AX Says · Oct 8, 2026Interfaces and lower prices
GPT-6 brings buttons and charts to ChatGPT answers
OpenAI gave ChatGPT buttons and charts through Intelligent UI, alongside GPT-6. Four user groups got access first; free users get it today. Anthropic cut Haiku 5.5’s cost by about three-quarters and gave Max and Team users API credits. Two darker stories: an Arizona court ordered resentencing because an AI victim video carried too much emotional weight. Common Sense rated ChatGPT’s teen version an “unacceptable risk”. Let’s take them one by one.
OpenAI gives ChatGPT an interface, and GPT-6 arrives
This caught me off guard. OpenAI did not just release a new model today. ChatGPT answers can now contain buttons, charts, forms, and interactive components. Called Intelligent UI, the feature rolls out alongside GPT-6: first to Pro, Plus, Business, and Enterprise on Wednesday, then to free users and the Go plan on Thursday. Demos include a wing-lift diagram, a recipe, an explanation of bicycle mechanics, and a hiking map. Users can also manually reduce how many visual elements appear in answers.
In internal tests, GPT-6 Instant starts answering questions requiring web search 44% sooner on average than GPT-5.6 Instant. It sounds like answering while thinking: give an answer first, then keep working and add to it. OpenAI also says it opened a public beta of the Decisions API today. It claims decisions can be up to 10 times faster than GPT-6 Luna through the Responses API.
“Intelligent UI lets users adjust inputs, explore diagrams, and use tools generated on the fly, right in the conversation.”
Here is what interests me. ChatGPT used to tell you what to do in text. Now it can hand you a clickable button or an interactive chart. I would temper the excitement: it just launched, and we cannot yet see its generation quality, latency, or cost. But the direction is clear. A conversation is becoming an interface for getting things done. Whether it pulls users away from dedicated tools is a question for a few months from now.
Anthropic lays down a price floor instead of flexing
Anthropic’s moves today fit in one sentence: it is handing you a cheaper model and credits together. First, the model: Claude Haiku 5.5 is out. Anthropic says it costs about 75% less to run on average than Haiku 4.5. Available on all platforms, it targets high-throughput work such as summarization, compression, database queries, and classification. It can also serve as a coding subagent alongside Opus 5.5 and Sonnet 5.5. The same day, Claude Code v2.1.293 added Haiku 5.5 to its default models. It also fixed a batch of context-compaction, background-session, and MCP-connection issues.
On credits, Anthropic added monthly Platform API credits to Claude Max and Team plans: Max 5x gets $100 a month, Max 20x gets $200, and Team gets up to $500, shareable across the team. They work with any model, including the new Haiku 5.5, in your own code or a third-party harness. Separately, Claude Sonnet 5.5 cache-read pricing was halved to $0.10 per million tokens. Anthropic says that cuts the cost of most long-running tasks by about 20%.
The context: Claude Code earned Anthropic plenty of goodwill for coding. But token bills for long tasks stung, so many people used it sparingly. A cheaper small model and free credits lower the barrier another notch. I would still temper the excitement. Complex coding needs Sonnet or Opus; Haiku is mainly for work that requires less thought. The point is clear: Anthropic is not flexing today. It is laying down a price floor.
OpenAI releases 372 math solutions that scholars have barely understood
This left me a little dazed. OpenAI released a pile of math papers and solutions last night, but human scholars have barely understood them. Yesterday, OpenAI published 372 results, including a solution to the Unique Games Conjecture; some include Lean certificates. The materials say its unreleased internal models generated 722 math papers and solved 90 of the first 500 open math problems, including results related to the Quasi-Riemann Hypothesis.
Scott Aaronson discussed this under the title “The Mathocalypse.” He compared how two AI companies share mathematical results: OpenAI publishes raw solutions, while Anthropic commissions scholars to organize them. In his words:
“Humans have not yet understood almost any of the solutions.”
It sounds like dumping a stack of unsolved exam papers on the entire math community. Who reads them, verifies them, explains them, and gets academic credit? All are new questions. I care more about how many are genuinely reproducible proofs, versus reasoning summaries that look right but nobody can check. Math is not code where a successful run settles it. Human peers need to verify a proof line by line. If this pace continues, peer review may have to reform itself first.
An AI video prompts resentencing, and ChatGPT fails a safety test
Two AI stories caused trouble in the real world today: one in court, the other in a teen safety assessment. First, the court. An Arizona appeals court ruled that an AI-generated victim video carried “undue emotional weight.” It ordered the defendant’s sentence reconsidered while leaving the negligent homicide conviction intact. Christopher Pelkey’s sister, Stacey Wales, wrote the script and made an AI likeness. In the video, the deceased forgave his killer, Gabriel Horcasitas. The trial judge said he liked the video and imposed the maximum sentence of 10 and a half years. The appeals court found that went too far.
Meanwhile, Common Sense Media’s teen AI safety institute tested more than 4,000 prompts. It rated the teen version of ChatGPT an “unacceptable risk” for minors. In tests using more than 12 accounts linked to parent accounts, none triggered notifications during explicit conversations about suicide, self-harm, or eating disorders. In more than a quarter of situations that called for a crisis referral, users were not directed to seek help.
Together, these cases make an interesting contrast. A court has recognized that emotionally charged AI-generated material is unreliable and ordered resentencing. But platforms have not kept up with their promises to protect teens. Frankly, the evaluators may take a strict line. Still, failure to alert anyone during a crisis conversation is far more serious than a model saying something wrong. Underneath both stories lie the ethical limits of synthetic media in emotional settings. Expect more disputes like these.
NVIDIA and Microsoft put AI agents on the Windows desktop
NVIDIA and Microsoft pushed local AI agents toward the Windows desktop today, with hardware and software together. The hardware is RTX Spark, aimed at Windows laptops and compact desktops. The software is Microsoft’s Microsoft Execution Containers (MXC), now generally available. It lets agents run safely and continuously in the background under operating-system control.
RTX Spark is not pitched as a top-end gaming card. It is meant to let thin laptops and small desktops run local models and agents reliably. Previously, running large models locally meant using a gaming GPU or calling a cloud API. There has not been a comfortable native setup for agents that stay on the desktop. Bringing model execution and agent containers into Windows lays the groundwork for agents that remain on office PCs. MXC’s secure containers are the key selling point. Without them, agents roaming the background would make people nervous.
Whether the combination runs smoothly is another question, but they have staked out a position. Microsoft clearly does not want Windows left behind in the AI agent era, even if most people still use cloud-based GPT and Claude.
Google makes toys and a counterfeit detector
Google was busy today. On one side, it is helping people who cannot code make games. On the other, it is opening AI-content verification to everyone. First, the games: Playground is Google’s experimental gaming platform. Users can create, play, and share custom games through text prompts. They can adjust physics, rules, characters, and environments. Some game types support multiplayer, and it works in phone or laptop browsers. Unity Spark integration remains in testing, and closed testing will open soon.
Now verification: Google upgraded its SynthID detector to support watermarks from all partners. Starting today, everyone can use it through SynthID.com. Previously, only Gemini could check. Now anyone can upload an image, video, or audio file to check for an AI-generated watermark. It can even recognize SynthID watermarks in ChatGPT images.
By advancing generation and detection together, Google seems to be adding an optional “origin label” to the AI content market. But honestly, watermarks are embedded voluntarily. Those willing to cooperate will label their work; those who are not will simply avoid them. Detection can help some people tell what is real, but it cannot catch every fake. This counterfeit detector mostly gives willing users something to work with. As for whether Playground keeps nonprofessional players around, its retention rate in a few months will tell us.
Small signals today
- Engineering tools are piling up. Docker Agent uses declarative YAML to define, run, and share AI agents. It supports automatic delegation across multiple agents, RAG, and MCP, and can push to OCI repositories. GitHub used a ModernBERT fine-tuned classifier to cut secret-detection latency below 2 milliseconds. Pinrail makes coding agents submit specified steps for human review before continuing, with data stored locally.
- A telegraphic-compression test found that compressing records saves 48.4% in tokens. But because gpt-5-mini cannot turn off reasoning, its write costs after compression were twice those of plain text.
- Bites lets users order directly from restaurants through ChatGPT. For the same order, The Verge found a total of $56.73 including tip, taxes, fees, and delivery, versus $70.29 on DoorDash.
- Wikimedia’s investigation found OpenAI agents making unauthorized edits, scraping, and running hundreds of thousands of queries. Meta and Microsoft are also reducing internal Claude use. Microsoft cut its monthly AI spending cap from $100,000 to $10,000. That is a budget cap, not actual spending.
- a16z found that Texas’s data-center grid-connection queue surged from 63 GW to 474 GW. About 90% is data centers, with many speculative applications and too little communication with communities.
Translated from the Chinese original. Read the Chinese original
Some sources via AI HOT
The day's picks
Model releases & updates
Product launches & updates
- ChatGPT adds Intelligent UI with interactive visuals inside conversationsTechCrunch · AI
- Anthropic adds monthly Platform API credits to Claude Max and Team plansAI HOT · Products
- Claude Code v2.1.293 ships Claude Haiku 5.5, fixes session and tool-call bugsAI HOT · Products
- Docker Agent builds, runs and shares AI agents from YAMLHacker News front page
- Common Sense rates ChatGPT teen version unacceptable risk; suicide chats miss parent alertsThe Decoder
- Google launches Playground, a text-prompt platform for making and playing custom gamesHacker News front page
- Google upgrades SynthID AI content detector, opens it to all via standalone siteArs Technica · AI
Research & technical reports
- Why GitHub's secret protection has to scale with software developmentGitHub Blog · AI & ML
Tips & opinion
- Scott Aaronson on how AI math breakthroughs strain human research and publishingHacker News front page
- Telegraph Test: LLMs compress text into telegraph style, and reading it back stays accurateHacker News front page