Yesterday, we shared how GPT‑5.6⁠ helped make itself more efficient to run⁠. Today, we’re passing those gains on to customers with lower prices for GPT‑5.6 Luna⁠(opens in a new window) and Terra⁠(opens in a new window) and faster performance with GPT‑5.6 Sol⁠ in the API. Together, these updates help customers get more from every dollar they invest in AI and move faster when time matters. — Read More

#performance

AI labs are moving faster than ever and releasing the most capable models we have ever seen. … But capability is only half of the picture. The other half is how much it costs these models to complete tasks, that is, the cost per successful task. Lower cost makes the model more affordable for users and less costly for the provider. A huge amount of effort inside the labs goes into making every component and layer of their AI applications optimized and more efficient, to reduce the overall cost. For example, GPT 5.6 Sol with max reasoning scores higher than Fable 5 on the Artificial Analysis Coding Agent Index while costing less than half as much.

To understand what techniques are adopted in frontier labs to make AI applications more efficient, we met with the OpenAI engineers who developed and shipped various efficiency techniques into the systems behind Codex and ChatGPT Work. — Read More

For the past few years, Simon Willison has tested every major LLM release with the same prompt: “Generate an SVG of a pelican riding a bicycle”.

… The benchmark is now famous enough that there’s plenty of discussion about its usefulness and about whether AI labs might be benchmaxxing on it.

I wanted to find out, so I put together a small experiment. I generated 1,008 SVGs across seven frontier models, scored them with an LLM judge, and used Claude Fable 5 for the analysis. — Read More

The panelists explain the realities of running AI systems reliably at scale. While building models is solved, maintaining production databases under constant pressure is not. They discuss the emerging architectural decisions separating teams that scale gracefully from those facing catastrophic outages, and what engineering leaders must rethink today. — Read More

#performance

We built a drawing arena: hand a model a blank white canvas and a set of colored-pencil tools, then get out of the way. The model sets a color, tip width, and pressure, lays down batches of strokes, smudges to blend, erases, and calls view_canvas to see its own work and decide what to fix. It either reproduces a target image or draws from a text prompt.

We ran four vision models, GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash, across two targets (the Mona Lisa and Van Gogh’s Starry Night, both scored objectively) and five open-ended prompts, for 28 drawings total. We’ll cover tool use, cost, output, and whether the models actually improved their work, with our opinion at the end. — Read More

Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to effectively fine-tune LLMs with an extreme reduction in trainable parameters. But, \emph{are their learned solutions really equivalent?} We study how LoRA and full-finetuning change pre-trained models by analyzing the model’s weight matrices through the lens of their spectral properties. We find that LoRA and full fine-tuning yield weight matrices whose singular value decompositions exhibit very different structure: weight matrices trained with LoRA have new, high-ranking singular vectors, which we call \emph{intruder dimensions}, while those trained with full fine-tuning do not. Further, we extend the finding that LoRA forgets less than full fine-tuning and find its forgetting is vastly localized to the intruder dimension — by causally intervening on the intruder dimensions by changing their associated singular values post-fine-tuning, we show that they cause forgetting. Moreover, scaling them down significantly improves modeling of the pre-training distribution with a minimal drop in downstream task performance. Given this, we should expect accumulating intruder dimensions to be harmful and lead to more forgetting. This will be amplified during continual learning because of sequentially fine-tuning, and we show that LoRA models do accumulate intruder dimensions here tend to perform worse in this setting, emphasizing the practicality of our findings. — Read More

#performance

Transformers have been established as the de-facto backbones for most recent advances in sequence modeling, mainly due to their growing memory capacity that scales with the context length. While plausible for retrieval tasks, it causes quadratic complexity and so has motivated recent studies to explore viable subquadratic recurrent alternatives. Despite showing promising preliminary results in diverse domains, such recurrent architectures underperform Transformers in recall-intensive tasks, often attributed to their fixed-size memory. In this paper, we introduce Memory Caching (MC), a simple yet effective technique that enhances recurrent models by caching checkpoints of their memory states (a.k.a. hidden states). Memory Caching allows the effective memory capacity of RNNs to grow with sequence length, offering a flexible trade-off that interpolates between the fixed memory (i.e., O(L) complexity) of RNNs and the growing memory (i.e., O(L²) complexity) of Transformers. We propose four variants of MC, including gated aggregation and sparse selective mechanisms, and discuss their implications on both linear and deep memory modules. Our experimental results on language modeling, and long-context understanding tasks show that MC enhances the performance of recurrent models, supporting its effectiveness. The results of in-context recall tasks indicate that while Transformers achieve the best accuracy, our MC variants show competitive performance, close the gap with Transformers, and performs better than state-of-the-art recurrent models. — Read More

#performance

Despite their widespread use, the role of reward models in shaping reinforcement learning is poorly understood. Reward models offer a tempting promise: they automatically estimate response quality in the absence of verifiers or human judges. Unlike “verifiable rewards” which typically produce binary scores, reward models typically produce continuous scores, allowing them to be sensitive to fine-grained differences in responses. However, we show this apparent strength is a serious weakness: many popular reward models are oversensitive, assigning different scores to equally good responses. Theoretically, we show that seemingly perfect reward models can be highly oversensitive; empirically, this oversensitivity can lead to bad policies. In place of existing notions of “reward model accuracy,” we propose evaluating reward models using distinct measures of “discriminative ability” and “specificity” (the complement of oversensitivity). As a solution, we describe a training-free algorithm that uses Monte Carlo dropout on any neural reward model to produce discrete reward clusters. Theoretically, we prove there exist discretizations that reduce oversensitivity at minimal expense of discriminative ability; empirically we show, in both controlled and natural RL settings, that discretizing rewards leads to less reward hacking and better policies than training on the original rewards. — Read More

#performance

In this post, I will present an algorithm that was able to compute an optimal tokenizer in some settings. This result is cool because optimal tokenization is theoretically intractable, but seems to be solvable in practice. My finding is very similar to various results on the Traveling Salesman Problem (TSP), where even difficult instances can be solved optimally using cutting-plane techniques.

I’ll highlight that, while this result is cool, there are a few reasons that it isn’t necessarily useful. First, the existing state of the art was already somewhat close to optimal (often within 1%). Second, even if a tokenizer is optimal on the training data, it may not generalize as well as other tokenizers when evaluated on held out test data. Finally, inefficient tokenizers are basically fine: you can pay for the cost of a less efficient tokenizer by slightly increasing your vocabulary size. — Read More

In Q1 2026, VentureBeat’s Pulse Research surfaced the “Governance Mirage”: the gap between the governance org charts enterprises had drawn and the control layers they had actually built. Forty-three percent said a central team owned AI governance; 23% couldn’t agree on who owned it at all; and 31% named vendor opacity as the single biggest obstacle.

This new wave of research asks the next question: Once you’ve admitted the governance problem, what breaks first when you try to fix it? The answer from our respondents is unambiguous. The failure point is not the model. It’s the runtime. — Read More