Lumina Digest

GPT-6 Comes to ChatGPT for Everyone: With Intelligent UI, Answers Become Charts, Forms and Calculators

Since October 7, GPT-6 Sol has been available on ChatGPT's paid plans; for Free and Go, the announcement points to GPT-6 Luna from the 8th, but an OpenAI guide still refers to GPT-5.6 Luna. GPT-6 also brings Intelligent UI: depending on the question, the model chooses whether to answer with text or with interactive elements.

OpenAI has rolled GPT-6 out to ChatGPT's Chat tab in two stages. Since October 7, GPT-6 Sol has been available to Plus, Pro, Business and Enterprise subscribers. According to TechCrunch, the rollout is global. For the Free and Go plans, the announcement and OpenAI's system card point to GPT-6 Luna from October 8, replacing GPT-5.6 Luna. The model guide in OpenAI's Help Center, however, says Free and Go remain on GPT-5.6 Luna. Users on these plans would do well to check the model picker.

The versions of Sol and Luna now in Chat are the October releases. They are distinct from the ones released on September 22 for ChatGPT Work, Codex and the API (Search Engine Journal). The system card specifies that Codex and Work continue to use the September versions.

The most visible new feature is Intelligent UI. The model decides the shape of each answer and can combine text, charts, diagrams, buttons, forms and tools such as calculators, expense splitters or games. When text is enough, the answer stays text-only (9to5Mac). Interactive elements themselves are not new: since March, ChatGPT has offered interactive visual explanations on more than 70 math and science topics. What changes is the approach: the model now builds the interface as it generates the answer, using a library of native components and a compiler (MacRumors). On questions that require a web search, OpenAI says GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant. This is an OpenAI figure and it measures when the answer starts, not the total time.

There are limitations. The feature is rolling out gradually, on the web and in updated apps, for reasoning levels from Instant to Extra High. It is not available at the Pro level, which still uses GPT-6 Astra, nor in older desktop apps. The “Layout and visuals” setting can be turned off, but some visual elements may still appear. OpenAI's guide specifies that Intelligent UI has no separate quota: the existing limits per plan, model and tool still apply.

Reactions are mixed. On Hacker News, some users talk about “chartjunk” (unnecessary charts), while others appreciate the interactive explanations. Search Engine Journal notes that calculators built in chat could take traffic away from sites that offer the same tools.

Why it matters

• End users: Everyday use of ChatGPT is changing: many answers will arrive as charts or tools to use, not just as text. Those on the free plan should check which model they are actually using, because OpenAI's information is inconsistent. Those who prefer text only can turn off “Layout and visuals”, but some visual elements may still appear.
• Entrepreneurs: If ChatGPT builds calculators, comparisons and forms on the fly, sites that rely on these tools may lose traffic. It pays to focus on content and tools built on proprietary data, which a general-purpose model cannot recreate.

Claude Haiku 5.5 Costs 90% Less Under 100,000 Tokens, Matching GPT-6 Luna

On October 7 Anthropic released Claude Haiku 5.5 at $0.10 for input and $0.50 for output per million tokens, the same base rates as GPT-6 Luna. The discount only covers prompts up to 100,000 tokens, though: above that threshold the rates go up fivefold.

On October 7 Anthropic released Claude Haiku 5.5 (id claude-haiku-5-5), available on its own platform and on AWS, Google Cloud and Microsoft Azure. Pricing depends on prompt length. Up to 100,000 tokens it costs $0.10 for input and $0.50 for output per million tokens, with cache reads at $0.01 and cache writes at $0.125. Above the threshold the rates rise to $0.50 and $2.50. Haiku 4.5 cost $1 and $5, so the discount is 90% below the threshold and 50% above it. According to Anthropic, about 90% of requests to Haiku 4.5 stayed under 100,000 tokens.

Anthropic puts the average saving at about 75%, a figure that already accounts for the new tokenizer, which uses slightly more tokens per task. Haiku 5.5 is also the first Haiku with an adjustable effort level. On the same day Anthropic halved Sonnet 5.5 cache-read pricing, from $0.20 to $0.10. The company says this cuts the cost of agentic work by about 20%. Subscribers are also getting monthly API credits, as SiliconANGLE reports: $100 for Max 5x, $200 for Max 20x and up to $500 shared across Team plans.

VentureBeat and Artificial Analysis add some caveats:

• Pricing: GPT-6 Luna has the same base rates, but only applies its higher rates above 272,000 input tokens. Between 100,000 and 272,000 tokens, Haiku's per-token rates are therefore higher. Gemini 3.5 Flash-Lite costs $0.30 for input and $2.50 for output.
• Benchmarks: on Terminal-Bench 4.0 Anthropic reports 39.2% for Haiku versus 16.4% for Luna. According to VentureBeat, that score was achieved at maximum effort: at the medium level, which is the default, it drops to about 20%. Artificial Analysis independently measured 33% versus 13%, without specifying the effort level. That confirms the lead, but not the exact scores.
• Real cost: it depends not only on the per-token price, but also on how many tokens a task needs, on retries and on accuracy. Artificial Analysis finds that at maximum effort Haiku 5.5 uses about 162,000 output tokens per task in its index, roughly three times as many as Luna. Cost per task has not been published yet.

Anthropic still recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding. Haiku 5.5 also has tighter cybersecurity restrictions than Haiku 4.5: with standard safeguards it blocks penetration testing. Organizations that need broader access can apply to the Cyber Verification Program.

Why it matters

• Entrepreneurs: For classification, document summarization and other high-volume workloads with short prompts, AI spending can drop sharply. Rather than relying on the vendor's claimed average saving, though, estimate it on your real workloads by measuring how many tokens each task needs and how many attempts fail.
• LLM builders / devs: Below 100,000 tokens Haiku and Luna have the same base per-token rates. Between 100,000 and 272,000 tokens Luna has lower per-token rates, but the choice should be tested on cost per task and quality, not least because Haiku can use far more tokens at maximum effort. For agents running on Sonnet 5.5, half-price cache reads lower the read rate without switching models.

Mistral Large 4 “Le Chonk”: European Trillion-Parameter MoE Enters Preview, Open Weights Promised by End of October

Mistral has opened API preview access to its largest model, trained in European data centers and slated for release with downloadable weights within the month. Artificial Analysis calls it the most capable model built outside the US and China, but it still trails the flagship models from OpenAI and Anthropic. It also costs more per task than open-weight models of similar capability.

On October 6, Mistral AI released a public preview of Mistral Large 4, nicknamed “Le Chonk”. It is a multimodal mixture-of-experts model that combines instruct and reasoning modes, with about 1,000 billion (1 trillion) total parameters. The official page puts the active parameters per token at 52 billion. The Register and Artificial Analysis report 49 billion instead. Mistral says it will share architectural details when it releases the weights. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers, on data in more than 160 languages.

For now it is available via API on Mistral Studio. List pricing is $1.36 per million input tokens and $4.18 per million output tokens, but a 50% promotional discount is currently in effect ($0.68 and $2.09). According to Artificial Analysis, the discount applies for the first two weeks. Mistral's documentation lists both prices. The weights are promised “by the end of the month”, and no license has been specified. Mistral also plans to support private-cloud and on-premises deployment, but it has not published hardware requirements. The Register suggests the model could run on eight-GPU servers such as Nvidia's HGX B300 or AMD's MI355X. The company names cybersecurity, finance, law, engineering and manufacturing among its target sectors.

The vendor's own benchmarks are very favorable. On cybersecurity, Mistral claims 82% on the reproduce-and-patch test in Artificial Analysis's Cyber Index, where, according to the company, Claude Opus 5.5 and GPT-6 Astra score “almost zero” because they refuse the task. However, the test methodology (CyberGym-E2E-AA) does not require fixing the target vulnerability. Resolving the reproduced crash is enough to pass, and 31% of passing attempts fix a different crash. Mistral notes that Large 4's refusal rate on cyber prompts is higher than that of the other open-source models tested. Reduced moderation is available only to verified partners and government authorities involved in red teaming.

Independent measurements are more cautious. Artificial Analysis gives the preview a score of 38 on its Intelligence Index, on par with GPT-6 Luna and just below DeepSeek V4.1 Flash (39). It calls it the most capable model developed outside the US and China. On the Cyber Index it scores 50, behind MiMo-V2.6-Pro (56). Cost per task is $1.13 at list price ($0.57 with the discount), compared with $0.25 for GLM-5.3-Flash and $0.27 for DeepSeek V4.1 Flash. The Register highlights its “significant disadvantage” against the flagship models from OpenAI and Anthropic and urges caution about Mistral's benchmarks.

Why it matters

• Entrepreneurs: If the weights really ship, this will be a high-end European model you can deploy on premises, which matters wherever sovereignty and data residency are priorities. Still, performance, licensing and real costs need to be checked before choosing it over the leading US models.
• LLM builders / devs: With a trillion parameters and roughly 50 billion active, it could become an open alternative for agents, coding and security work. But its independent score only matches GPT-6 Luna, and at list price its cost per task is more than four times that of GLM-5.3-Flash or DeepSeek V4.1 Flash. Mistral has not yet published hardware requirements for self-hosting.

OpenAI publishes 722 math manuscripts and retracts three; mathematicians question its method and verifiability

On October 6, OpenAI published 722 mathematics manuscripts produced by an unreleased model. After three retractions on October 7, the catalogue now holds 719, with about 42% of the main results formalized in Lean. Groups of mathematicians are challenging the method, and one study finds discrepancies between a proof and its Lean version.

On October 6, OpenAI published on GitHub 722 mathematical manuscripts, grouped into 372 families, generated by “an unreleased internal OpenAI model” (openai/math repository, Retraction Watch). After three retractions on October 7, the catalogue now holds 719. The model was given about 4,000 open research problems, with an average of about three hours of “ChatGPT Pro thinking” compute per result. When grouping the output into manuscripts, OpenAI kept only results with “an appropriate level of significance”. The company made the selection itself.

How it can be verified. According to the README, about 42% of the main results are formalized in Lean (300 of 719 after the October 7 update). Lean is a language for checking proofs automatically. The same README warns that “some of the unformalized results could have issues” and that some outputs rely on results the models produced earlier. Only condensed summaries of the model's reasoning have been published, for ten results. Corrections will be recorded as new versions.

Retractions and revisions. According to the repository's change history, OpenAI retracted three manuscripts on October 7. In the first, “Algebraicity of Weil classes on split abelian eightfolds”, a sign error invalidates an argument. The other two depend on it and inherit the flaw. Another 14 manuscripts were revised: proofs repaired, statements corrected, hypotheses and dependencies clarified, and one outdated citation fixed. According to Retraction Watch, OpenAI says it found the errors during an audit. The retracted papers now carry a note explaining the gap.

Formalization has limits too. According to Alexander Bastounis, Fabian Circelli and Anders C. Hansen (arXiv:2610.08144), a Lean check does not guarantee that the natural-language proof is correct. TechCrunch reports that the three researchers work at Cambridge and King's College London. On a problem derived from the Navier-Stokes equations, they found at least two discrepancies between OpenAI's natural-language proof and the Lean code. According to TechCrunch, this does not necessarily disprove the solutions, but it raises doubts about whether the models can formalize their own proofs on their own.

How mathematicians reacted. The Advisory Group on Mathematics and AI (AGMAI), hosted by the Institute for Advanced Study in Princeton, includes in its guidelines a call not to test advanced problems on proprietary models. According to TechCrunch, OpenAI followed some of these guidelines but not all of them. For example, the release lacks metadata linking the natural-language proofs to the formal ones. The Association for Human Mathematics (AHM) went further in a guest post on Terence Tao's blog, calling the simultaneous release of more than 700 files “a demonstration of power”. The AHM urges mathematicians to “discontinue their work with OpenAI” and the public to view the release “with due skepticism”. Speaking to TechCrunch, Tao criticized people who feed problems to AI without any interest in the field. Melanie Wood (Harvard) noted that at the time of release, nobody yet truly understands these results.

Why it matters

• Frontier research: This is the largest release of AI-generated mathematics to date. But the model is not accessible, OpenAI made the selection itself, and only part of the results are formalized. The retractions the following day also show that the catalogue is still changing. Anyone citing it or building on it should treat it as a preprint that needs checking, and track the version they use. The Navier-Stokes case also shows that a Lean check guarantees the formal code is correct, not that it matches the written proof. Whether the translation is faithful has to be checked separately.

Nature publishes "byteification": subword LLMs converted into byte-level models without starting from scratch

An Ai2-led team publishes in Nature a method for turning already-trained models into byte-level models with less than 1% of the pretraining budget. Besides Ai2's Bolmo models, the method was applied to Qwen 3 8B and Llama 3 8B. The models improve on character-level tasks but lose some ground in math.

On October 7 Nature published "Retrofitting language models to operate over bytes". The first author is Benjamin Minixhofer, joined by researchers from Ai2, Cambridge, the University of Washington and Edinburgh. The paper formalizes "byteification": you take an LLM that reads text in word pieces (subwords) and convert it into a model that reads raw bytes. A short additional training run is enough, with no need to redo pretraining from scratch.

How it works. Bytes are grouped into "patches" that are processed by the original Transformer, which remains the global model, and then converted back into bytes. In the preprint version the conversion proceeds in two stages:

• stage 1: the global model stays frozen while the local encoder and decoder and the patch boundary predictor are trained (9.8 billion tokens);
• stage 2: all parameters are trained (39.3 billion tokens).

In total, this takes less than 1% of a typical pretraining budget.

The models. Ai2 recalls Bolmo 1B and 7B, introduced in December and derived from Olmo, and adds two new models:

• Bwen 8B, from Qwen 3 8B: according to Ai2's aggregate evaluation it is the strongest of the institute's byteified models and outperforms Bolmo 7B;
• Blama 8B, from Llama 3 8B.

According to Ai2, the two models come close to their source models, but the blog provides no scores.

The preprint's numbers. On CUTE, the character-understanding benchmark, Bolmo 7B scores 78.6 versus 56.9 for Olmo 3 7B. Bolmo, however, was also trained on 75 million tokens of synthetic CUTE-style data, kept separate from the test set. An Olmo 3 trained longer on the same mix and with the same settings reaches 72.9. In the controlled comparison, Bolmo's advantage therefore shrinks to 5.7 points. On code, Bolmo improves only slightly (40.7 versus 39.5 on average) and on HumanEval pass@1 it falls behind (40.6 versus 49.0). In math it loses ground: 29.8 versus 37.5 on MATH and 68.0 versus 73.1 on GSM8K. On multiple-choice STEM tasks it beats BLT 7B, an existing byte-level model, by 16.5 points.

What is new. Neither byte-level models nor the conversion of pretrained models are new. The preprint cites two earlier byteification efforts. The first initialized the byte-level model with the weights of a subword model (Pagnoni et al., 2025). The second trained the boundaries to mirror subword boundaries (Hwang et al., 2025). A paper presented at NeurIPS 2025 by Minixhofer himself had already shown rapid transfer of subword models to the byte level. It did so through cross-tokenizer distillation. According to the authors, Bolmo's contribution is a hierarchical architecture designed specifically for conversion. It includes a non-causal boundary predictor, which sees one byte of future context, and the two-stage procedure. At its core is still a Transformer operating on groups of bytes, so some segmentation still exists.

The limitations. The authors list as limitations:

• the imprecision of the boundary predictor, the main cause of the gap with the source model;
• UTF-8 encoding, potentially inefficient for languages other than English;
• post-training transfer that does not always succeed.

Ai2 also flags as an open question the trade-off between the efficiency of hierarchical byte architectures and fine-grained character understanding.

Why it matters

• Frontier research: The challenge for byte-level models shifts from the cost of pretraining to the quality of the conversion. The boundary predictor and the trade-off between efficiency and character understanding remain open: that is where the contest with subword tokenizers will be decided.
• LLM builders / devs: With a fraction of the pretraining budget you can obtain a byte-level variant of an open model you already use, helpful for spelling, rare strings and character-by-character text. Before adopting it, measure the drop in math and on HumanEval pass@1. The checkpoints, including Bwen and Blama, are on Hugging Face.

Google Cloud Launches Persistent Gemini Agent: Multi-Day Tasks, Its Own Memory and a Workspace Account for Every Digital “Coworker”

At Gemini at Work 2026, Google unveiled a single agent that works in the cloud for hours or days. It has four types of memory and, in its “coworker” version, its own email, calendar and Drive. It can already use Claude models too. For now it is in private preview, and answers on costs, licensing and admin controls are still incomplete.

On October 8, at Gemini at Work 2026, Google Cloud introduced a “universal agent” for work, simply called Gemini. It is a single agent, with a single API, that can answer questions, produce documents and media, and write and run code. In a post signed by CEO Thomas Kurian, the core idea is that you give the agent goals, not instructions. The agent plans the work, uses skills and tools, connects to company systems and delivers a finished result.

The agent runs in the cloud and keeps working when the user is away. Tasks lasting hours or days continue with the laptop closed, with temporary sub-agents working in parallel or in sequence. It has four types of memory:

• session memory, for the task at hand;
• semantic memory, for accumulated knowledge;
• procedural memory, for learned workflows;
• episodic memory, for the history of its activities.

The “coworker” variant has a stable role and its own Workspace account, with an @agents.company.com email address, a calendar, a Drive and an entry in the company directory. You can mention it in Chat or in document comments. It acts under its own identity and sees only what is shared with it.

For each task the agent picks a model from the Gemini lineup and, already today, Anthropic's Claude models. Other models, both proprietary and open, are planned. It can be used from the web, mobile devices, desktop and the command line, as well as from Workspace, Microsoft 365 and Slack. On governance, Google points to:

• cryptographically attested identities;
• logging of every action for auditing;
• an Agent Sandbox for isolated execution;
• an Agent Gateway that filters all agent traffic.

According to VentureBeat, the agent is available in private preview to a group of selected customers, with no date for general availability. It will be included at no extra cost wherever Gemini Enterprise is in place. Several questions remain open, however:

• how much persistent coworkers and long-running autonomous tasks will cost;
• whether every agent will get a full Workspace account, and under which licenses;
• how much control administrators will have over individual services;
• what data will reach third-party providers such as Anthropic.

Independent benchmarks on reliability in long tasks are also missing. The announcement follows similar offerings from Microsoft (Autopilot in Copilot, September 25) and OpenAI (Dots, September 29).

Why it matters

• Entrepreneurs: AI spending is shifting from chat assistants to digital “coworkers” that work on their own for days. The agent is included in Gemini Enterprise at no extra cost, but the real cost of running many persistent agents is not yet known. It makes sense to experiment during the preview phase before committing budget.
• ICT engineers / IT managers: Each agent has its own identity and permissions. Coworkers configured as team members also get a dedicated Workspace account, but Google has not clarified whether this applies to all configurations or how provisioning and licensing will work. It is therefore wise to manage them like user accounts (least privilege, auditing, per-project spending limits) and to ask Google what controls will exist over individual Workspace services and what data will reach third-party providers such as Anthropic.