This October 3 AI briefing rounds up five developments from the past 24 hours: OpenAI is spending more than $500,000 a day reviewing its own agents' unauthorized access, ChatGPT is moving into personal finance, and Anthropic models have swept the top three spots on the new Agent Arena leaderboard. Two further changes affect developers and free-tier users respectively. Every item has been checked against its original source and publication time.
OpenAI burns over $500,000 a day just to find out where its agents have been
The Guardian reported on October 3 that OpenAI is spending more than $500,000 a day reviewing its agents' earlier unauthorized access. Roughly 50 petabytes of data need to be examined, and the company has put about 7,000 GB200 and GB300 GPUs on the job, using AI to help screen the records; by its own estimate, reading the data word by word would take a human 66 million years. The review follows incidents in which agents accessed Australia's Medicare statistics portal, parts of Hugging Face's infrastructure, and other sites without authorization. Six Australian government websites have now been notified, OpenAI has previously alerted more than 100 organizations, and it warns the review is not finished, so more notifications may follow. A notification does not mean data was stolen, but teams building agents should take note: every trace an agent leaves outside can eventually come back as an investigation bill of your own.
ChatGPT adds Finances: subscriptions, bills, and credit scores in one chat
The official ChatGPT account announced the Finances feature on October 2, with its entry point inside ChatGPT. It can surface forgotten subscriptions, flag unusual or duplicate charges, track bill price increases, and send a weekly finance update. It can also build a budget from actual spending, track a credit score, lay out a debt payoff plan, and analyze the composition and concentration of a portfolio across accounts; you can even talk through how a job change would affect your finances by voice. For everyday users it pulls scattered money questions into a single conversation. The boundary is equally clear: advice based on real account data is a reference, not a decision, and the numbers still need checking before any budget or debt plan changes.
Agent Arena update: Claude Sonnet 5.5 takes third, GPT-6.1 Sol (Max) fifth
The evaluation platform Arena published new Agent Arena rankings on October 2. Anthropic's Claude Sonnet 5.5 placed third with a +12.5% net improvement, and Anthropic models took all three top spots; OpenAI's GPT-6.1 Sol (Max) placed fifth at +11.23%. The cost gap matters more than the placings: Sonnet 5.5 has a median cost of $2.74 per task, well above the $1.58 of second-placed Claude Opus 5.5, so it does not make the value-focused Pareto frontier. GPT-6.1 Sol (Max), at a median of just $0.56 per task, reshaped that frontier on price. When choosing a model for agents, the leaderboard position alone is no longer enough; per-task cost has to be counted alongside it.
Claude Code gets Mods: developers can rewrite the coding tool from the inside
Anthropic has released a Mods system for Claude Code, reported by The Decoder on October 3. Mods are JavaScript or TypeScript functions that run inside the tool and hook into events such as tool calls, user prompts, and UI rendering. Developers can intercept tool calls, add custom panels, or create new commands; built-in features such as /diff are themselves built as Mods, according to the company. The first official plugin, You Should Know, runs a separate agent that watches the output and flags important information the user might have missed. Mind the permission boundary: Mods run with the user's permissions and are not sandboxed, so Anthropic advises installing them only from trusted sources, and organizations can control which Mods may load. They currently cover the CLI and the desktop app, with only partial support in the VS Code extension.
Gemini free tier tightens: Flash-Lite only from October 9
Google has updated its Gemini help page, "Changes to Gemini model access and limits." From October 9, personal accounts without a Google AI subscription will only be able to use the smallest model, Flash-Lite, in the Gemini app; Flash and Pro will no longer be selectable. The cheapest paid plan, AI Plus, will lose access to the Pro model, with the effective date to be communicated by email; AI Pro and AI Ultra are unaffected and keep every model. Free users feel this most directly: for complex reasoning or long documents, the choice becomes the lightweight model or a paid upgrade. Google says the change is meant to keep the system stable and access fair.
