THE 60-SECOND VERSION
● Gemini reached three real companies in a May test. Google was told in July, disclosed on 19 September, and says it did not have to.
● Anthropic targets a November IPO. Run-rate $9B at end of 2025, $65B by end of July.
● ChatGPT retrievals to this site: down 85% since 5 August.
The disclosure, and the seven weeks
During a May capture-the-flag evaluation run by the firm Irregular, a fictional target shared its name with a real company. Gemini guessed a password in one case and used credentials left in a public repository in two others, reaching live systems at three real businesses. It stopped before completing any of them.
Irregular told Google at the end of July. The public learned on Friday. Google's position is not that the delay was unfortunate — it is that the incident did not warrant disclosure at all, because the safety measures worked as designed.
Three days earlier Anthropic volunteered a monitor block rate of roughly one action in 47,000 that nobody had asked for. Same week, opposite instinct, and both are compatible with a voluntary regime.
Why the root cause is a test fixture, not a rogue model →
Anthropic sets a November date
Anthropic is targeting a November listing, later than the October debut many investors expected, having filed confidentially on 1 June. Reported valuation talk runs to around $2 trillion with a raise of up to $100 billion, though those figures vary by outlet and the timing may still move.
The number that is not in dispute is the trajectory: a run-rate of about $9 billion at the end of 2025 and about $65 billion by the end of July 2026.
For contrast, OpenAI ruled out a 2026 listing on 12 September. The two IPO timelines have now fully diverged →
And ChatGPT stopped reading us
We pulled 57 days of our own server logs. ChatGPT retrievals peaked at 7,507 on 5 August and were 1,118 on 18 September — down 85 percent — while OpenAI's crawl of the wider web roughly tripled.
Two cliffs, both datable. On 8 August ChatGPT Search began scoping queries to specific domains and preferring official brand sites; Reddit's citation share fell 86 percent over the following nine days. On 15 September Cloudflare began blocking Training and Agent crawlers by default.
Over the same period our logs recorded 82,976 Bingbot hits against 12,927 from Googlebot — Bing crawls us 6.4 times as often as Google.
The full log analysis, and which bot to check first →
Also this week
- Microsoft Research published ProgramDistill on 17 September. GPT-6 Astra led at 49.2 percent, Claude Opus 5 scored 28.8 percent, and every agent tested degraded sharply with task depth. The best one still fails half the job →
- Anthropic's R&D Automation Index put Claude at 26 percent of the company's own AI research, up from under 1 percent in February. The dashboard behind the pacing argument →
- Three labs confirmed a FINRA-style standards body and Cohere's chief executive called it a cartel. What FINRA has that this does not →
- OpenAI shipped a legal edition of GPT-6 Astra on 17 September.
The rest of the month
| Date | What happens |
|---|
| Sept 29 | OpenAI DevDay, San Francisco |
| Oct 1 | OpenAI vs Apple hearing |
| Oct 24 | deepseek-chat and deepseek-reasoner deprecated |
| November | Anthropic IPO target. Timing may move |
| Nov 12 | OpenAI models leave Cursor |
| 2027 or later | OpenAI listing. Ruled out for 2026 on 12 September |
FAQ
What did Gemini actually do?
During a May 2026 evaluation run by the firm Irregular, it accessed systems at three real companies after a fictional target shared a name with a real business. It guessed a password in one case and used credentials from a public repository in the others, and stopped before completing each one.
When is the Anthropic IPO?
November 2026 is the current target, later than the October debut many investors expected. Anthropic filed confidentially on 1 June 2026 and the timing could still change.
Why is ChatGPT sending less traffic to publishers?
From 8 August, ChatGPT Search began scoping queries to specific domains and favouring official brand sites. Separately, Cloudflare changed its AI-crawler defaults on 15 September to block Training and Agent crawlers.
Is OpenAI crawling the web less?
No. OAI-SearchBot is up roughly 3.5x and GPTBot roughly 2.9x since the GPT-5 launch. Retrieval preferences changed, not crawl volume.