Xiaoliu BOT
中文EN

X Platform September 30 AI Brief | GPT-6.1 Sol and Ultrafast Tiers Released, Dots Cloud Agent Launched, DeepSeek Open-Sources Ascend Components

GPT-6.1 Sol and Ultrafast Tiers Differentiate Performance, Pricing, and Subscription Levels

At DevDay, OpenAI introduced the GPT-6.1 Sol model and the Ultrafast speed tier for Astra, with a focus on making high-capability models more suitable for high-frequency, long-duration work, while setting a higher price point for extremely high-speed usage. The company states that Sol approaches Astra’s capability level at about one-fifth the price, with a 95% discount for cached reads; Sol is now available to multiple categories of paying users and via API. Ultrafast offers up to approximately 8x faster generation speed for Astra and up to about 6x faster via API; Ultrafast for Codex and ChatGPT Work is tied to the new Pro 500 plan, with the official usage cap for the plan being 25 times that of the Plus tier.

Sources:

Dots Introduces Persistent Cloud Agents to Personal and Team Workflows

OpenAI released Dots: persistent agents powered by GPT-6 Astra, featuring independent cloud computers and browsers, capable of connecting to user-selected applications and continuously advancing tasks in the background. The company claims coverage of over 4,000 applications, with the first Dot available to eligible Pro, Business Premium, and Enterprise plans; users can configure which actions can be performed autonomously, when prior approval is required, and which actions are prohibited. OpenAI executive Tibo stated that initially, one Dot per person is being made available, with team configurations to follow later; the billing boundaries for a standard Dot working autonomously differ from those for invoking Codex to execute tasks, and it should not be broadly interpreted that all autonomous work incurs no usage limits.

Sources:

ChatGPT Space and Plugin Extensions Target Collaboration and App Distribution

DevDay extends ChatGPT from a conversational interface to collaborative spaces and native applications: the Space page can contain notes, charts, checklists, and other content, and allows team members to update it alongside ChatGPT, Codex, or a Dot; the OpenAI developer account also stated that plugin extensions can provide interactive panels, file viewers, and sidebars within ChatGPT, improve moderation feedback, and connecting application events can trigger automation via the proposed MCP Events specification. OpenAI product lead Tibo stated the platform has 1.2 billion weekly users and mentioned developers can build native applications embedded within ChatGPT; these are official descriptions of distribution capabilities and do not indicate every plugin is already widely available.

Sources:

Codex Updates Cover Cloud Execution, CLI, and Security Review

OpenAI announced that the Codex cloud environment is beginning to roll out to Plus, Pro, Business, and Enterprise users, allowing tasks to run in the cloud and be continued from a phone or another computer; team editions can also share environments. CLI updates add a full-screen interface, parallel job management, managed worktrees, Mermaid/LaTeX rendering, and collapsible diffs, among other capabilities. Additionally, Codex Security Cloud continuously scans GitHub repositories, organizes and deduplicates findings, and prepares fixes for human review. Taken together, the updates simultaneously improve long-running task continuation, in-terminal collaboration, and cloud security checks, rather than being a single interface overhaul.

Sources:

DeepSeek Open-Sources Multiple Training and Inference Foundation Components Adapted for Ascend

According to a compilation by MaxForAI, DeepSeek has open-sourced components for Huawei Ascend including TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, covering underlying aspects such as operator writing, matrix computation, MoE communication, and attention mechanisms; one of the goals of TileLang is to adapt to different hardware through a higher-level interface. Test data cited by the blogger states that DeepGEMM Ascend achieved 99.8% of the hardware’s theoretical upper limit in some Dense GEMM tests; Gorden Sun also referenced the open-sourcing and performance data of DeepGEMM Ascend. The performance figures here belong to the test results cited in the post and should not be extrapolated to all models or workloads.

Sources:

Apex Math Demonstrates Research Workflow of “Mathematician Sets Problem, AI Assists in Proof, Human Reviews”

Gorden Sun introduced that Apex Intelligence launched Apex Math, where mathematicians first select research problems, then the system attempts to advance proofs, and finally experts review them; he stated that five research projects are currently being advanced, with two having passed expert verification, and three candidate arguments being publicly solicited for review, mentioning that these include the Nivat conjecture. The value of this case lies in placing AI outputs within a verifiable research workflow, rather than solely presenting “solving difficult problems” as a conclusion; the aforementioned progress and verification status in this material come from the blogger’s compilation and should be regarded as their report, not independently confirmed.

Sources:

Google CEO Confirms Signing White House “Super Intelligence” Accord

Google CEO Sundar Pichai posted to confirm that he met with U.S. government and tech industry representatives and signed the White House Accord on Super Intelligence; he mentioned that Google will responsibly develop and release models through testing, evaluation, red team exercises, and other means. A summary of on-site remarks by Dotey added that representatives from Meta, Anthropic, and other companies also participated in the signing, and described the accord’s arrangements for internal controls and internal/external audits; however, their post noted that the full text of the accord and the complete list of signatories were not disclosed on-site. Therefore, what can be confirmed is Pichai’s statement regarding the signing and Google’s safety measures; the complete terms of the accord should still be based on the official text.

Sources:

Phonon-2 Focuses on Long-Form Audio Transcription on Lightweight Devices

Gorden Sun introduced the Fermion open-source speech recognition model Phonon-2: the model is about 164MB, and the post claims that transcribing one hour of audio on a standard thin-and-light laptop takes about 20 seconds, and it only supports English. If the described speed and accuracy performance can be reproduced on common devices, it provides a practical option for local transcription without requiring high-end GPUs; the current material only provides the blogger’s introduction to the model and its test performance, with no independent benchmark comparisons seen.

Sources:

Statistics: Scanned timeline items=669 Matched blogger count=68 Matched tweet total=466 Weighted tweet score=359.5 Original tweet count=219 Retweet count=108 Crawl attempt count=5 Boundary coverage status=tail_confidently_crossed_target_boundary

Usage statistics and your choice

We use short visits, source and network-derived identifiers to measure visible pages and outbound clicks. This usage telemetry does not record page text. Visits expire after 30 minutes of inactivity or 24 hours. This choice controls the added usage reporting and follows browser privacy preferences; existing access statistics remain.