At Airbnb, we use AI models for classification tasks, such as detecting risky user-generated content, identifying why customers contact support, and recognizing topics in reviews. These tasks require subtle language understanding and the common sense to interpret what people leave unsaid.

Frontier models, with hundreds of billions of parameters and internet-scale pretraining, are often good at these tasks. But for interactive products and high-volume workloads, accuracy is only part of the requirement. We also need low latency, manageable serving costs, and control over the tradeoff between false positives and false negatives.

… GLOWC, an internal framework used by engineers at Airbnb, makes it easy to fine-tune open-weight language (and vision-language) models into fast, accurate classifiers. In this blog post, we put one of these classifiers head-to-head with open-weight and proprietary models and find that it matches frontier-level quality in under 100ms at roughly 1% of the cost. — Read More

Modern LLMs achieve impressive performance via pre-training and in-context learning (ICL). However, once deployed, they remain largely static, operating with a fixed knowledge cutoff date that leaves them unable to acquire long-term capabilities or adapt to new information.

While updating models with new data seems like the obvious solution, continual parameter updates often trigger catastrophic forgetting, a fundamental issue where learning new tasks severely degrades performance on previously learned ones. Alternatively, relying strictly on ICL limits knowledge retention to the immediate context window: once the context session ends, the acquired short-term information is forgotten.

nspired by how the human brain processes and stabilizes memory during rest, we introduce a sleep framework in continual learning: a learning paradigm that equips large language models (LLMs) with a continuous periodic lifecycle, enabling them to continually learn, consolidate fragile short-term context into stable long-term parameters. With this framework, we demonstrate that alternating between active interaction and offline memory consolidation allows LLMs to continuously acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning. — Read More

ChatGPT, Claude and Gemini all came out of America. Yet America couldn’t produce a free model that stayed in the race.

For close to two years, anyone hunting for the best AI they could download and run on their own hardware ended up with a Chinese model. DeepSeek, Qwen, Kimi, GLM, MiniMax, Xiaomi’s MiMo. They kept taking turns passing each other, while American labs locked their strongest work behind subscriptions.

… Then on October 5, a New York startup, already known in some circles as “the American DeepSeek,” finally released its response. The model is called Beam. — Read More

A lot of recent progress in language models has focused on better reasoning, larger reinforcement-learning runs, and increasingly sophisticated agent training. One of the many remaining bottlenecks is the context. Coding agents, research agents, and tool-using systems repeatedly process hundreds of thousands of tokens while generating comparatively little output. At that point, prefill compute and the KV cache become the most important infrastructure problems.

DeepSeek-V4.1-Flash is essentially an attempt to redesign the Transformer around that workload. Its 552B-parameter MoE backbone uses a new Causal Encoder-Decoder (CED) architecture, activating only about 8B parameters per token during prefill and 16B during decoding. Compressed Sparse Attention 2 (CSA2) shares KV representations and sparse-attention decisions across layers; SWA Bounded Replay avoids persistently storing sliding-window KV states; and FP4 KV quantization pushes the global cache down to only 890 bytes per token — about one quarter of DeepSeek-V4-Flash’s. The model still supports a 1M-token context, native vision, and strong agentic capabilities. It is trained on 45T multimodal tokens, with much of the improvement coming from scaling automatically generated agent tasks, environments, and rollouts. — Read More

Language models are changing how we build software and solve computational problems [1] [2]. Coding agents put these capabilities to work through harness, a software system that manages a model’s tools, context, and task execution [3] [4]. While models provide the core intelligence behind coding agents, harnesses are increasingly seen as central to how effectively that intelligence is used [5] [3]. Choosing a coding agent therefore means selecting both a model and a harness, even when the explicit focus is only on the model [6].

… We evaluate 21 model–harness pairs spanning seven models and three harnesses—Claude Code, Codex CLI, and Pi—on SWE-bench Lite and Terminal-Bench 2.0 [8] [9]. Our study reveals three surprising findings on these two open-source benchmarks:

— Models may perform better with other harnesses than with their own. So it turns out that your Claude models may not need Claude Code… 
— Harness choice has little effect on task success rate, but can significantly affect the cost on the benchmarks we test. The same model can achieve similar success rates at up to 5x costs.
— A simple harness can be competitive. Pi, a minimal, open-source harness can be competitive on both cost and task success rate.

— Read More

Pricing is 2.5x GPT-5.6 Sol’s current prices across the board, up from $4/$20 to $10/$50 per million input/output tokens, with the same 90% discount for cache reads and 25% premium for cache writes.

We see distinct stories across our two flagship Indices. In the Artificial Analysis Coding Agent Index, GPT-6 Astra equals Fable 5 at less than half the cost, driven by significant token efficiency gains. In the Artificial Analysis Intelligence Index, GPT-6 Astra is more token efficient than its predecessor for similar performance, but this is offset by the price increase. — Read More

This article grew out of the comments on my previous piece, Why AI Research Tools Struggle to Fact-Check the Web, where, using GPT Researcher as an example, I showed how common search engines and context-building methods undermine the very foundations of AI-based fact-checking.

Not that those tools were bad. They’re brilliant at standard (re)search tasks. But when finding genuinely diverse and independent sources matters, they start to show their limits. So we fixed that part. Adding source assessment through an admission policy was enough to make a difference.

But as the readers’ comments quickly showed, I was only touching the tip of the iceberg. — Read More

The rumors were true, all of them (and then some): OpenAI today is releasing GPT-6 Astra, a new frontier model that the company says likely marks the onset of artificial generalized intelligence (AGI), its long sought goal of “highly autonomous systems that outperform humans at most economically valuable work.”

In a closed a press briefing earlier today, OpenAI co-founder and president Greg Brockman offered an unusually direct formulation of that message, ending the session with: “Welcome to the AGI era.”

That is an unusually consequential framing even by the standards of frontier AI launches. But for enterprises, the more immediate significance of Astra may be considerably more concrete: OpenAI is positioning GPT-6 Astra as a new era of computing in which users, including employees, no longer have to click around a mouse or type on a keyboard ever again (if they don’t want).

… Astra is designed to navigate software much as a person does — working across browsers, spreadsheets, websites and desktop applications, producing finished documents and presentations, and carrying out multistep workflows rather than merely telling a user how to complete them. — Read More

ArtificialAnalysis is a website that benchmarks the intelligence of various LLM models. They publish a headline Intelligence Index, which is calculated as the mean output of the curated selection of benchmarks they run on each model. It’s a decent finger-in-the-air measure of how smart a model is overall.

AA also records useful information — namely, how much it cost them to run the benchmarks. Since the benchmarks are the same across all models, this offers a good indicator of how much it will cost a user to run each model, in relative terms.

One of their main plots is the Intelligence vs. cost plot, which shows the Pareto frontier, i.e. the cheapest model that can achieve each intelligence score. This frontier is important, because using a super-intelligent and super-expensive model to accomplish menial tasks that could be done by a much dumber and cheaper one is just a waste of money.

Over time, I’ve become progressively more irritated by this plot, for a few reasons. — Read More

A closely followed measure of artificial intelligence token prices touched fresh lows this week, the latest sign of deflating prices in an increasingly competitive landscape.

The LLM Token Expenditure Index, a key gauge of daily prices from intelligence firm Silicon Data, fell to 97 cents on Monday. That marked the index’s lowest reading since its creation late last year and has fallen by more than half from the high recorded earlier this summer.

….The recent drop is driven in part by the rise of open-source Chinese models like Moonshot’s Kimi K3 that can fetch lower prices than alternatives from leading frontier labs, according to a Tuesday post from Charles-Henry Monchau, investing chief at Syz Group. — Read More