Claude Haiku 5.5 (High) debuted at #30 with 1587 points in Code Arena: WebDev, just outside the Pareto frontier. It matches GPT‑6 Luna's price at $0.10/$0.50 per 1M input/output tokens while scoring six points higher. Arena says it costs 90% less than Haiku 4.5 for +257 points, and 95% less than Sonnet 5.5 (High) for -128 points.
Unfortunately, newer models have seen the original innovation during their training, making the task significantly easier. However, even with this information, GPT-6 Astra and Claude Fable 5.1 struggled to reimplement it.
Our recipe adapts existing models to bytes with a short additional training run. It keeps the model's core, adding components that group bytes into variable-length patches for it to process, then expand its outputs back to byte-level representations to predict the next byte.
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 https://www.nature.com/articles/s41586-026-11111-4
Ai2 has published its Bolmo byte-level language model research in Nature and released new checkpoints on Hugging Face. The byteifying process converts an existing subword model into a byte-level one with a relatively short additional training run, and the paper reports that it also works for Qwen 3 8B and Llama 3 8B, producing Bwen 8B and Blama 8B. Ai2 also released Stage 1 checkpoints for researchers extending the architecture.
@SpaceXAI On ARC-AGI-3, Grok 4.7 scores 1.8% (vs Grok 4.7's 2.1%) in the standard harness, which lets models carry forward notes between turns, and 10.0% in a new provider adapter harness, which preserves opaque reasoning and enables auto compaction.
DeepSeek V4.1 Flash from @deepseek_ai on ARC-AGI (Verified): - ARC-AGI-2: 72.9%, $0.13/task - ARC-AGI-1: 94.5%, $0.07/task DeepSeek V4.1 Flash beats V4 Flash's best scores by 5.5 points on ARC-AGI-1 and 11.5 on ARC-AGI-2, but costs about 250% more per task.
Benchmarking: We used WMT26 benchmarks, which were released after we created the model, ensuring we couldn't train North Small Translate to its measurements. (5/6)
This is a notable verification because the ~27B model is about as big as you can fit into the official ARC Prize Kaggle competition run time.
@Alibaba_Qwen @baseten - Leaderboard: https://arcprize.org/leaderboard - Reproduce the public results: https://github.com/arcprize/arc-agi-benchmarking - Testing policy: https://arcprize.org/policy - Full Qwen3.8-27B results: https://arcprize.org/results/alibaba-qwen3-8-27b
Over ~17 hours, Ling-3.1-flash built a Lua-to-x86-64 ELF compiler from scratch and passed 178/182 tests (97.8%). That’s within 1.2 percentage points of the Claude Fable 5 result shown (99.0%) and above GLM-5.3-Flash (95.9%).
Databricks tested recent AI models across 2,400 engineers and found Opus 5.5 offers the highest quality mid-tier performance, with about 20% lower same-task costs than Opus 4.8. The company is now encouraging Opus 5.5 as a default model for coding, and reports that GPT-6 Luna is at least 20 times cheaper per task than Opus 5.5, roughly matching Opus 4.6 on one difficult evaluation suite. The Luna findings are preliminary.
Google Developers reproduced Ai2's Olmo 3 7B from scratch in MaxText on Google Cloud TPUs, covering both the stage-1 pre-training run and the stage-2 mid-training anneal. The match was checked on held-out C4 loss, an 8-task accuracy suite, multi-domain perplexity, and token-level KL, not just the training loss curve. The post also describes a data-loader bug that made training loss look better than the reference while held-out metrics did not move.
AIWhy it matters: The post documents how a faithful reproduction was verified on held-out metrics, including a data bug that training loss alone would have hidden.
Shanghai AI Lab and SJTU's LUMIA Lab released NCP-ArchPreview, an 8.9B open-weight language model under Apache 2.0. The model reportedly reaches OLMo-3-7B's final Stage 1 loss using 51.3% of the tokens from the 5.73T Dolma 3 corpus, a 1.95× convergence gain. Its concept module jointly predicts tokens and concepts, and domain adaptation updates only its 17M parameters while the token backbone stays frozen.
METR's preliminary evaluation concludes that Claude Opus 5.5 likely gives slightly higher AI R&D productivity uplift than Fable 5.1 but is unlikely to fully automate AI R&D. The evaluation used five capability tasks over 10 business days of API access, and METR says Anthropic reviewed and edited the summary before sign-off.
AIWhy it matters: The report separates two claims about AI R&D acceleration and discloses that Anthropic reviewed the summary, which helps readers weigh its independence and evidence.
Finance is a particular focus for Step 5 Preview. Financial work has to stand up to scrutiny. We evaluate Step 5 Preview on its ability to identify and verify reliable information, reconcile differences across reports, make assumptions explicit, and produce internally consistent forecasts and reproducible valuations. FinStepBench tests these capabilities across LiveSearch, CorporateValuation, and DeepResearch. We also evaluate Step 5 Preview on FrontierFinance across six investment use cases.
SenseTime released its SenseNova U1.5 technical report, describing an open-source 8B native MoT unified model that connects understanding and generation through shared attention. The model reports 68.2% on VBVR-Pro-Bench, ahead of Nano-Banana-Pro (56.4%) and GPT-Image-2 (50.7%), and its full training recipes, including SFT, RL, and multi-expert on-policy distillation, are open-sourced.
Astra for antibody developability prediction:
Google's RT-2 model, announced in July 2023, trained a multimodal LLM to output robot actions directly, and the article argues this approach launched the current robotics boom. The author follows later work from Physical Intelligence, including action chunking with flow matching, reinforcement learning on real robots, and visual subgoal generation, and notes that the field is debating whether vision-language-action models will give way to world models.
Intern Large Models introduces Visual Pretraining, a pretraining paradigm for foundation models that learns directly from visual documents. The post says it outperforms text-only pretraining across backbones and benchmarks, and links the arXiv paper 2607.09657 along with Intern-S2-Preview (35B) and Intern-S2-Preview-397B on Hugging Face, the latter presented as a multimodal foundation model trained with this recipe.
Still waiting on the tech report but a few innovations: 1/ KDA swaps Gated DeltaNet's single scalar decay for learned per-dimension forgetting. 2/ AttnRes retrieves selectively across depth. 3/ LatentMoE activates 16 of 896 experts, balanced from router score quantiles.
DataChef, an AI4AI framework, uses reinforcement learning to automatically generate optimal data recipes for adapting LLMs. Its DataChef-32B model, using an efficient proxy reward system, matches Gemini-3-Pro in recipe generation, with its recipes surpassing expert-curated ones on AIME'25 and ClimaQA benchmarks.