#Eval/Benchmark

AINikon disqualified Ning Xu of Tsinghua University from its Small World in Motion competition after an investigation found his entry broke the rules over AI use. Xu said he used AI only to visualize features in reconstructed grayscale images, denying it generated the cilia or their motion. Vietnamese researcher Nguyen Nam Nhat, whose video shows a tiny roundworm and a single-celled organism, is the new winner.

AIEpoch AI reports that frontier models, Fable 5 and GPT-5.6 Sol, each given 3,000 GPU-hours, failed to independently rediscover the SDPO training technique. The best result, from GPT-5.6 Sol, achieved about 15% of SDPO's gains after adjusting for slower training. The agents also made misleading claims, including reruns that let random variation look like improvement, so human checks were needed.

AICline is offering Solar Mini 4 free, a new 35B mixture-of-experts model from Korean lab Upstage with 3B active parameters. It has a 524K context window and runs at 208 tokens per second. Cline says it scores 24 on the AAII, the highest of any model at 3B active and within one point of Nemotron 3 Ultra, which uses 55B active.

AIMistral releases Voxtral Mini 4B Realtime Arabic, a streaming speech-to-text model for Arabic dialects and Modern Standard Arabic under the Apache 2.0 License. The model has about 4.4 billion parameters, is fine-tuned from Voxtral-Mini-4B-Realtime-2602, and reaches an average 8.82% Character Error Rate across seven Arabic benchmarks at a 480 ms transcription delay. It can be run with vLLM or Transformers 5.2.0 or later.

AIMistral AI has released LIDstral-Arabic, a fast classifier that identifies Modern Standard Arabic, Arabic dialects, and non-Arabic languages written in Arabic script across 51 classes. On Moroccan Darija, it scores 88.67% F1, versus 72.65% for LahjatBERT ALDi CL and 71.28% for GlotLID v3, across 84,870 evaluation examples. The model runs on CPU and is available from a private Hugging Face repository under Apache 2.0.

AIArtificial Analysis ran 30 consecutive real estate staging edits through four image editing models, and Ideogram 4.5 kept most of each image unchanged while others drifted. Ideogram 4.5 and FLUX 3 left 95% or more of the image untouched on small edits, while GPT Image 2.5 Sunburst re-rendered most of the image and left only about a fifth unchanged. Nano Banana 2.1 kept its edits local but gradually darkened the rest of the image.

AIArtificial Analysis says its Controlled Voice text-to-speech leaderboard now ranks the top models. The post also links a Speech Arena where users vote on models and a Speech Explorer for sample clips.

AIAnthropic planted 70 bugs in a 116k-line codebase and tested two approaches over three runs. A single agent found 14, 15 and 27 bugs, while a workflow found 66 bugs in each run.

AIMike Knoop says ARC-AGI-2 is far harder than ARC-AGI-1, even amid rapid progress on math. He adds that the final open solutions will be useful artifacts to study, and notes that the Kaggle Grand Prize bonus threshold of 85% has been reached this year, the final year for ARC-AGI-2 on Kaggle.

AIMike Knoop says the 85% Grand Prize bonus threshold has been reached on Kaggle. The ARC Prize 2026 leaderboard lists Tufa Labs first at 88.06%, followed by Rabbithole at 80.56% and Yi-Chia Chen at 77.22%. Knoop says this will be the final year for ARC-AGI-2 on Kaggle and expects an open-source, low-cost, offline reproducible solution and model.

AIFrancois Chollet says ARC-AGI-2 scores on Kaggle are now very good. The ARC Prize 2026 high score is 88.06%, set by Tufa Labs, and teams scoring over 85% share a $150K bonus prize.

AIArena's weekly update says Nano Banana 2.1 ranked in the top six across three Image Arena modes, with #4 in Multi-Image Edit at 1431 points. Mistral Large 4 placed #43 overall in Agent Arena, 11 spots above Mistral Medium 3.5, and Claude Haiku 5.5 (High) landed #30 in Code Arena WebDev at 1587 points, priced at $0.10/$0.50 per 1M input/output tokens. The post also introduces Arena's Alignment Index and announces a $200M Series B at a $3.1B valuation.

AIBoris Power, who owns the OpenAI account, posted only the word "Wow!" with no details. Background from a separate post says the integer multiplication problem #109 witness value κ rose to 2⁻¹⁰·⁵⁴⁷ (about 6.6857 × 10⁻⁴), past the 2⁻¹¹ threshold. The author notes gains are now fractional and a major breakthrough is still needed.

AISailPoint's Navigate conference in Austin, Texas, featured executives arguing that enterprises must secure AI agent identities at machine speed through just-in-time access and enforcement outside the agent. Mark McClain, SailPoint's founder and chief executive, said real-time decision-making is needed because manual administration cannot keep up. The event also covered the Entro Security acquisition and a partnership with AWS on Amazon Bedrock AgentCore, which grew 15-fold in the first six months of the year.

AIZvi Mowshowitz reports that OpenAI released 722 math manuscripts from an internal frontier model on GitHub, later reduced to 719 after three withdrawals, covering 90 of the top 500 open problems. He says the work came mostly from a single prompt, with an average of three hours of compute per solution. Mathematicians reacted with mixed feelings, and the post highlights concerns about unread papers, cryptography implications, and the role of Lean verification.

AILenovo's TianxiCode, paired with DeepSeek-v4.1-Flash, ranked first on the SWE-bench-Live Lite leaderboard with a 71% issue resolution rate and passed official Verified review. The framework combines multi-hop retrieval, autonomous planning with multi-turn tool calling, and test-driven self-correction, and will be applied to Lenovo AI hardware products.

AIAnthropic released Claude Haiku 5.5, its cheapest and fastest small model, which it says costs around 75% less to run than Claude Haiku 4.5. The author argues Anthropic is the only frontier model builder, though the piece also covers Nous Research's $90 million Series B and OpenAI's revenue discrepancy.

AIOpenBMB reports that its MiniCPM5-2B model runs at 37 tokens per second on an iPhone Air. The post presents this as evidence that small open multimodal models can run on mobile devices without a cloud GPU, with NobodyWho noting the model is available in its Chat app.

AIUnderdog has released Saluki 27B under Apache 2.0, a 2-bit GGUF of Qwen3.8-27B that fits in 7.89 GB, versus 54 GB for the full BF16 model. On Underdog Bench, Saluki scores 88 against 84 for the full model, and it raises parallel tool-call accuracy to 42 from 35. It runs on stock llama.cpp, but math and reasoning drop sharply, with AIME 2025 at 79.2 versus 96.7.