Special Edition

Cheaper Models, Record AI Financing and Computing in Orbit

A Special Edition round-up: GPT-6.1 Terra halves prices, Mistral opens Large 4, Anthropic files for a Nasdaq IPO, Meta and SpaceX line up record financing, Google runs Gemma 4 in orbit, and new energy and publishing rules in China and the UK.

By H.AI Tools Team4 stories
Products in this issue: ChatGPT, Claude, Gemini, Meta AI, DeepSeek, Qwen Chat, Perplexity
In this issue

The short version

  • OpenAI's GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens, half the price of the previous Terra model, and targets everyday computer-use tasks.
  • Mistral released the weights of its roughly one-trillion-parameter Large 4, giving companies a frontier-scale model they can host themselves.
  • Anthropic filed to list on Nasdaq as ANT, Meta set out up to $97.5 billion of share sales, and SpaceX lined up a $40 billion lease for Nvidia chips.
  • Google's Suncatcher satellite ran Gemma 4 on TPUs in orbit, China forecast AI computing will use about 6% of national electricity by 2030, and the UK's CMA opened an investigation into Meta's terms for publisher content.

This Special Edition covers October 26 to 31. Three themes ran through the week. Capable models got cheaper and more open. Record sums of money were lined up for the computing behind them. And computing began to test new limits, from satellites in orbit to national electricity grids and the content that models learn from. The sections below connect the stories and suggest what to do with each.

01

Cheaper and more open models

What was released

OpenAI released GPT-6.1 Terra on October 26 as its everyday model for computer use, writing and visual reasoning, available in the API, ChatGPT and Codex. API pricing is $1 per million input tokens and $5 per million output tokens, half the price of GPT-5.6 Terra. OpenAI says it completes agent tasks more than three times faster than its larger models. Benchmark figures are the company's own.

On October 29, Mistral released the weights of Mistral Large 4, three weeks after an API-only preview. It is a multimodal mixture-of-experts model with about one trillion parameters in total and roughly 50 billion active for each token, released under Mistral's own open-weight licence rather than a standard open-source one. Mistral says it continued training after the preview, so the downloadable model differs from the one that was tested in the API. Running it still requires data-centre hardware.

In China, Apple announced on October 31 that Apple Intelligence will arrive with Alibaba's (阿里巴巴) Qwen (通义千问) models, which we cover in our November 4 issue.

Re-check your model tiers

When a mid-tier model halves in price, the cheapest sensible setup changes. List the tasks your team sends to AI each week and mark which ones truly need the top model. Move the rest to a mid-tier model such as GPT-6.1 Terra in ChatGPT or the API, and compare results on ten real examples before switching.

Open weights such as Large 4 matter for a different reason: control. If your data cannot leave your own cloud, a self-hosted model may be the only option, and the licence decides how you can use it. Read it before building on it. Compare hosted open models from DeepSeek (深度求索) and Qwen too. Our ChatGPT vs Claude guide covers the leading closed assistants.

Get HAI Daily in your inbox

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

02

Record financing for AI computing

What was announced

Anthropic filed its registration statement with the US Securities and Exchange Commission on October 26 to list on the Nasdaq Global Select Market under the ticker ANT. The share count and price range were not set, and the filing had not yet taken effect. An amended filing a week later added third-quarter figures, covered in our November 3 issue.

Meta priced share offerings on October 27: $25 billion of Class A shares and $12.5 billion of mandatory convertible preferred shares. It also plans to sell up to $60 billion more through an at-the-market programme over time. Together that adds up to about $97.5 billion of potential funding for AI infrastructure, not money already raised.

On October 26, SpaceX disclosed a binding commitment for a roughly $40 billion finance lease with an entity managed by Apollo. The entity will buy Nvidia Vera Rubin GPU systems and lease them to SpaceX for its ground-based computing. The deal is expected to close before the end of 2026, with equipment arriving from the first half of 2027.

What the money means for AI users

Deals of this size show that the largest companies expect demand for AI computing to keep growing, and that they are now using public markets and equipment leasing to pay for it. For users, the effect shows up slowly: more capacity eventually means looser usage limits and more competitive prices, but new hardware takes months to arrive.

In the meantime, avoid depending on a single provider for critical work. Keep a tested alternative, export your prompts and settings regularly, and check usage limits before committing a team to a plan. Assistants such as Claude, Gemini and Meta AI all change plans often, so compare them each quarter. This is general guidance, not investment advice.

03

Computing in orbit: Google's Suncatcher answers back

What Google reported

Google reported on October 30 that its Project Suncatcher prototype satellite had run a prompt on the Gemma 4 model using TPU chips in orbit and sent the answer, a haiku, back to the ground. Google calls it the first AI response from a data-centre-class TPU in space. The satellite, built with Planet, launched on October 1 carrying four TPUs. Google plans to test how the chips handle heat and radiation, and a pair of linked satellites is planned for 2027. Google published no speed or power figures, and reports say the satellite can run one TPU for only around 20 minutes at a time.

Keeping expectations realistic

Space computing is an experiment, not a product. The appeal is constant solar power and no land or cooling water. The obstacles are heat, radiation, launch cost and slow links to the ground. China presented its own orbital computing plan in early November, covered in our November 10 issue. Neither will affect the AI services you buy for years. Treat these stories as signals of where the industry sees its next limit, which is power.

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

04

Policy: electricity in China, content rights in the UK

What was announced

On October 30, China's National Energy Administration forecast that electricity used by computing facilities will reach about 800 billion kilowatt-hours by 2030, around 6% of the country's total use. By the end of September, 24 provinces had issued rules for connecting renewable power directly to users, and 12 such projects for computing facilities had been approved.

The same day, the UK's Competition and Markets Authority opened an investigation into whether Meta's proposed business terms would let it use content from publishers' and advertisers' websites to build its own products, including AI features, without a fair say for those businesses. Meta paused the terms on October 28 and says there are no grounds for an investigation. The case has only just started, and no breach has been found.

What to take from it

The two stories show the two inputs AI depends on: electricity and content. Both are becoming regulated resources. If you publish online, review the terms you have accepted with large platforms, especially any clauses about using your content for AI, and decide what your site's crawler settings should allow. If your organisation runs its own AI workloads, ask your cloud provider how its electricity is sourced, because energy rules increasingly affect where and how computing can be built. Answer engines such as Perplexity help you track the official documents behind each policy.

What to try next

What to watch next

Watch for Anthropic's price range and listing date, how quickly Meta draws on its share programme, and whether more companies release frontier-scale open weights. For your own work, the lesson of the week is to re-check which model tier each task needs, keep a backup provider tested and know the terms that govern your content.

Explore Chatbots, Code Generation and Search Engines to build your shortlist. Our Perplexity vs ChatGPT comparison helps you choose a research 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