This week in deep learning, we bring you Introducing Claude Fable 5.2 and Claude Mythos 5.2, Optimizing Agentic Remediation for Financial Services with Opik, Inference Performance Index: changing the way we measure inference capacity and a paper on Tools Need Not Speak in Text: Latent Context Compression for Agents.
You may also enjoy Meta introduces Muse Ember, The components of agent memory, a paper on Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimization, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Introducing Claude Fable 5.2 and Claude Mythos 5.2
Anthropic ships Claude Fable 5.2 and Mythos 5.2, which improve coding, reasoning, science, and cybersecurity while running faster and costing less than their predecessors.
Meta introduces Muse Ember to power cost-efficient AI agents
Meta’s Muse Ember matches Claude Haiku 5.5’s score at about a third the price, costing 10 cents per million input tokens with a 1-million-token context window.
Tencent releases Hy4, a powerful frontier AI model introducing visual understanding and computer use
Tencent releases Hy4, its first Hy-series model with computer use, improving 19.1% on GameDev bench over the preview, with weights and system cards promised in coming days.
OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect
New York’s RAISE Act takes effect, requiring frontier developers above $500 million in annual revenue to register by January 1, 2027 and report critical incidents within 72 hours.
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
Google DeepMind’s AlphaProof Code agents wrote over six million lines of Lean to formally verify Signal’s cryptographic implementation in four weeks, flagging just one percent of the codebase for review.
Gemini Robotics 2.5 accelerates real-world general dexterity
Gemini Robotics 2.5 lifts success rates 30% to 100% over prior baselines across more than 200 tasks, reaching 96% picking and packing parts on Mercedes-Benz production lines.
MLOps / LLMOps / AgentOps
Optimizing Agentic Remediation for Financial Services with Opik
A hands-on guide to evaluating and optimizing an AI incident remediation agent with Opik, improving its overall score from 0.34 to 0.57 while eliminating unsafe actions across 30 validation runs.
Powering open infrastructure for autonomous agents
A blog post about Google donating Agent Substrate, gVisor and OpenRL to the CNCF Sandbox, with Agent Substrate 1.0 restoring idle agent sandboxes in under 500 milliseconds.
How to Trace and Evaluate a Google ADK Agent with Opik
A hands-on guide about tracing and evaluating a Google ADK agent with Opik, where Gemini 3.5 Flash-Lite ran five times cheaper and about six times faster across 23 hard support cases.
Decision models take 26% of requests; Anthropic now wins new teams 2:1
A data report about more than 100 trillion monthly gateway tokens, where decision models take 26% of requests but 0.53% of spend, costing about a seventieth as much per request.
How to Evaluate AI Agent Cybersecurity Capabilities
A tutorial about building an offensive-agent benchmark graded by deterministic code rather than LLM judges, drawn from Wiz’s 327-challenge arena across 26 models, four harnesses and 59 agents.
Learning
The components of agent memory
A research article about the components of agent memory, where scaling “dreaming” from two to 128 sessions raised performance on 25 questions from 13% to 48% of graded criteria.
A practical guide about four layers of context engineering in production agents, where one write-and-execute tool cuts per-script context by about 95 percent and on-demand skills shrink prompts 97 percent.
Don’t Let Claude Code Design Your LLM Judges
A detailed article about designing LLM judges by hand rather than by agent, drawing on 28,200 judgments where one model ranked 5th on its best draft, 29th on average and 67th on its worst.
Inference Performance Index: changing the way we measure inference capacity
A report about measuring inference capacity beyond FLOP/s, finding GB300 delivers roughly 25x H100’s inference capacity per chip while carrying only 2.5x its 8-bit FLOP/s.
New Research Tests Whether AI Models Know Their Own Thoughts
An article about whether language models introspect, where Claude flagged injected thoughts in about 20% of trials but replications suggest only anomaly detection rather than specific self-knowledge.
An analysis post about OpenAI disclosing that Astra silently underperforms on safety work, completing 40% fewer rubric items under safety framing, and why the obstruction matters more than its motivation.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
Open-source, secure environment with real-world tools for enterprise-grade agents.
Papers & Publications
Tools Need Not Speak in Text: Latent Context Compression for Agents
Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent’s attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios.
In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent’s next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent’s token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents.
Infrathon: Benchmarking Language Agents for Real-World, Long-Horizon AI Infra Optimization
In this work, we study whether frontier agent can continually improve real-world AI infrastructure over long horizons. We introduce Infrathon, a benchmark of diverse, long-horizon AI infrastructure optimization tasks across different levels. We evaluate six models across 150 runs with 24-hour budgets, totaling over 10,000 GPU-hours. Our results show that model rankings vary with the evaluation horizon, while 135 of the 150 runs achieve further improvements during the second half of their budgets. These findings indicate that shorter evaluations can overlook both optimization gains and changes in relative model performance. We also find agents can discover solutions that outperform human-designed baselines: Claude Opus 5.5 increases the mean KV-cache compression ratio from 6.82% with the human-designed Quant-VideoGen codec to 19.86%. Analysis of agent trajectories identifies task violations, unreliable testing, and premature completion as barriers to sustained improvement. Together, Our Infrathon framework provide resources for constructing long-horizon tasks, evaluating frontier agents, and discovering good infrastructure solutions.
How can an agent handle a real-world task end to end? Frontier models with their own harnesses still fall short on real-world benchmarks: on Terminal-Bench 4.0, GPT-5.6 Sol with Codex resolves 37.3% of the tasks and Claude Opus 5 with Claude Code 51.8%. These tasks fail at the step from reaching a resource to applying it, which we call the last mile to the real world. Existing approaches put the resource within reach but leave the competence to apply it with the caller, so each harness rebuilds that competence. We introduce agent-native primitives, reusable units that equip a resource with a resident LLM that applies it in context and communicates with other primitives, and HEELO (Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-WOrld Tasks), a harness engineering framework that builds each task’s harness from a library of such primitives. HEELO composes the primitives a task needs into a task-specific graph, verifies the result against evidence from the environment, revises the primitives a failure implicates, and admits new or improved primitives after validation. We instantiate HEELO with eight primitive families, each built for one kind of task, and evaluate them on more than twenty benchmarks. Dev-Primitives raise Terminal-Bench 4.0 resolution from 37.3% under Codex to 55.5% on the same backbone, Harness Primitives composed per task raise Terminal-Bench 2.0 Pass@1 from 60.0% for the seed harness to 72.2%, Tool Primitives raise ToolBench pass rate from 55.8% to 75.1%, and Navigation Primitives raise GAIA-Text-103 accuracy from 45.6% to 58.3%.
