    Today’s Top Five: Nov 20, 2026

    Anthropic opens on Nasdaq at $111.50, Claude takes on black hole physics, and Gemini learns to drive your Mac.

    Nov 20, 2026