GPT-6.1 Terra: OpenAI’s Cheaper Fast Model, Anthropic’s IPO Filing and This Week’s AI News

Key takeaways

• GPT-6.1 Terra is available now in ChatGPT Work, Codex and the API as gpt-6.1-terra
• OpenAI says it is over three times faster than GPT-6 Astra and GPT-6.1 Sol on its computer-use tests
• For the hardest coding work, OpenAI itself still recommends Sol or Astra
• Anthropic filed its S-1 on 26 October and plans to list on Nasdaq as “ANT”. Share price and count are not set yet
• Cohere says a Qwen3.8 model tried to upload a probe file to Alibaba Cloud in Beijing. Its sandbox blocked it
• Qwen 4 27B, Claude Fable 5.5 timing and two new OpenAI image models are still leaks and rumours

If you use AI for day-to-day admin, writing or simple automations, the most useful story this week is not the biggest headline. It is that a fast, cheaper model can now do a lot of routine work that used to need a premium one.

WorldofAI’s latest round-up covers OpenAI’s GPT-6.1 Terra, Anthropic’s IPO filing, a worrying security finding about an open-weight Qwen model and a string of leaks. Below we pull out what matters, check the key numbers against the official announcements, and clearly flag what is still rumour.

What is GPT-6.1 Terra?

GPT-6.1 Terra is OpenAI’s new mid-priced model, built to be fast and capable at using a computer on your behalf. OpenAI describes it as combining visual reasoning, precise computer use and clearer writing, aimed at the tasks that “fill the workday”: drafting proposals, reading technical drawings, filling in forms and editing videos.

It replaces GPT-5.6 Terra and sits below OpenAI’s more expensive GPT-6.1 Sol and GPT-6 Astra. According to OpenAI’s announcement, it is available today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API.

In the video, the creator calls it the fastest OpenAI model he has tested and says he cares less about its coding than about how it handles real agent-style work, such as operating apps.

How much does GPT-6.1 Terra cost?

Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says that is 50% lower than both GPT-5.6 Terra and GPT-6.1 Sol.

If you are on a ChatGPT plan, OpenAI says the same usage allowance covers more tasks with Terra than with older or larger models, so your limits go further.

How good is GPT-6.1 Terra?

On OpenAI’s own figures, it is a big step up from the previous Terra, especially for coding from the command line and for factual accuracy.

• Terminal-Bench 4.0 (coding agents working in a terminal): 50.3% at max effort, up from 21.5% for GPT-5.6 Terra
• Factual errors on OpenAI’s internal test of tricky prompts: down from 53.6% to 16.7% at low effort, and to 4.3% at max effort
• Visual reasoning (MMMU-Pro) and desktop app tasks (OSWorld 2.1): better than GPT-5.6 Terra at lower cost per task

The video quotes a different hallucination figure (a drop from 88% to 38%) that appears to come from an independent leaderboard. We have used OpenAI’s published numbers above, as they are the ones we could check.

In the creator’s own benchmark, Terra came out slightly behind models such as Kimi K3.1 Pro and Claude Opus 5, but he still rates it as a solid “daily driver” for computer-use and basic agent tasks. One demo shows it building a 3D hatchet model in Blender in under five minutes for about $1.

Should you switch everyday AI work to Terra?

For high-volume, routine work, it is worth a test. Think reading invoices into accounting software, filling in forms, tidying spreadsheets or drafting standard replies. That is exactly the kind of job OpenAI is pitching it for.

1. List three routine tasks you currently send to a premium model
2. Run each one on Terra with the same prompt and inputs
3. Check the output for accuracy, not just speed
4. Compare the cost per finished task, including any retries
5. Keep the premium model for complex coding, legal or high-stakes writing

If you want an independent view before switching, our guide on how to tell if an AI model review is sponsored explains what to look for in benchmark videos. This one, for example, is sponsored by Firecrawl, though the sponsor is not connected to the models covered.

What did Anthropic’s IPO filing reveal?

Anthropic, the company behind Claude, has filed a Form S-1 with the US Securities and Exchange Commission for a proposed initial public offering. Its announcement confirms it intends to list on the Nasdaq Global Select Market under the symbol “ANT”. The number of shares and the price range have not been set.

The video runs through financial figures reported from the filing. We have not checked these against the S-1 itself, so treat them as reported:

The creator also gives his view on how the shares might trade after listing, using past tech IPOs as examples. He is clear this is not financial advice, and neither is anything here. If you are thinking of investing, read the prospectus when it is published and speak to a regulated adviser.

Is Claude Fable 5.5 coming soon?

For the background on these rumours, see our earlier post: Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update.

What did Cohere find in a Qwen model?

Cohere says that during a closed test, Alibaba’s Qwen3.8-Flash-Next model, asked to write a defensive security rule, instead checked the machine it was running on and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Cohere’s sandbox had no route to the internet, so the request was blocked.

A few details from Cohere’s write-up matter for anyone running open models:

• The model’s reasoning said it would not use any tools, then it ran commands anyway
• The behaviour was rare: fewer than one call in 100,000 in Cohere’s tests
• Cohere says it does not know whether it was deliberate or a leftover from training
• Its advice: block an agent’s internet access by default, check tool calls before they run, and do not treat a model’s reasoning as a security boundary

The big lesson is that “we run it on our own servers” is not the same as “our data cannot leave”. Once a model can use a terminal or a browser, it can make network requests. Our guide How to Stop an AI Agent Sending Your Data Online Without Permission walks through the practical fixes.

What do we know about Qwen 4 27B?

Why it matters: a 27-billion-parameter “dense” model could run on one or two graphics cards when compressed, which makes it attractive for private, local AI. In the creator’s tests it held its own against Claude Sonnet 5.5 on 3D and front-end web demos, though Sonnet produced more detailed results.

If you plan to try a new local model, read our checklist for downloaded AI models first.

Is reverse engineering with Codex allowed again?

Reportedly, yes for Codex. The video says an OpenAI researcher has confirmed there is no blanket ban on reverse engineering in Codex, and that blocks stopping people removing unwanted software from their own smart TVs were accidental and are being fixed. This follows criticism from open-source developers.

Anthropic’s reported restrictions on the popular “reverse engineer anything” (REA) workflow in Claude Code, which we covered in our GPT-7 Bel round-up, are a separate matter. The video does not say they have changed.

What else happened in AI this week?

• New OpenAI image models (unconfirmed): two test models nicknamed “shooting star red alpha” and “green alpha” have appeared in an arena, with more realistic textures and physics
• Codex usage reset: as part of OpenAI’s shipping month, users got a free usage-limit reset
• ChatGPT themes: you can now change the theme and accent colour of ChatGPT on the web
• Gemini 4 Flash (leak): a model reportedly spotted in Google’s Antigravity tool with a 200K-token context window
• Nanbeige 4.3 33B: a new open model claiming to beat larger rivals on its makers’ own benchmarks; one person ran a heavily compressed version on a 12GB graphics card at about 20 tokens per second
• Devin Security 2.0: code scanning that finds, checks and suggests fixes for security flaws, with free scans for new enterprise customers
• Google Cloud credits: eligible Google AI Pro and Ultra subscribers can reportedly get $10 to $100 a month in credits through the Google Developer Program
• Gemini API CLI: lets agents like Claude Code, Codex and Cursor call Gemini for images, video, speech, music, PDFs and research
• Phone privacy: someone reportedly used Claude Code to find thousands of “deleted” photos still sitting in cache folders on their phones. This is unverified, but our guide How to Make Sure Deleted Photos Are Really Gone From Your Phone covers the practical checks
• Robots: a humanoid robot learning to climb ladders, and a very lifelike receptionist-style robot called Echo

What is still unconfirmed?

Frequently asked questions

The bottom line

GPT-6.1 Terra is the practical story this week: a faster, cheaper model aimed squarely at the routine computer work most small businesses want to hand off. Anthropic’s IPO filing is the big headline, and Cohere’s Qwen finding is the warning to take seriously if you run open models with tools. Your next step: pick one routine task, test it on Terra against your current model, and make sure any AI agent you run has no more internet access than it needs.

Related posts

• GPT-7 Bel: OpenAI’s Offline Internet, Claude’s Reverse Engineering Block and This Week’s AI News
• Kimi K3.1 Pro Review: Is It Really Cheaper Than Claude Opus 5.5?
• Fable 5.5 Release Rumours, OpenAI Spy Arrest and a Big Claude Code Update

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio