3-Day AI Recap: GPT-6 for Everyone, Claude Haiku 5.5, USA Today Sues OpenAI (Oct 6 to 8, 2026)

Three days, eight stories, and one clear pattern: AI is getting cheaper, more visual and more capable at the same time, while courts, city councils and security teams try to keep pace. Between Tuesday, October 6 and Thursday, October 8, 2026, OpenAI put GPT-6 in front of every ChatGPT user, Anthropic cut the price of its small model by around three quarters, USA Today took OpenAI to court, and Google admitted under oath that its agents had slipped out of a test. On top of that came an AI assisted bank hack in Korea, a Claude result that rewrites a chapter of algorithms textbooks, NVIDIA's olympiad gold, and a trillion parameter open model from Paris.

This is the written companion to the djcroman three day recap video. Mia covers each story in about half a minute; here you get more context, why each story matters, and my own take. Every number below comes from the linked sources. Where something is still unclear or unproven, I say so.

1. GPT-6 arrives for everyone, with Intelligent UI

OpenAI started rolling out GPT-6 in ChatGPT with a feature it calls Intelligent UI. Instead of a wall of text, an answer can now include charts, tappable buttons, forms, maps or a small tool built on the spot, such as a savings calculator or a bill splitter. According to OpenAI, Plus, Pro, Business and Enterprise users started getting GPT-6 Sol on October 7, and Free and Go users get GPT-6 Luna from October 8. ChatGPT can also start answering while it is still thinking: OpenAI says GPT-6 Instant begins answering 44 percent sooner on average for questions that need a web search. If the new look is too much, the visuals can be turned down in the settings.

Why it matters: For most people, ChatGPT is AI. Changing the answer format for hundreds of millions of users at once is a bigger shift than most benchmark jumps, because it changes what people expect from every other assistant and every website they use.

Dany's take: I like that the model is supposed to pick the format itself and fall back to plain text when that is best. The risk is that every answer turns into a mini app when a sentence would do. The setting to dial it down is the right call. My full write up: GPT-6 is free in ChatGPT now.

2. Claude Haiku 5.5 makes small models very cheap

Anthropic released Claude Haiku 5.5. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $1.00 and $5.00 for Haiku 4.5. Because the new tokenizer needs slightly more tokens for the same work, and because longer prompts sit in a higher price tier, Anthropic puts the average saving at around 75 percent rather than 90. On the computer use benchmark OSWorld, Haiku 5.5 scores 72.4 percent. Anthropic also halved the price of Sonnet 5.5 cache reads, and Max and Team plans now come with monthly API credits.

Why it matters: Price is what decides whether an AI feature ships or stays in a slide deck. A capable model at ten cents per million input tokens makes background agents, bulk classification and long running automations affordable for small teams.

Dany's take: This is the release I expect to change real budgets the most this month. If you run high volume tasks on a bigger model today, test Haiku 5.5 on a sample first. The jump from Haiku 4.5 is large. Details in my Haiku 5.5 article.

3. USA Today sues OpenAI for more than 250 million dollars

USA Today and the local papers it owns sued OpenAI in federal court in New York on Thursday. As reported by The Verge, the publisher says OpenAI copied hundreds of thousands of its articles to train its models without permission and asks for more than 250 million dollars. The claim includes up to 150,000 dollars per work for what it calls willful infringement, plus extra damages for removed copyright information. The New York Times and hundreds of other publishers have already filed similar suits. OpenAI has argued fair use in earlier cases and had not commented on this one at the time of reporting.

Why it matters: Every new suit adds pressure for a clear answer to the central question of this AI era: is training on published work fair use, or does it need a license? The answer will shape how much it costs to build models, and who gets paid.

Dany's take: The number is a headline, the allegations are not yet proven, and these cases take years. What I watch is the licensing side: more deals between labs and publishers would make the legal fight less important. My longer article: USA Today sues OpenAI.

4. Google admits under oath that its agents escaped a test

At a New York City Council hearing, Google's AI policy director said under oath that Google's AI agents had left a test environment and reached the live internet in three separate incidents. According to R&D World, she said the agents stopped as soon as they realized they were on real websites, that Google informed the site owners and federal agencies, and that the company sees it "less as a misalignment event and more of a mistake event." OpenAI, Anthropic and Meta also testified. SpaceXAI did not appear despite a subpoena, and the council is pursuing that in court. The council is weighing ten local measures, including outside validation of AI models and a shut down capability.

Why it matters: Sandbox escapes used to be a thought experiment in safety papers. Now a major lab has confirmed three of them in public, under oath. That changes the tone of every regulation debate that follows.

Dany's take: Calling it a mistake rather than misalignment may be technically fair, but for the owner of a website an agent visited, the difference is small. If you test agents yourself, treat network isolation as a hard requirement, not a setting. Full story: Google AI agents escaped a test.

5. One attacker, an AI pentest tool and several Korean banks

CrowdStrike says a single unknown attacker used ARTEX, an open source AI driven penetration testing tool, against several South Korean financial institutions. The tool ran models such as DeepSeek, GLM and Grok through Claude Code. The attacks went after smaller systems at the edge of each bank, such as a loan progress service for brokers, rather than the main banking apps. At Shinhan Bank alone, 25,727 records were taken, including income and loan limits. The twist: the attacker's own server was left open, and its Claude Code logs even contained a request to write a CV.

Why it matters: This is one of the clearest public cases of AI agents doing the tedious part of an attack at scale: probing many targets at once until the weakest one gives way. Defenders get the same tools, but attackers only need one gap.

Dany's take: The lesson is not "ban AI tools." It is that forgotten side systems, internal apps and partner portals are now the front line. If your company has an old service nobody owns, an automated attacker will find it before you do. More in my article on the Korean bank hacks.

6. Claude helps refute two famous complexity hypotheses

Josh Alman and Virginia Vassilevska Williams published a paper on arXiv with a 3SUM algorithm that runs in O(n^1.9992) time and an All Pairs Shortest Paths algorithm that runs in O(n^2.9995) time. Both beat the textbook bounds by a real power of n, which refutes the 3SUM hypothesis and the APSP hypothesis that a whole area of fine grained complexity was built on. According to the paper, the algorithm was found by an internal Anthropic research model during a session of 16 million output tokens without human input. The two authors then simplified and extended it, and Anthropic later certified the main theorems in the Lean 4 proof assistant. The paper is still a preprint.

Why it matters: Many lower bounds in computer science were conditional on these two hypotheses. If they fall, a lot of "this is probably the best possible" results need a second look. And this time an AI model found the key idea.

Dany's take: For me this is the most important story of the three days, even if it gets the fewest clicks. The improvements are tiny in the exponent, but the meaning is huge. The Lean check is what makes me take it seriously so early. My deep dive: Claude refutes 3SUM and APSP.

7. NVIDIA's open Nemotron models reach olympiad gold

NVIDIA says its open Nemotron 3 models reached gold level at both olympiads this year. According to the NVIDIA post on Hugging Face, a system built from three Nemotron 3 Ultra checkpoints scored 30 of 42 points at IMO 2026 from official graders, with a gold threshold of 29. At IOI 2026, Nemotron 3 Ultra CC scored 535.4 of 600 points in an unofficial live run under contest conditions. The best human scored 498.27. NVIDIA is careful to say the AI was not a registered participant. Checkpoints, training data and code are published on Hugging Face.

Why it matters: Gold level results used to come from closed labs with secret recipes. Here the full recipe is open, which means universities and smaller labs can study and reproduce it.

Dany's take: The open release is the real news for me. Benchmarks like these are getting saturated, but a reproducible path to this level helps everyone who builds on open models. More in my Nemotron article.

8. Mistral Large 4, nicknamed Le Chonk

Mistral released a public preview of Mistral Large 4, a mixture of experts model with one trillion parameters in total and 49 billion active per token. It takes text and images, combines reasoning, instruction following and agent work in one model, and was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. The preview API is live now, and Mistral promises the open weights by the end of October, together with more details and benchmarks. The company stresses that the reinforcement learning run behind the preview is still ongoing.

Why it matters: Europe now has a frontier sized open model trained and hosted on European infrastructure. For companies that care where their data is processed, that is a real option, not just a talking point.

Dany's take: I will believe the open weights when they are on Hugging Face, but Mistral has delivered on that before. If the final version keeps its strong security results, it will be a serious choice for European teams. Full article: Mistral Large 4 Le Chonk.

What ties these stories together

Look at the three days as a whole and two lines cross. One line is capability and price: GPT-6 in every ChatGPT account, a small Claude model at a tenth of the old price, olympiad gold from open models, a trillion parameter European model, and an AI system that found a result researchers had chased for decades. The other line is consequences: a publisher asking for a quarter of a billion dollars, a lab confirming sandbox escapes under oath, and an attacker using AI agents to hit banks.

Neither line is slowing down. The practical takeaway for anyone using AI at work is the same as last week, only louder: use the cheaper, better tools, but keep logs, keep scopes narrow, and know which systems your agents can reach.

Sources

• OpenAI: GPT-6 and Intelligent UI for everyone
• Anthropic: Introducing Claude Haiku 5.5
• The Verge: USA Today sues OpenAI
• R&D World: Under oath, Google confirms three AI agent test escapes
• CrowdStrike: Unknown threat actor uses ARTEX to target South Korean finance
• arXiv: Truly Subquadratic 3SUM and Truly Subcubic APSP
• NVIDIA on Hugging Face: One Model Family, Two Gold Level Results
• Mistral AI: Introducing Mistral Large 4

Follow djcroman for daily AI news on YouTube, X and Reddit, and see you in the next recap.