State of AI Engineering

What The State of AI Engineering Report Reveals About AI in Production

AI engineering has moved past experimentation, but are your observability practices keeping up? Datadog analyzed LLM telemetry from 1,000+ customers to reveal what's actually happening at scale: model adoption shifts, hidden token costs, and what the rise of agentic frameworks means for reliability. Download the report to benchmark your AI stack against production reality.

pnpm experiments with Rust! - pnpm 12 is being rebuilt in Rust to tackle the startup and filesystem overhead that can make JavaScript package management painfully slow. Benchmarks show dramatic improvements, with some warm-install tests dropping from 472ms to just 15ms. Read more.

Shopify’s LLM gisting! - Shopify explores gisting, a technique that compresses long prompts into smaller representations while keeping the important information intact. It explains how this can reduce context size, lower inference costs, and make AI agents faster. Read more.

GPT-5.6: OpenAI’s New Frontier Model Is Built for Real-World AI Work

OpenAI has introduced GPT-5.6, a new family of AI models focused on combining advanced reasoning with better speed, efficiency, coding, and agentic capabilities. The GPT-5.6 family includes GPT-5.6 Sol, the flagship model, alongside GPT-5.6 Terra for balanced performance and GPT-5.6 Luna for cost-efficient workloads. OpenAI’s main goal is not simply to make AI smarter, but to make advanced intelligence more useful and affordable for real-world work.

What Is GPT-5.6?

GPT-5.6 is designed for complex tasks that require models to reason, use tools, work with large amounts of information, and complete multiple steps with less human guidance. OpenAI says GPT-5.6 Sol delivers state-of-the-art performance across coding, knowledge work, cybersecurity, and science while using fewer tokens and lowering the estimated cost of completing tasks.

The model family is also designed around different levels of capability. Sol targets demanding workloads, Terra provides a balance between capability and cost, while Luna focuses on making AI intelligence more affordable at scale.

GPT-5.6 Coding and AI Agents

Coding is one of the biggest areas where GPT-5.6 aims to improve. OpenAI describes GPT-5.6 Sol as its best coding model yet, with strong results on coding-agent evaluations including Terminal-Bench 2.1 and DeepSWE.

GPT-5.6 can also write and execute lightweight programs that coordinate tools, process intermediate results, monitor progress, and decide what to do next. This makes it particularly useful for AI coding agents and long-running software engineering workflows.

Another important feature is Programmatic Tool Calling, which allows the model to process large amounts of intermediate information without sending every result back through the model. This can reduce unnecessary tokens and model round trips.

More Reasoning When You Need It

GPT-5.6 introduces different levels of reasoning investment. The max setting gives the model more time to explore alternatives, perform checks, and revise its approach. The ultra setting goes further by coordinating multiple agents working in parallel.

This approach is important because not every problem requires maximum reasoning. Simple tasks can remain efficient, while complicated research, coding, and analysis workflows can receive additional compute when it matters.

GPT-5.6 Goes Beyond Coding

OpenAI is also positioning GPT-5.6 as a model for broader knowledge work. It can work with documents, presentations, spreadsheets, browsing tasks, and computer interfaces while producing more polished final outputs.

The model also brings stronger design and computer-use capabilities. Instead of simply generating an interface, GPT-5.6 can inspect the rendered result, identify visual or functional problems, and refine the output.

The Bigger Picture

The most interesting part of GPT-5.6 is its focus on performance per dollar. OpenAI is moving beyond the idea that every AI improvement must come from a larger and more expensive model. Instead, GPT-5.6 is designed to complete more useful work with fewer tokens, fewer tool calls, and less time.

For developers building AI agents, coding assistants, research systems, and business automation, that efficiency could be just as important as raw intelligence. GPT-5.6 therefore represents another step toward AI systems that do not just answer questions, but can independently work through complex tasks from beginning to end.

Broadcom saves Open Source - Broadcom has introduced TrueSource, a new suite aimed at helping enterprises securely manage and support the open-source software they depend on. It focuses on verified, human-checked code alongside enterprise support, giving organizations more confidence in their open-source dependencies. Read more.

Miscrosoft jealous of Google - Microsoft Defender for Office 365 is mistakenly flagging legitimate Google Search links as malicious, showing users security warnings when they try to open them. Microsoft has acknowledged the issue as an inaccurate security classification and is investigating the cause. Read more.

Buzz of the Week!

Prompt Caching

Prompt caching is a technique where an AI system reuses the computation from a repeated part of a prompt instead of processing those tokens from scratch every time. It is especially useful for AI agents that repeatedly send the same system instructions, tool definitions, or large context alongside changing user requests. By reusing that previous computation, systems can reduce latency, GPU work, and inference costs. The catch is that caching only helps when requests contain enough repeated context and the cached data remains valid. As AI applications move toward long-running agents with huge system prompts, prompt caching is quietly becoming an important piece of the inference stack.

Things that launched. Things that went viral. Things you'll pretend to try.

pkl

pkl is Apple's configuration language designed to make complex configuration safer, typed, reusable, and easier to validate.

grype

grype scans container images and filesystems for known vulnerabilities. It pairs particularly well with Syft.

cosign

cosign signs and verifies container images and other software artifacts, helping prevent tampering in software supply chains.

Build Braincells, Not Just Features

This weekend’s read: The new Hiring Filter.

This week’s watch: How Did Humans Invent Cooking?

Meanwhile…