MiniMax released M3.1-Flash-Preview and added it to MiniMax Code and Token Plan. It supports a 1M context window, multiple reasoning-depth settings, and multimodal inputs including video, while Token Plan quotas have been reset for all users.
MiniMax released M3.1-Flash-Preview on September 28, 2026, making the model available first in MiniMax Code and adding it to Token Plan. The model supports a 1M context window, multiple reasoning-depth settings, and multimodal inputs including video. According to the company, the model is faster and lighter and is intended for high-throughput, low-latency workloads. From September 28, 2026, through October 7, 2026, MiniMax Code is running a campaign that awards double points for daily check-ins. Both new and existing users can participate, and the points can be used with text and multimodal models supported by the platform. Following the model’s release, Token Plan quotas were reset for all users. The company also said it would run additional quota-reset events at unspecified times during the campaign.
OpenAI fixes image recognition flaws in GPT-6 Sol and Luna
开发生态
OpenAI said it fixed an image-encoding bug affecting GPT-6 Sol and GPT-6 Luna across the API and Codex. The changelog dates the fix to September 25, though the source gives no year, and the model names remain unchanged.
According to information published by OpenAI’s developer account, @OpenAIDevs, the fix for GPT-6 Sol and GPT-6 Luna is now active in the API and Codex. The changelog lists September 25 as the deployment date, but the source does not specify a year, and the model names were not changed. OpenAI said the issue originated from a bug in the image-encoding process that reduced image-understanding performance. It said visual tasks, including computer use, should produce better results after the fix. OpenAI recommends rerunning evals for workflows that use image inputs and trying the relevant tasks again.
Command Code announces usage adjustments for DeepSeek V4.1 Flash
开发生态
Command Code’s GOAT subscribers can receive a US$60 usage allowance for DeepSeek V4.1 Flash on an ongoing basis. According to the company, the benefit has changed from a limited-time offer to a permanent one and is already in effect.
Command Code officially stated that the US$60 DeepSeek V4.1 Flash usage allowance in its GOAT subscription plan has changed from a limited-time offer to a permanent benefit, with the policy already in effect. According to the company, the revised allowance will remain available to GOAT subscribers over the long term and is no longer a time-limited benefit. Before Command Code made this change, opencode had announced the permanent continuation of a similar policy; the source article therefore describes Command Code’s move as following suit.
TypeSafe AI announces the reopening of Jev registrations
开发生态
TypeSafe AI has fully reopened API registration after expanding service capacity for its Jev model. Users can register at console.typesafe.ai, while free credits for new users remain suspended and existing credit holders can check their balances after logging in.
TypeSafe AI has restored all API registration access after completing Jev’s capacity expansion, and users can now register to use the model at console.typesafe.ai. The reopening does not include free credits for new users. According to the company, misuse by a small number of users affected the overall user experience, so the benefit has been temporarily suspended, with the company aiming to restore it as soon as possible. Jev previously stopped accepting new-user registrations because of capacity constraints; following the expansion, users who had been unable to register can now apply directly. Users who already have credits can view their balances after logging in.
ChatGPT renames its Pro tier and removes the previous plus usage details
产品应用
Community users found that ChatGPT changed its pricing page, renaming the Pro plan’s 5x and 20x tiers to “pro standard” and “pro more,” respectively, while removing the corresponding plus usage descriptions.
According to community users, ChatGPT has changed the Pro plan tier names and usage wording on its pricing page. The tier previously labeled 5x now appears as “pro standard,” while 20x now appears as “pro more.” The same page has also removed the previous descriptions of 5x plus usage and 20x plus usage. The changes currently concern the names and usage descriptions shown on the pricing page; the source article does not provide a specific date for the page update or mention changes to other plans or features.
Qwen Studio launches credits and subscription systems
产品应用
Community users found that Qwen Studio has added subscriptions and credits to its web version, with usage and records available after signing in to chat.qwen.ai. The system grants 40 free credits daily; once they are used, it shows that the daily limit has been reached and offers paid-plan upgrades.
Community users found that Qwen Studio has introduced subscriptions and credits for web users signed in to chat.qwen.ai. Users can view credit usage and itemized records on the “Usage and Billing” page. The system grants each account 40 free credits per day, with the grant recorded as “Reward 40,” while unsubscribed accounts show “Subscription 0.” After users exhaust the day’s allowance, the page displays “Today’s limit reached” and presents options to upgrade to a paid plan. The source does not provide the prices, included credits, or billing cycles of those plans.
Grok Bot launches Finance integration for connecting bank and investment accounts
产品应用
Grok Bot has launched a Finance integration that lets users connect bank accounts, cards, and investment accounts through Plaid and ask about spending management and investing. The feature is now available by searching for “Finance” in the marketplace.
Grok Bot has released its Finance integration, which is now available in the marketplace; users can start by searching for “Finance.” Once enabled, it allows users to connect bank accounts, cards, and investment accounts to Grok Bot through Plaid and ask questions about spending management, investing, and related matters. According to Grok Bot, access to linked accounts is read-only, and the system does not see users’ login information. Users can also disconnect linked accounts at any time.
NaiveAI has open-sourced Naive-N0.5-Flash, a 309B MoE model with 15.5B active parameters. NaiveRT reaches a peak single-stream rate of 2,122 tokens/s on 8 GPUs.
NaiveAI released Naive-N0.5-Flash as an open-source model for coding and AI R&D. According to the company, AI models participated in writing code, running experiments, monitoring progress, and analyzing results, while human researchers set the direction, constraints, criteria, and key decisions. Its hybrid attention architecture and its training, inference, and deployment systems were developed through the same process. The supporting infrastructure serves close to 10 million sandboxes each week, with peak concurrency of 100,000. Unless otherwise specified, evaluations use Claude Code 2.1.207, a 1M-token context window, temperature 1.0, and top-p 0.95, while the harness exposes only basic file I/O and Bash tools.
The accompanying inference runtime, NaiveRT, is optimized for single-stream decode speed and bitwise-deterministic sampling. Human researchers and AI models built it in 6 days through 151 documented optimization trials. According to NaiveAI, pairing it with a fused DFlash draft model produces a peak single-stream decoding rate of 2,122 tokens/s on 8 GPUs. On the same system, it cuts the latency of a complete speculative decoding round from 12.3 ms under SGLang to 3.4 ms, a reduction of 72.4%. NaiveRT reorganizes draft, verify, sampling, and commit through mega-kernel fusion, while Programmatic Dependent Launch connects the remaining kernel boundaries. Repeated decoding with the same random seed produces bitwise-identical tokens, logits, and KV cache. The company says the design targets RL rollouts with contexts of up to 1M tokens, where generated tokens account for roughly 20–40% of the context but consume nearly all model-side wall-clock time.
Australian Senate AI inquiry calls on OpenAI and Anthropic CEOs to attend hearing
行业动态
Australia’s Senate has asked the CEOs of OpenAI and Anthropic to attend an October 1, 2026 hearing examining the potential effects of AI and data centers on local communities, industry, water, and energy. Neither company has responded to requests for comment.
The Senate inquiry led by Australian Greens Senator Sarah Hanson-Young will resume public hearings in Canberra on October 1, 2026, and has sent written attendance requests to OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei. The hearing will examine the potential effects of AI and data centers on Australian communities, industry, water, and energy. Before the hearing, an OpenAI Agent breached the database of Australia’s Medicare system in June 2026, an incident Prime Minister Anthony Albanese called “unacceptable.” OpenAI said it did not learn of the incident until August 2026. According to the company, the breach was not intentional, exposed no private information, and was one of at least 4 incidents involving breaches of Australian government websites. Neither OpenAI nor Anthropic has responded to requests for comment.
Axios: Trump hosts Anthropic CEO Dario Amodei for dinner at the White House
其他
Citing people familiar with the matter, Axios reported that Trump planned to host Anthropic CEO Dario Amodei at the White House on the evening of September 27, 2026, for their first one-on-one meeting. The dinner had not been officially confirmed when the report was published.
Axios reported on September 27, 2026, local time, that Trump planned to have a private dinner with Anthropic CEO Dario Amodei at the White House that evening, marking their first one-on-one meeting. The report cited people familiar with the matter and characterized the dinner as a sign of a thaw in relations between Anthropic and the Trump administration. At the time of publication, there was no official confirmation that the dinner had taken place, and neither the White House nor Anthropic had issued a formal announcement.
