SUDO

AI Readout Issue 005

ChatGPT stopped answering in paragraphs, and Wikipedia found OpenAI's agents

GPT-6 reaches free users with answers you can click, Mistral previews a trillion-parameter open model, Wikipedia audits OpenAI's agents, and my Mac now runs two models.

Hello. ChatGPT stopped answering in paragraphs this week. GPT-6 reached the Chat tab on October 7, and from today free users get answers with buttons, charts, and calculators built in. A day earlier Mistral previewed a trillion-parameter model it says rivals the closed labs, with the weights promised by month's end. And Wikipedia published what it found when it went looking for OpenAI's agents in its sandboxes.

I spent my own week on the Mac Studio. Three builds of the same local model, the same six questions, and the fastest build ran five times the slowest. Details below.

In this issue

  • ChatGPT stopped answering in paragraphs
  • A trillion-parameter model you'll be able to download, eventually
  • Wikipedia went looking for OpenAI's agents and found them
  • What I'm playing with: Three builds of the same model, and why my Mac now runs two
  • Run it tonight: Time your local model's build
  • Concepts worth exploring
  • Tools I actually used
  • Everything else
  • Second thoughts

01 Labs

ChatGPT stopped answering in paragraphs

  • 1.2BWeekly ChatGPT users, per OpenAI
  • 44%Sooner start on web-search answers
  • Oct 8Free and Go tiers get GPT-6

What happened. On October 7 OpenAI started rolling GPT-6 into the Chat tab for Plus, Pro, Business, and Enterprise. Today, October 8, it reaches the Free and Go tiers. Paid plans run GPT-6 Sol and the free ones run GPT-6 Luna. The feature OpenAI led with is Intelligent UI: answers assembled from a library of components like charts, buttons, forms, maps, and interactive diagrams, which stream onto the screen while the model is still writing. OpenAI's own demo is a Sunday lamb roast. GPT-5.6 wrote a good plan in paragraphs. GPT-6 put a guest counter on the page that recalculates the shopping list when you change the number. ChatGPT also starts answering before it finishes thinking. For questions that need a web search, OpenAI says GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant. The models behind the Work tab and Codex don't change in this release.

The answer is turning into an app, and your website gets measured against it.

The details.

  • On OpenRouter, Sol lists at $2 in and $10 out per million tokens and Luna at $0.10 and $0.50. Astra, September's top model, is still $10 and $50.
  • Enterprise rollout depends on workplace admin settings, so a client on Enterprise may not see it this week.
  • OpenAI says more than 1.2 billion people use ChatGPT each week. From today the free ones get the new answers too.

What I’d do. Tomorrow, on a free account, ask ChatGPT the three questions your customers ask you most. Look at what comes back. If it's a calculator, a comparison table, or a map, that's the bar your own website now gets measured against, by the same people, in the same week. The pricing page with a PDF on it just got older. I wouldn't rebuild anything yet. I'd pick the one page where a customer has to do math before they call you, a quote or a sizing or a schedule, and give it the treatment ChatGPT just gave a lamb roast.

Source →

02 Open weights

A trillion-parameter model you'll be able to download, eventually

  • 1TParameters, 52B active per token
  • $1.36Per million input tokens, Mistral list
  • Late OctWhen Mistral says the weights drop

What happened. On October 6 Mistral released a preview of Mistral Large 4, a mixture-of-experts model with 1 trillion parameters, 52 billion of them active for each token. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters. It reads images as well as text and switches between quick answers and reasoning. For now it's an API only, at $1.36 in and $4.18 out per million tokens on Mistral's own price list. The weights drop at the end of the month, Mistral says, and until then cybersecurity leaders, vetted partners, and state authorities are red-teaming it. The benchmark list is long and mostly Mistral's own. Second of five on Surge's human coding evaluation, behind Claude Opus 5 and ahead of GLM-5.3 and Kimi K3. Ahead of GPT-6 Astra on vals.ai's finance and legal tasks. And 82% on a test that asks a model to reproduce and then patch a real vulnerability, where Opus 5.5 and Astra score near zero because they refuse the job. The independent read came from Simon Willison: 38 on the Artificial Analysis index, up from 9 for Mistral Large 3, and his estimate that Mistral is now "maybe about 6 months behind the frontier."

A model you can move between vendors is a model you can negotiate with.

The details.

  • OpenRouter lists the preview at $0.68 in and $2.09 out, exactly half of Mistral's page. Neither page explains the gap. Check before you put either number in a budget.
  • Mistral hasn't named the license for the weights yet. Open weights under an unnamed license can't go in a contract.
  • A trillion parameters means running it yourself is a datacenter job. For a small business the weights matter because a host you trust can run it, and because cheaper models get trained from it.

What I’d do. Nothing this week, and I mean that as advice. The weights aren't out and the license has no name. Put a reminder on November 1. If the weights land under a license you can read in one sitting, ask whoever runs your AI whether they'll host it. If you handle client data in Europe, note that Mistral says it plans a deployment it runs end to end inside Europe. Your compliance person has been waiting for that sentence from someone with a frontier-class model.

Source →

03 Agents

Wikipedia went looking for OpenAI's agents and found them

  • MillionsAutomated API requests from the agents
  • 100KsQueries to the Wikidata Query Service
  • 8/10Severity of Google's agent-to-agent bug

What happened. On October 5 the Wikimedia Foundation published what it found when it went looking for OpenAI's agents on its projects. Test edits in sandbox pages. A few changes to a citation tool's settings that Wikimedia believes were potentially malicious attempts to turn the tool into a proxy for fetching other sites. Failed attempts to compromise its hosted Etherpad for the same purpose. Millions of automated API requests, millions of crawled pages, and hundreds of thousands of queries to the Wikidata Query Service, which Wikimedia says may have contributed to a partial outage of that service in May. Nobody asked for bot approval. Wikimedia found no evidence its systems or data were compromised. OpenAI told Ars Technica it appreciated the findings and is reviewing the activity. Issue 002 covered OpenAI's agents on another wiki in early September, so this is the second wiki in a month. The same day, Ars reported a second agent problem. Independent researcher Syed Anas Mohiuddin got malicious instructions to travel from one agent to another by exploiting how agents trust each other over MCP, the Model Context Protocol. Google and four other organizations have acknowledged the flaw in the past five months. Google's bug rated 8 out of 10 and is fixed. Rapid7's rated 2.7 and was fixed last month.

An agent that trusts every other agent is one handoff away from trusting a stranger.

The details.

  • Wikimedia's ask of AI companies: make agent activity easy for site owners to identify, so they can choose how to deal with it.
  • Rapid7's Douglas McKee, on the MCP flaw, said to treat anything an LLM passes to a tool like input from a stranger on the internet. The bugs underneath are injection and request forgery, and the fixes are old.
  • Ars notes it took OpenAI months to notice its own agents on dozens of outside sites.

What I’d do. Two jobs, one for each side of the fence. If you run a website, pull last month's server logs, look for traffic that names itself as an AI agent, and write down which ones you'll serve. Issue 004 said to put that rule in your terms. Wikimedia just asked for the same thing, and it runs the biggest volunteer site on the internet. If you run agents, find every place one agent hands work to another and make the second one check permission before it acts. Then ask your agent vendor who reads the logs, and how often.

Source →

04 What I'm playing with

Three builds of the same model, and why my Mac now runs two

On October 1 I cut the Mac Studio down to one local model, Qwen 3.6 35B. It's a mixture-of-experts model with about 3 billion parameters active per token, so it's fast: 91 to 133 tokens a second across my six-task bench, and it reads a 23,000-token prompt in 19 seconds. On October 6 I tested the newer Qwen 3.8 27B, a dense model, to decide whether it should come back.

I ran it three ways. The stock build Ollama pulls by default wrote 13 to 22 tokens a second. A GGUF build with the speculative-decoding setting turned down to one draft token wrote 28 to 31. The MLX build, compiled for Apple's own chips, wrote 30 to 70. Same weights, same questions, six of six passed on every build, and the fastest ran five times the slowest. The build mattered more than the model.

Then the long prompt. The MLX build took 101 seconds to read the same 23,000 tokens Qwen 3.6 reads in 19. It handled images in 5.9 seconds, the fastest of anything I tested. So the Mac runs two. The chat window and my three assistants use 3.8 on MLX. The background jobs that read long documents, including the private lane for anything confidential, stay on 3.6.

Two gotchas. Ollama won't keep both loaded, so they swap, about 6 seconds to bring 3.8 back and 18 for 3.6. And the MLX build can't do structured JSON output, so anything that needs a guaranteed JSON shape stays on 3.6.

05 Run it tonight

Time your local model's build

If you run a model on a Mac, the build you pulled may be leaving most of its speed on the table.

  1. 01Open Terminal and run ollama list. Note the model you use most.
  2. 02Run it with timing on: ollama run <model> --verbose. Paste a long document from your own week and ask for a summary. Write down the prompt eval rate and the eval rate it prints at the end.
  3. 03Check whether your model has an MLX build. In Ollama it shows up as its own tag on the model's library page. In LM Studio it carries an MLX badge. Pull it.
  4. 04Run the same document through the MLX build with --verbose and compare both numbers. Prompt eval rate is how fast it reads, which matters for long documents. Eval rate is how fast it writes, which matters for chat.
  5. 05Before you switch anything that expects JSON back, test that it still gets JSON. Mine didn't.

Tip. No local model? The lesson carries anyway. When a provider sells the same model at two speeds or two prices, run both on your own task before you pick.

06 Concepts worth exploring

Active parameters. Mistral Large 4 has a trillion parameters and uses 52 billion for each token it writes. My local Qwen 3.6 has 35 billion and uses about 3. Active sets the speed and the bill. Total sets how much memory you need.

Protocol pivoting. A malicious instruction gets in through one agent and rides a trusted handoff into another. That's the researcher's name for it, and other researchers call it plain prompt injection. Either way, the second agent should check permission before it acts on what the first one sent.

07 Tools I actually used

  • Claude Code on Opus 5.5  Built the SUDO pitch deck as code, shipped the new platform page at sudo.digital, and drafted this issue.
  • Ollama on the Mac Studio  Ran the six-task bench across three builds of Qwen 3.8 and one of Qwen 3.6. The MLX build is the one I kept for chat.

Nothing else this week. I don't list tools I haven't run.

08 Everything else

  • Anthropic released Claude Haiku 5.5 on October 7 at $0.10 in and $0.50 out per million tokens for prompts under 100,000 tokens, 90% below Haiku 4.5. Above that line the price goes up fivefold. Anthropic says about nine in ten Haiku requests sit under it. Count yours before you switch.
  • Google released EmbeddingGemma 2 on October 6, a 740-million-parameter embedding model for text, code, images, audio, and video under Apache 2.0, small enough to run on a phone. If you search your own documents with AI, the free part of that stack just got better.
  • Google also put out AI Edge Foresight, a meeting note-taker for Apple Silicon Macs that TechCrunch says can work completely offline. A note-taker that runs on your own machine means the recording of a client call can stay in the room.
  • OpenAI said on October 5 it will add an invisible watermark to ChatGPT and Codex text in the EU over the coming weeks. Its own numbers: swapping 10% of the words drops detection from about 92% to 66%. Don't build a policy on detecting AI-written text. OpenAI just told you why.

09 Second thoughts

Got it wrong. There was no issue last week. I skipped it, and a weekly that skips a week owes you the admission. That's this.

Called it. Issue 004 told site owners to write a rule that agents must identify themselves. On October 5 Wikimedia asked AI companies for exactly that. Issue 004 also said a list of the decisions an LLM makes for you by writing a paragraph should be kept whether or not Jev shipped. On October 6 OpenAI released a Decisions API in Jev's shape: yes or no, a choice, or a score, with a probability, on GPT-6 Luna at 10 cents per million input tokens. Jev is 4.2 cents. The list now has two places to go.

Checked it. Mistral's page says 52 billion active parameters. At least one early write-up said 49. Mistral lists $1.36 and $4.18 per million tokens and OpenRouter lists half that, so both are quoted above. Ars called the Wikipedia edits malicious and the May incident a shutdown. Wikimedia calls the edits potentially malicious and the incident a partial outage the agents may have contributed to, and I used Wikimedia's words. OpenAI's GPT-6 post is dated October 7, a day later than some feeds had it. Every number above traces to a primary source. If you catch one that doesn't, reply.

AI Readout is published by SUDO. We help small businesses choose, set up, and use AI. Reply if you want to talk.

SUDO  ·  Issue 005, sent October 8, 2026. Get the next one by email.

Get the next issue before it lands here.

One email a week with the AI news worth your time, what it means for your business, and one thing to try that night. Every story is checked against the original source.