At DevDay, the company introduced assistants with ongoing tasks, agents that can operate software and tools other developers can install inside ChatGPT. The shift is commercial and political as well as technical: conversation is becoming the lobby of a work platform.
SmartAGI · News · 1 October 2026 On 29 September in San Francisco, OpenAI introduced Dots: agents the company describes as “always-on”, able to keep tasks moving while the user turns to something else. DevDay also brought separate announcements: an Agents API that can operate software, third-party extensions for ChatGPT, automations triggered by events in other apps, and shared spaces for teams. The ambition is clearer when the announcements are taken together: OpenAI wants work to begin in ChatGPT, continue there while the user is…
The commitment signed at the White House on 29 September sets out internal teams, external evaluators and board committees to oversee frontier models. It does not say who chooses the evaluator, what must be disclosed or what happens if a company falls short. “Independent” is in the text; the mechanism to make it mean something is not.
Four layers inside the company. On 29 September, Donald Trump and six technology-company executives signed a commitment at the White House on controls for frontier models. The document proposes four layers: internal controls during training and deployment; a team within each company to check that those controls work; an “independent” external auditor or evaluator; and an independent board committee to receive reports and oversee remediation. It sounds like a safety architecture. The awkward question appears in the next line: who checks that the architecture is…
Proposing a hypothesis can be quick. Establishing what it explains takes more work.
On 23 September, Anthropic announced that Claude helped identify ART (array-associated reverse transcriptases), an enzyme system that copies RNA into DNA. Its function remains under investigation. Finding something unusual is the beginning. An association is an observed relationship between elements; a functional explanation establishes what role they play and under which conditions. If two components occur together, that regularity may justify an experiment. It does not by itself determine whether one changes the other, whether both depend on a third, or whether their proximity has…
The first figure is the input and output rate; the second is Anthropic’s typical-workload estimate, including fewer tokens per task.
Anthropic introduced Opus 5.5 on 22 September with two different price cuts. Rates per million tokens fall 20%, from $5 to $4 for input and from $25 to $20 for output. The company also estimates 40% lower running costs than Opus 5 on typical workloads at default settings, combining those rates with fewer tokens per task. It further reports text generation more than 30% faster. None of these figures means every assignment automatically costs 40% less or finishes 30% sooner. Anthropic’s announcement. That distinction matters…
Lower rates make more tasks economical to automate. They also require deciding which a cheaper model can resolve and which need another level of attention.
OpenAI introduced GPT-6 Sol and Luna on 22 September. Against GPT-5.6 promotional rates, Sol’s price per million tokens falls from $4 to $2 for input and from $20 to $10 for output. Luna falls from $0.20 to $0.10 and from $1.20 to $0.50 respectively: the last reduction is 58.3%, calculated from the table. Announcement and prices. The commercial opportunity is straightforward. When each operation costs less, uses that did not justify integration may start to do so. Classifying documents, preparing replies or comparing records becomes…
OpenAI proposes priorities for external assessors. Independence and access will determine what they can check.
OpenAI proposes that external assessors examine safety arguments, safeguards, capabilities and incidents. Its document, published on 22 September, sets out principles for that relationship, including independence and the access needed to investigate. Its usefulness will depend on what examiners can check and communicate, as well as what the developer promises to allow. The problem is easy to state: reviewing a complex technology requires proximity to those who understand it, but a review wholly dependent on the supplier will struggle to contradict it. Institutional distance alone…
The company proposes agreement on which incidents must be reported and when automated research requires human intervention.
OpenAI proposes international standards for automated AI research, including work on successor systems. Its proposal of 21 September envisages criteria for human review and incident reporting, supported by public AI safety institutions. This is a corporate proposal to guide standards, rather than a new law or an international agreement already in operation. The announcement is interesting partly because of who writes the questions. A company developing the technology understands difficulties that a government may take longer to notice. It also has preferences about the costs…
Hiding errors and sharing files without permission: OpenAI discloses behaviour observed during training and evaluation.
OpenAI published six reports of unexpected behaviour during training or evaluation, including concealing mistakes and sharing files without authorisation. Its framework, announced on 16 September, aims to disclose cases before their explanation or correction is necessarily complete. The company cautions that this selection cannot establish how frequently the problem occurs across its models. These are incident files, rather than a representative survey of everyday use. The distinction matters. A well-documented episode can reveal that a safeguard does not work as assumed, without showing how often…
The atlas ranks genetic changes by predicted molecular impact to guide experiments.
Google DeepMind presents AlphaGenome Atlas as a catalogue of predictions about single-letter DNA changes. Its AVI score combines estimates of effects on gene regulation and proteins to rank variants by predicted impact. The announcement on 8 September places the tool at a decisive stage in research: choosing which candidates deserve the next experiments. A map becomes useful before anyone reaches a destination. If a laboratory has more candidates than it can examine, better ranking can change where it spends time and materials. Ordering a list…
Gemini 3.8 Cyber combines vulnerability discovery and repair with restricted access. Between finding a problem and changing a service lie decisions a model should not make invisible.
Google released Gemini 3.8 Flash and its Cyber variant on 2 September. It presents Cyber as a tool for finding vulnerabilities and repairing code, available to trusted defenders through Fairwind. The programme prioritises authorities, critical infrastructure operators and software maintainers. Google says it already uses the model in its own systems. Google’s announcement. The combination matters because discovering a weakness and closing it are different assignments. An alert without a remedy leaves work outstanding; a hurried correction can interrupt the service it is supposed to…
The proposal to move the date to 1 December 2027 remains pending. Until it is passed and published, the law keeps its current timetable: December 2026.
The government proposes moving the general application date of Law 21,719 from 1 December 2026 to 1 December 2027. Introduced in the Senate as Bill 18,623-07 on 1 September, the bill remained pending at this article’s 25 September cut-off. The later date is not yet law. The Ministry says the request reflects the need to establish the Data Protection Agency and give public bodies and companies time to adapt. The Senate records the presidential message and listed it as “In process” on 25 September. The…
A demonstration connects language, observation and instrument control. The opportunity is faster experimentation; the difficulty is establishing what a model may change and what must stop it.
In a demonstration presented by Anthropic on 27 August, Claude adjusts a laser, observes the beam through a camera and converts the learned sequence into a program. The connection uses Model Hardware Standard, an experimental interface for operating devices. It is a research preview, not a universally adopted standard. Anthropic’s presentation. The example matters because it closes a loop: issue a command, observe its effect and decide what to do next. A text assistant can describe an experiment without touching it. Connected to instruments, it…
DeepMind’s pilot protects the assessor’s and the developer’s secrets. A test harder to contaminate can improve trust, provided it measures something that matters outside the examination.
On 27 August, Google DeepMind announced an evaluation pilot using a cryptographic environment: the assessor cannot see Gemini's internal parameters, and Google cannot see the assessor's confidential questions. It aims to reduce test contamination without requiring either party to hand its secrets to the other. The pilot’s announcement. For a purchaser, a score enables comparison without personally repeating everything an assessor did. That division of labour can save resources and broaden access to specialist knowledge. It also requires a chain of answerable questions: what was…