DeepMind Opens WeatherNext Models, Claims More Than a Day’s Cyclone Forecasting Gain

01DeepMind Open-Sources WeatherNext Cyclone Models, Claims More Than a Day of Additional Forecasting Lead Time

Google DeepMind has opened its WeatherNext cyclone-forecasting models to researchers and weather organizations, releasing code and model weights alongside a Nature paper. WeatherNext produces multiple possible forecasts to represent uncertainty around dangerous storms, but DeepMind cautions that its output does not replace forecasts or warnings from local meteorological agencies and national weather services.

The release includes WeatherNext Cyclones, which ran during the hurricane season; WeatherNext 2, a later version operationalized in October; and WeatherNext 2-mini, a compact model that can run on a single TPU through a free public Colab notebook. By making the models and their weights freely available, DeepMind says it wants to support academic research, operational forecasting and the development of specialized or localized systems.

WeatherNext uses Functional Generative Networks, or FGNs, to create an ensemble: a collection of different predictions showing a range of plausible outcomes rather than a single deterministic forecast. Last year, the system generated 50 predictions at a time, matching the ensemble size of global physics-based models. DeepMind has now expanded that to 1,000 members, which it says gives forecasters a better view of rare but damaging possibilities, including rapid cyclone intensification.

The company says WeatherNext can generate one 15-day forecast on a TPU in less than a minute. That speed allows users to examine the probability distribution of low-likelihood, high-consequence outcomes without waiting for a large set of conventional simulations.

The models also operate with comparatively coarse input data. WeatherNext Cyclones uses a spatial resolution of 28 by 28 kilometers, which DeepMind describes as 100 times coarser than traditional models. WeatherNext 2-mini works at 111 by 111 kilometers. DeepMind says both still perform strongly, although it acknowledges that researchers do not yet fully understand how the models produce accurate predictions at those resolutions.

DeepMind characterizes the results as adding more than a full day of lead time for cyclone prediction and as equivalent to a decade of meteorological progress. The source, however, does not provide the complete evaluation definition behind that comparison, identify the benchmark model, or give regional results. The concrete claim is therefore limited to DeepMind’s reported forecasting gain, rather than an independently established extra day of warning for every storm or location.

02SoftBank Donates $50 Million to Trump Presidential Library, Announces Federal-Land Data Center Project Two Months Later

SoftBank said it donated $50 million in January to the Donald J. Trump Presidential Library Foundation. Two months later, the Japanese company announced that its SB Energy subsidiary had formed a public-private partnership to lease federal land for an artificial intelligence data center in Portsmouth, Ohio.

The disclosure has drawn scrutiny because the project involves a US Department of Energy site and was announced while Donald Trump was in office. SoftBank revealed the timing in response to a June letter from Senator Elizabeth Warren, Senator Richard Blumenthal and Representative Melanie Stansbury, who raised concerns about possible bribery. The available source does not establish that the donation influenced the lease or any government decision.

SoftBank described the planned facility in March as “the world’s largest” AI data center. The partnership also calls for SB Energy, SoftBank’s energy subsidiary, to build a nearby gas-fired power plant with at least 9.2 gigawatts of capacity. The source does not provide the data center’s physical dimensions, computing capacity, lease terms or approval history.

Warren questioned whether the two-month interval was coincidental and accused the Trump administration of favoring private arrangements with corporate executives over the public interest. The lawmakers also contrasted the contribution with SoftBank’s donations to the presidential libraries of Ronald Reagan and George W. Bush, which they said were made only after those presidents had left office and the relevant institutions had been established.

Their letter noted that a separate entity called the Donald J. Trump Presidential Library Fund had been dissolved by Florida in September after failing to file an annual report. SoftBank said its contribution instead went to the Donald J. Trump Presidential Library Foundation, Inc.

In its response, SoftBank said it understood that the $50 million would support the construction and operation of Trump’s presidential library, consistent with its previous library contributions. The source does not include a fuller company response to the lawmakers’ concern about a possible exchange of favors, leaving the relationship between the donation and the federal-land project unproven.

03ChatGPT to Remove Text-Only Chat Limits for Free and Go Users Next Week

OpenAI says ChatGPT users on its free and Go tiers will be able to hold unlimited text chats starting next week. Those users can currently encounter rate limits after sending many text messages, but OpenAI is removing that restriction for text-only conversations.

The change does not make every kind of ChatGPT interaction unlimited. Messages containing file uploads or images will remain subject to limits. OpenAI has not disclosed the specific allowances for those messages, so the scope of the expansion is limited to pure text chats.

The rollout also changes which models and reasoning controls are available across ChatGPT’s subscription tiers. This week, OpenAI is upgrading the default model for free and Go users to the recently released GPT-5.6 Luna. Next week, those users will also receive a “Think” button that the company says will provide access to higher reasoning for harder questions.

OpenAI has not said whether use of the Think feature will have its own cap. That leaves unanswered whether free and Go users will be able to invoke the higher-reasoning mode as often as they can start ordinary text chats.

Plus and Pro subscribers will receive a different set of upgrades beginning Thursday. OpenAI is updating GPT-5.6 Sol for those customers and adding a slider that lets users choose how much thought ChatGPT puts into an answer.

The company describes the revised Sol model as more reliable with facts. OpenAI says it is designed to use retrieved sources more effectively and make fewer mistakes when an answer depends on dates, numbers, sources, rules or assumptions. It also says the model will produce more direct responses, use tighter formatting and avoid details that do not help.

Those reliability improvements remain company claims: OpenAI did not provide quantitative results showing how much error rates have changed. The confirmed schedule is narrower and concrete—Luna becomes the default for free and Go users this week; updated Sol and its reasoning slider reach Plus and Pro subscribers Thursday; unlimited text chats and the Think button arrive for free and Go users next week.