Daily LLM Briefing: Thursday 29 October 2026
Good afternoon. Google leads today with Gemini 4 Carbon, a new frontier model that tops the freshly reset Artificial Analysis index but is going to cyber defenders first. Axiom says swarms of AI agents turned the longest proof in mathematics into machine-checked code in 24 days, and found a real gap along the way. Cartesia has new evidence that models reading raw bytes can beat ones using tokenizers. Particle-1 shows how to make an open model do the same coding work with far fewer tokens. A Swiss "sovereign" model turns out to inherit Chinese weights. And OpenAI is now giving paying ChatGPT subscribers free API credit.
LEAD: Google's Gemini 4 Carbon takes first place, but only cyber defenders get it for now
Background: Google DeepMind launched Gemini 4 Argon just two weeks ago. On Tuesday Artificial Analysis rebuilt its Intelligence Index around mostly private tests, with Claude Opus 5.5 leading at 58.
What happened: On 28 October Google announced Gemini 4 Carbon, which it calls its strongest model, with a focus on cybersecurity, long tasks, coding and agent work. Google says Carbon scores 59 on Artificial Analysis Index v5, putting it at number one. It reports 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploiting them) and 71.0% on CWE-bench (patching them), and says internal tests show less misbehaviour and fewer made-up claims in live coding. Access is deliberately limited: a subset of cyber defenders get it this week, more trusted partners follow through Google's Fairwind Program, and Google is taking part in the US government's voluntary pre-release testing. There is no date yet for developers or consumers. When it opens, the introductory API price matches Argon at $2 per million input tokens and $10 per million output tokens (cached input 95% off), rising to $4 and $20 after the introductory period.
Why it matters: This is the second frontier lab in a month to hold back its best model because of what it can do in hacking, after Anthropic's unreleased Claude exploit work. The benchmark lead is narrow and comes from Google's own announcement, so independent results will matter. For builders, the useful signal is the planned price: top-tier capability at mid-tier cost, once it actually ships.
Sources: Google blog, The Decoder, The Neuron
Axiom's AI agents machine-check "the enormous theorem" and find a gap in a published proof
Background: The Classification of Finite Simple Groups lists the basic building blocks of all finite symmetry. Its proof runs to about 15,000 pages across more than 500 papers by roughly 100 mathematicians. Lean is a programming language where a computer checks every step of a proof.
What happened: On 28 October Axiom released what it calls the first complete computer-checked proof of the classification. Over 24 days its engineers and mathematicians directed swarms of dozens of AI agents, producing about 43.1 million lines of Lean. The work was split into 108 milestones and about 380 tasks that agents claimed through GitHub issues, handing missing prerequisites to other agents. It follows the original first-generation proof and is published on GitHub. Translating it exposed 34 noteworthy deviations from the published record. The biggest broke the standard route to the Unbalanced Group Theorem, and Axiom says it replaced that step with a new result. The Agent Report notes the real bottleneck was Lean rechecking the agents' output fast enough, not model ability, and that humans still had to confirm the formal statement says what mathematicians mean.
Why it matters: Unlike this month's headline maths claims, this result checks itself: Lean either accepts the proof or it doesn't. It shows agent swarms can do huge, verifiable technical work, and that checking, not generating, is becoming the limit.
Sources: Axiom, GitHub, The Agent Report
Cartesia: models that learn from raw bytes, with no tokenizer, are up to 10 times more efficient
Background: Almost every LLM first chops text into tokens using a fixed, hand-built tokenizer. Cartesia introduced hierarchical networks, or H-Nets, in 2025, which learn their own chunks from raw bytes inside the model.
What happened: On 28 October Cartesia published new results from an improved H-Net. In a carefully matched comparison that only swapped the tokenizer, the H-Net needed about 1.28 times less compute to match a standard model, about 2 times on broader evaluation suites, a median of about 2 times on code and around 3 times on programming languages it never saw (up to 5.37 times on Prolog). A larger 3.67B two-stage H-Net reached the final score of Hugging Face's SmolLM3-3B after about 1 trillion training tokens instead of about 11 trillion. Cartesia also shows the model learning new words on the fly within a prompt, and early gains on DNA and images. It says it plans to release some of the architecture.
Why it matters: Tokenizers are one of the last hand-designed parts of an LLM and a known weak spot for code, non-English text and new formats. These are Cartesia's own results at small scale, but if they hold up larger, it could change how future models are built.
Sources: Cartesia blog, original H-Net paper
Particle-1: the same coding answers from DeepSeek, with about 45% fewer tokens
Background: Reasoning models often over-think, re-check and write long confirmations, and every extra token costs money and time.
What happened: On 28 October Particle, the new company from Mixpanel founder Suhail Doshi, launched Particle-1, a version of the open DeepSeek V4.1 Flash trained with reinforcement learning to finish tasks in fewer tokens. Across Terminal-Bench 2.1, ReactBench, ADE Bench and DeepSWE 1.1, Particle says it uses 28.4% to 52.2% fewer output tokens and costs 34.4% to 51.6% less per task, with pass rates within the base model's margin of error. General scores barely move, for example GPQA Diamond 90.4% against 90.9% for the base. Most of the savings come from shorter reasoning and stopping once the job is verified. It is available through Particle's API at $0.25 per million input tokens and $1.00 per million output tokens.
Why it matters: For agents that run all day, token count is the bill. This shows a cheap add-on training step can cut it sharply without hurting results, though all figures so far are Particle's own.
Sources: Particle blog, Particle model catalogue, Neural Brew
Swiss "sovereign" Giotto models turn out to inherit Kimi weights
Background: European governments and companies want AI they control, hosted in Europe and not dependent on US or Chinese labs.
What happened: On 27 October Lausanne startup Giotto launched Giotto Alto, a flagship reasoning model that needs eight Nvidia B300 GPUs and is free on its shared European cloud, and Giotto Solo, which runs on a single GPU for on-premises use. After questions online, its CEO acknowledged the models inherit some weights from Moonshot's Chinese Kimi model, saying the team retrained it with its own methods. Reports say Giotto hasn't said which Kimi version, or whether this applies to both models, and the weights aren't public, so the lineage can't be checked.
Why it matters: "Sovereign" can mean where a model is hosted, or who built it. This case shows the two are different, and that buyers who care should ask about lineage, not just data location.
Sources: Giotto press release (EQS), Neural Daily, LLMs Blog
OpenAI gives ChatGPT Pro and Business subscribers up to $500 a month in API credit
Background: ChatGPT subscriptions and the OpenAI developer platform have always been billed separately.
What happened: OpenAI's API changelog for 28 October says eligible ChatGPT Pro and Business subscribers can now claim up to $500 a month in free API credits by linking an API organisation they own. At its Berlin DevDay Exchange the same day, OpenAI said its Dots assistant should reach more people in Europe "in a few weeks" after talks with regulators; personal Pro accounts in the EEA, Switzerland and the UK are still locked out. It also said weekly Codex users in Germany are up twentyfold since January, without giving absolute numbers.
Why it matters: Bundling API credit into subscriptions makes it cheaper to try OpenAI models in your own tools and pulls developers further into OpenAI's platform. UK users waiting for Dots now have a rough timeline.
Sources: OpenAI API changelog, The Decoder
Also worth a look
- Safety research: DatologyAI's SimpleLoRIF traces which harmless-looking fine-tuning examples erode safety; removing the worst 20% from Tulu 3 cut attack success by up to 22% while keeping capability. arXiv
- Nvidia Nemotron 3.5 Transcribe: an open model combining transcription, translation and speaker recognition. The Decoder
- Claude for Legal: Anthropic added ready-made legal artifacts that check citations and link each statement to its source before a lawyer reviews. The Decoder