GPT-6 Intelligent UI, Haiku 5.5 Price War, and Machine-Checked Math

James Turner

Welcome to 48 Hour AI. I'm James Turner. From October 7 to 9, 2026, we got a new ChatGPT that builds little apps for you mid answer, a price war in small models with a catch buried in the fine print, a trillion parameter model from Europe, and 722 math papers dropped on GitHub. Let's start with the ChatGPT change.

James Turner

OpenAI launched GPT 6 in ChatGPT with something it calls Intelligent UI. Here's the concrete version. Say you ask for a meal plan. Instead of a wall of text, you can get a planner with a slider for guest count, and the quantities change as you drag it. Ask about a bike repair and you can get an interactive schematic, an exploded diagram where you tap a part. Ask how to split a dinner check, and the model builds you a tip and bill splitter on the spot.

James Turner

So how does that work without making you wait? OpenAI built a component library and a real time streaming compiler. The interface renders progressively as the model generates, so pieces appear while the answer is still being written, instead of after the whole turn finishes.

James Turner

The second piece is interleaved thinking. Normally a reasoning model goes quiet, thinks, then answers. GPT 6 delivers partial responses while it keeps reasoning in the background. OpenAI says web search queries start answering 44 percent faster than GPT 5.6 Instant. And in its internal evaluations of complex agentic tasks, GPT 6 at Extra High starts answering at the speed of GPT 5.6 Medium, while beating GPT 5.6 Extra High on total quality. Those are OpenAI's internal numbers, so keep that in mind.

James Turner

Rollout matters here. This covers the Chat tab only. Work and Codex are unchanged. Paid subscribers on Plus, Pro, Business, and Enterprise get GPT 6 Sol. Free and Go users get GPT 6 Luna.

James Turner

Developers are split. Fans call it Bret Victor style responsive computing, the early version of software that adapts to you instead of you adapting to it. Critics point to generated images that don't match, like recipes with missing ingredients, plus visual clutter. And there's a bigger worry: if ChatGPT generates the tip splitter on demand, what happens to the standalone utility tools people built? That's a platform centralization concern, and it's an opinion about the future, not a result yet.

James Turner

Now to the small model price war. On October 7, Anthropic released Claude Haiku 5.5, its fastest and most economical small model. The headline is $0.10 per million input tokens and $0.50 per million output tokens. That's a 90 percent cut from Haiku 4.5, which was $1 and $5, and it directly matches OpenAI's GPT 6 Luna.

James Turner

But here's the catch. The 90 percent discount applies only to prompts under 100,000 tokens. Go past that, and the price jumps five times, to $0.50 input and $2.50 output per million. Anthropic says 90 percent of Haiku 4.5 requests historically stayed under 100K. Compare that to Luna, where the step up doesn't hit until 272,000 tokens, at $0.20 and $0.75.

James Turner

There's a second wrinkle. Independent analysis from Simon Willison says Haiku 5.5 uses an updated tokenizer that needs roughly 1.25 times more tokens for the same prompt compared to Haiku 4.5. So the same text costs you more tokens, and that eats part of the nominal savings. Still a big cut, just not quite 90 percent in practice.

James Turner

What do you get? A 1 million token context window and 128K max output. It's Haiku's first adjustable effort setting, low, medium, high, extra high, and max, defaulting to medium, and reasoning can't be turned off. On OSWorld 2.1, the computer use benchmark, it scores 72.4 percent, up from 15.7 on Haiku 4.5. On Humanity's Last Exam it hits 45.9 percent without tools and 57.4 with tools, and 39.2 on Terminal Bench 4.0.

James Turner

Anthropic also cut Sonnet 5.5 cache read prices by half, to $0.10 per million tokens, which lowers typical agent costs by about 20 percent. Max 5x, Max 20x, and Team plans now get monthly Claude API credits, $100, $200, and $500 pooled. And the Python and TypeScript SDKs picked up beta browser and computer use support.

James Turner

Next, Europe. On October 6 and 7, Mistral released Mistral Large 4, nicknamed Le Chonk. It's a one trillion parameter hybrid instruct and reasoning Mixture of Experts model, activating about 49 to 52 billion parameters per token. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and Mistral frames it heavily around European AI sovereignty under EU law.

James Turner

There's a public preview API in Mistral Studio at $1.36 per million input tokens and $4.18 output. Open weights are slated for the end of October. Context is 500,000 tokens, half the 1 million standard from the US closed models.

James Turner

The interesting part is cybersecurity. Large 4 ties GLM 5.3 Flash for first place on the Artificial Analysis Cyber Index. It hits 82 percent on reproducing and patching real open source vulnerabilities, the highest recorded, and 93 percent on Cybench. The reported difference is behavior. US closed models like Claude Opus 5.5 and GPT 6 Astra refuse many real world vulnerability reproduction tasks, while Large 4 carries them out. Vetted partners and state authorities get a reduced moderation version.

James Turner

Now the nuance. On the broader Artificial Analysis Intelligence Index, Large 4 ranks 25th out of 25 tracked models, with a score of 38, behind top US proprietary models and Chinese open weights like GLM 5.3 and Kimi K3. And early testers in OpenCode hit runaway chain of thought loops that burned through the context window without a graceful error. So it's a specialist with a strong niche and rough edges, not an overall leader.

James Turner

Last big story, and the one I keep thinking about. On October 7, OpenAI published 722 research manuscripts on GitHub under Apache 2.0, generated by an unreleased internal frontier model. They picked problems from roughly 4,000 unsolved open math questions, spent about three hours of test time compute per problem, and the results cluster into 372 topic families.

James Turner

What makes this more than a pile of AI text is verification. Every result comes with a formal proof in the Lean proof assistant, so a machine can check the correctness. Results include new bounds on the irrationality exponent of pi, partial proofs of the Hodge Conjecture for specific geometric objects, and refined zero free regions for the Riemann zeta function.

James Turner

The context is governance. The release followed an open letter signed by 25 Fields Medalists urging labs not to privately monopolize fundamental mathematical discovery. OpenAI set up an independent advisory board of top mathematicians to oversee the disclosure. Whether the papers hold up to human scrutiny beyond Lean is the next thing to watch.

James Turner

Two quick hits. On October 8, Google Labs launched Playground, a no code, natural language game creation platform for US adults, 18 and up. You describe a game and get something playable, no Unity or Unreal needed. And on October 7, Google released EmbeddingGemma 2, an Apache 2.0 licensed 740 million parameter multimodal embedding model. It maps text, code, images, audio, and video into one shared space, built for private, on device search and retrieval.

James Turner

So the pattern: interfaces generated on demand, small model prices falling with fine print attached, sovereign European models, and AI math you can machine check. Read the pricing tiers before you celebrate. I'm James Turner, and I'll see you in 48 hours.