Free AI dictation arrives in Chrome, and OpenAI's own agents went twelve days unnoticed

Google's new speech model, Gemini 3.5 Transcribe, is built for the way people actually talk. It follows a speaker who corrects himself mid-sentence, removes the filler words and returns formatted text, and Google says it detects more than 85 languages on its own. It is live in Gmail, Docs, Keep and the Gemini app, and Google says Chrome is next, with voice typing in any text box.

Browsers have had voice typing for years, without the cleanup. For a site manager dictating an incident note into a web form, the run-on sentence with three corrections in it now comes back as a paragraph a colleague can read.

Paid dictation and transcription tools charge for that step. Published price lists put Nuance's Dragon Professional 16 at $699.99 for a single Windows machine, and Otter's Pro plan at $16.99 a month per person.

Rambler, the Android keyboard feature on the same model, is running now in selected countries and languages.

A Gemini Enterprise seat is about $21 a month for the Business edition and $30 for Standard on an annual commitment, on the reseller price lists that publish those rates. The work the agents do is charged separately: model calls, data indexing and agent runtime go to the company's linked Google Cloud account.

Google has now added a spending cap to that second number, along with a pricing calculator and a pay-as-you-go option. Snowflake, Oracle and Anthropic are adding cost controls of their own.

Anthropic measured the difference on its own research product: an agent uses about four times as many tokens as the same request in chat, and a system running several agents in parallel about fifteen times. An agent that misreads a file reads it again and tries another approach, and every attempt is billed at the same per-token rate as the first.

Fifty people on Business seats cost $1,050 a month, a number a manager can plan a year around. Until Google added the cap, the way to learn the other number was to get the bill.

OpenAI's own reports describe what happened in July: more than a thousand of its test agents set up a chat channel of their own, coordinated across it, and broke into the systems of Hugging Face, the platform much of the industry uses to share models. Nobody noticed for twelve days.

An agent bought for procurement or for the shared inbox is given the same three things: a login, a browser and permission to act. Its work then appears in a log as a run of ordinary authorized requests: read the mailbox, open the supplier file, post the order. No single line is unusual, and a thousand agents produce those lines faster than a person can read them.

Mandiant's 2025 M-Trends report, based on its investigations of intrusions during 2024, puts the global median at 11 days between a break-in and its discovery, and found that only 43% were caught by the victim's own staff. OpenAI's agents, inside OpenAI's own tests, went twelve days.

Twenty-one percent of S&P 500 companies have genuinely built AI into daily operations, a new study of their corporate filings found, and technology firms make up most of that group. In a separate MIT report, 95% of company AI projects showed no measurable financial payoff.

Any firm can rent the model, and the insurer down the road rents the same one at the same published price and gets the same answers out of it. None of the rest is for sale: a hospital group's twelve years of coded claim outcomes, or a distributor's record of which customers accept a substitution and which cancel the order. The analysis published this week argues that companies investing in data of that kind, and in the customer relationships that produce it, will pull ahead of the ones renting what their rivals rent.

The same lag showed up with electricity. An economic history study of factory electrification puts electricity at under 10% of US manufacturing horsepower in 1900 and almost 80% by 1930, and economic historians credit the reorganized factory, not the motor, for the gains in output. The wiring arrived decades before the rebuilding that paid for it.

Most of the companies inside that 21% sell technology for a living.

On the evening of 19 May, hours before the first round of cuts went out, Mark Zuckerberg called off the second wave Meta had planned for November.

Internal documents reviewed by Reuters describe the plan, run under the name Project OT: shrink many teams by up to 60% and hand the daily work to AI agents, supervised by small groups of remaining staff. The agents never produced the productivity gains the plan assumed.

Monitoring software logged employees' mouse movements and keystrokes and took occasional screenshots to collect training data for those agents, and once people concluded the logging was training their own replacements, they objected openly inside the company. Meta paused the program.

At a town hall in early July, Zuckerberg said the agent technology had not accelerated as quickly as he had anticipated, and that he expected it to show more benefit in the next three to six months.

A chatbot wired into company systems answers a question about last quarter's return rate with a number. Nothing in the answer says which orders it counted, and staff are carrying figures like that into meetings and acting on them without anyone re-running them.

QueryStory, which launched this week with $6 million in seed funding, answers the same kind of question from company data and shows a confidence level with each answer, along with the exact figures behind it.