Axiom says a cluster of agents working under human guidance generated roughly 50 million lines of Lean code in 24 days to complete the formalization of the classification theorem for finite simple groups, while also finding a counterexample to an assertion used in a published proof approach. Anthropic says a research version of Claude found deterministic subexponential-time algorithms for 3-SAT and #3-SAT, but the result concerns asymptotic complexity and does not imply faster solving in practice. Anthropic disclosed six drug development partnerships. One version of the genetic instructions for CAR-T designed by Generate using Claude Code was about 70% shorter and produced roughly 10 times as many receptors as the standard version in laboratory tests.
Axiom works with an existing body of proofs, while a research version of Claude proposes a new complexity result.
Axiom announced that it has completed the Lean formalization of the classification theorem for finite simple groups, with a cluster of agents working under human guidance generating roughly 50 million lines of code in 24 days. The original proof spans about 500 papers and 15,000 pages. Agents claimed tasks through GitHub, reviewed one another's work, and merged the results. The project also found a counterexample to a closure assertion used in a published approach to proving the Unbalanced Group Theorem, and supplied an alternative argument for the simple-group case.
Read original →Anthropic says an internal research version of Claude found a deterministic subexponential-time algorithm for 3-SAT that would refute the Exponential Time Hypothesis (ETH). The algorithm can also count satisfying assignments, solving the harder #3-SAT problem. The paper and an AI-written Lean 4 formal proof have been made public. The result concerns only asymptotic complexity and does not imply better performance when solving formulas in practice.
Read original →🔍 Analysis: Axiom's task was to translate an existing proof spanning roughly 500 papers into Lean code, while Claude's ETH claim comes with a Lean 4 formal proof of a new algorithm. What they share is a machine-checkable form of delivery, not the same discovery process. For developers of research tools, this means the output cannot be merely a persuasive derivation: definitions, assumptions, and reasoning must also be connected into a verifiable proof chain. That verification is no substitute for testing an algorithm's actual runtime performance.
Generate has laboratory results for its CAR-T design, while Anthropic has also launched its own drug discovery program.
Anthropic disclosed research partnerships with six drug development organizations and launched its own drug discovery program. Six scientists at Generate used Claude Code to design segments of genetic instructions for CAR-T therapy. One version was about 70% shorter than the standard version, leaving room for additional instructions. In laboratory tests, the design caused immune cells to produce roughly 10 times as many receptors as the standard version, kill tumor cells more effectively, and remain active longer.
Read original →FrontierCode 2.0 joins the AA Intelligence Index v5, bringing coding trade-offs under incomplete requirements into overall rankings.
Artificial Analysis released Intelligence Index v5, with 60% of its evaluations drawn from private or held-out sets. The update adds tests including FrontierCode 2.0, with four evaluation categories each accounting for 25%. Claude Opus 5.5 (max) leads with 58 points, ahead of GPT-6 Astra at 57. The gap between open-weight and closed-source models has widened from 3 points to 12 points.
Read original →Cognition released the coding benchmark FrontierCode 2.0, whose current top score is 45%. Unlike most benchmarks, which test whether code meets explicit specifications, it deliberately uses prompts with incompletely specified requirements and scores changes on whether a maintainer should merge them. About 80% of submitted PRs were rejected during the benchmark's construction.
Read original →🔍 Analysis: FrontierCode 2.0 reports standalone scores and is also included in the AA Intelligence Index v5, so the question of whether a maintainer should merge a change affects overall model rankings, not just coding leaderboards. The current top score is only 45%, in contrast to the mere one-point gap between Claude Opus 5.5 and GPT-6 Astra in the updated index. For enterprises choosing coding models, similar overall scores do not mean the models can all reliably handle ambiguous requirements; their ability to produce merge-worthy changes needs to be evaluated separately.
LFM3 lowers deployment requirements, Nemotron expands audio capabilities, and TokenGate cuts prefill computation.
Liquid AI released six LFM3 models with around 2 billion parameters, covering text, vision, audio, omni-modal capabilities, and decision-making. The LFM3-1.7B text model supports a 256K context window, takes up about 1 GB with 4-bit quantization, and targets local agents and tool calling. The family unifies multiple modalities on a shared backbone, and all models can be fine-tuned on a single GPU. The vision model adds video input support for the first time.
Read original →NVIDIA introduced Nemotron 3.5 Transcribe, an open-weight speech model supporting transcription in 40 languages with roughly 4.3 billion active parameters. The same model can follow text instructions to perform transcription, speaker labeling, translation, and audio question answering, and can accept specialized terminology as context. It supports recordings up to one hour long, comes in BF16, FP8, and NVFP4 versions, and can be used for commercial or noncommercial purposes.
Read original →Researchers introduced TokenGate, which accelerates long-context prefill through learnable token pruning, and released the code. It adjusts attention weights based on token importance rather than fully processing every input token through every layer. The paper reports a 1.4–11.3× improvement in time to first token without a loss of accuracy on the tasks tested. The component supports architectures including Qwen3 and Qwen3.5, with adaptation overhead comparable to standard parameter-efficient fine-tuning.
Read original →🔍 Analysis: LFM3-1.7B's 256K context window and Nemotron's support for recordings up to one hour long both make long inputs a deployment consideration. But fitting a model onto a device and processing its input quickly are two separate cost questions. TokenGate addresses the latter by reducing prefill computation, complementing the approach of LFM3's roughly 1 GB footprint under 4-bit quantization. It currently supports architectures such as Qwen3 and Qwen3.5, so its reported speedup of up to 11.3× cannot be directly assumed for the new models above. For local application developers, GPU memory usage and time to first token must be assessed separately.
The public script tests URL continuation after a supplied 2,323-token prefix.
CohereLabs released investigation traces, raw token IDs, and a reproduction script for Qwen3.8-Flash-Next. The script prefills the first 2,323 tokens from the record, stopping just before the URL, then lets the model generate up to 48 more tokens. Example outputs include Alibaba Cloud and other domains. This reproduction tests continuation from a supplied prefix, and its outputs vary, so reproducing the target behavior may require multiple attempts.
Read original →