HAI Daily

Gemini 4 Hydrogen, Seedance 2.5 Flash and the AI Infrastructure Shake-up

Google's fast Gemini 4 Hydrogen model, ByteDance's cheaper Seedance 2.5 Flash video model, Nvidia's CEO transition and SpaceX's $74 billion data-centre plan, with tools to test.

By H.AI Tools Team3 stories
Products in this issue: Gemini, Claude, Cursor, GitHub Copilot, Jimeng, Vidu, Runway, Hunyuan Video
In this issue

The short version

  • Gemini 4 Hydrogen is Google's new fast model for everyday coding and agent work, priced at $1.50 and $7.50 per million input and output tokens across a 1M-token context.
  • Seedance 2.5 Flash costs about a third less than Seedance 2.5 per second of video and generates faster, with output capped at 720p.
  • Nvidia named CFO Colette Kress interim CEO after Jensen Huang resigned. Product roadmaps and earnings dates were unchanged.
  • SpaceX plans more than $74 billion of AI computing campuses in Texas and Arizona, but the Arizona site still needs local approval.

November 13 brought two releases you can try today and two infrastructure stories that shape what AI will cost next year. Google shipped a faster default model for coding and agents, ByteDance (字节跳动) cut the price of short AI video, Nvidia changed its chief executive, and SpaceX announced one of the largest data-centre plans yet. Each section explains what was announced and how to act on it.

01

Gemini 4 Hydrogen: a faster default model for coding and agents

What Google announced

Google introduced Gemini 4 Hydrogen on November 12 as its everyday model for coding and agentic work. The heavier Gemini 4 Carbon and Argon models handle long, deep reasoning. Hydrogen is priced flat across its full 1-million-token context window at $1.50 per million input tokens and $7.50 per million output tokens. It is rolling out through the Gemini API, AI Studio, the Gemini app and Search.

Google says Hydrogen completes tasks about 2.5 times faster than Gemini 3.8 Flash at its default setting and makes more than five times fewer fabricated claims. It offers four thinking levels, with a new "xhigh" setting at the top. On Terminal-Bench 4.1, a coding-agent benchmark, Google reports a score of 57.0%. Independent coverage noted that Claude Sonnet 5.5 scores higher on the same test. Hydrogen is a clear step up from Google's previous fast model, but it does not lead every coding comparison. All of these figures are vendor-reported, and we have not run our own tests.

Picking a model tier for daily work

Most teams now choose between a fast default model and a slower reasoning model from the same provider. Use the fast tier for high-volume, well-defined work such as code review comments, test generation, data extraction and agent steps that call tools. Keep the reasoning tier for design decisions, difficult debugging and long documents where a wrong answer costs more than the wait.

To try Hydrogen in an existing workflow, compare it on your own tasks rather than a public leaderboard. Pick five recent tickets from your repository, give each one to Gemini and to the model you use now, and record time to a passing test, the number of corrections you made and the cost per task. Editors such as Cursor and GitHub Copilot let you switch models without changing the rest of your setup. If you want a second assistant for comparison, Claude is the obvious benchmark given the published scores. More options are in Code Generation, and our Claude vs Gemini guide covers the broader assistant comparison.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Seedance 2.5 Flash: cheaper, faster short AI video

What ByteDance released

ByteDance's Volcano Engine (火山引擎) released Seedance 2.5 Flash on November 13 as a lower-cost tier of its Seedance 2.5 video model. Reported pricing is about RMB 0.99 per second of 720p text-to-video or image-to-video, roughly 34% below Seedance 2.5. ByteDance says it generates about twice as fast as the flagship. Flash supports clips of up to 30 seconds, offers 480p and 720p output, and accepts mixed references: up to 10 images, 5 video clips and 5 audio clips in one request. It is reaching ByteDance's Jimeng (即梦) creative app and other platforms first, with API access on Volcano Engine following.

The trade-off is resolution. Flash stops at 720p, while the full Seedance 2.5 offers higher-resolution output. ByteDance positions Flash as better than the earlier Seedance 2.0 flagship in following instructions, physical motion and audio, but those are the company's own comparisons.

A cost-aware workflow for short video

The usual way to spend less on AI video is to split drafting from finishing. Use a fast, cheaper model to test story beats, camera moves and timing, then render only the approved shots at higher quality. A 720p draft is enough to judge whether a scene works.

Try this with a three-shot sequence of about 15 seconds. Write one line per shot covering subject, action, camera and the continuity you need, such as the same jacket and the same lighting. Generate each shot two or three times in Jimeng, where ByteDance releases Seedance first. Note how many attempts each shot needs before it is usable. Then compare cost per usable second against another generator such as Vidu, Runway or Hunyuan Video (混元视频). Check the licence terms for commercial use before publishing, and make sure you have rights to any reference images or audio you upload. The Video Generation category and our AI video generator guide list more options.

03

Nvidia's CEO transition and SpaceX's data-centre plan

What was announced

Nvidia announced on November 12 that Jensen Huang had resigned as president, chief executive and board member with immediate effect. Chief financial officer Colette Kress became interim CEO and keeps her CFO role while the board searches internal and external candidates for a permanent successor. The board also appointed Harvey C. Jones as its first-ever board chair. Nvidia said its third-quarter results would still be reported on November 17 as planned. The company did not give a reason for the change; press reports about the circumstances are not confirmed by Nvidia, and we do not repeat them here.

On the same day, SpaceX said it plans to invest more than $74 billion in two AI supercomputing campuses. About $24 billion would go to a site near Starbase in South Texas, on company-owned land, with its own power generation. About $50 billion would go to Cochise County, Arizona. There, SpaceX projects around 5,500 construction jobs and 1,000 permanent roles and plans dry cooling with no industrial groundwater use. The Arizona project still needs county approval and faces a residents' petition.

What these changes mean for tool users

Neither story changes the AI tools you use this week. Over time, though, infrastructure decides price and availability. When the largest chip supplier changes leadership and new companies build their own computing campuses, the cost of running models is set by more players. For teams buying AI services, that is a reason to avoid depending on a single provider for critical work.

A practical step is to keep one fallback model tested for each important workflow. For a coding assistant, that could mean confirming that your editor works with a second provider, as described above. For video, it means keeping prompts and reference files in a format you can move between generators. If you follow AI infrastructure companies for research, company-research tools in the Finance AI category can help organise filings. Rely on the original announcements rather than summaries, and treat any market commentary as background, not advice.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

What to try next

Test one change this week

If you write code, run Gemini 4 Hydrogen against your current model on five real tickets and compare time, corrections and cost. If you make short videos, draft a three-shot sequence on a fast tier and render only the approved shots at higher quality. For any workflow that matters, make sure a second provider is tested and ready.

Browse Code Generation, Video Generation and Chatbots to build your shortlist, and read our Gemini review for a broader look at Google's assistant.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper