5 min read · 6 sources

Anthropic just dropped Claude Haiku 5.5, cutting execution costs by roughly 75% compared to Haiku 4.5 while beating or matching previous flagship checkpoints on low-latency workloads. It arrives alongside a 50% price cut to Claude Sonnet 5.5 cache reads, live across AWS, Google Cloud, and Azure.

At the same time, the infrastructure around agentic execution is hardening. Microsoft is moving inference off centralized clusters and straight onto endpoint silicon with unified-memory Windows workstations, while AWS is shipping kernel-level sandboxing tools to keep local agent runtimes from running arbitrary shell code.

Here is what shipped today, what it costs, and what it breaks in your stack.

A LoRA-tuned 2B classifier cut Bedrock injected records from 147 to 25 and dropped median gating latency to 115 ms on an RTX 3090.

Anthropic undercuts mid-tier routing with Claude Haiku 5.5

Source: anthropic.com ↗

The core pitch for Haiku 5.5 is throughput and economics. It runs at a quarter of the price of Haiku 4.5, includes a tunable effort dial for adjusting inference spend against complex reasoning, and posts a 1620 score on GDPval-AA compared to Haiku 4.5’s 735. Anthropic is positioning the checkpoint as a replacement for GPT-6 Luna and Sonnet 5.5 in latency-sensitive pipelines like live support, classification queues, and multi-tenant sub-agent swarms.

For teams building tiered agent patterns, the architecture is straightforward: you point code generation, test iteration, and refactoring tasks at Haiku 5.5, keeping Opus 5.5 or Sonnet 5.5 as the upstream planner. If your system relies on heavy retrieval-augmented generation where token ingestion dominates your invoice, the updated pricing makes Haiku 5.5 the default ingestion engine. Anthropic also halved prompt cache read fees for Sonnet 5.5, which directly rewards stacks running long system prompts and stable tool schemas.

Why your agent pipelines should stop polling long-term memory

Source: blog.mikegchambers.com ↗

Most production agents leak tokens because they blindly query vector databases on every interaction loop. Engineer Mike Chambers published an architecture teardown demonstrating decider-gated memory, using a local 2B-parameter model called strands-decider to evaluate whether an AWS Bedrock agent actually needs external context.

The setup replaces the traditional language model head on a Qwen 3.5-2B torso with a pointer head fine-tuned via LoRA. In testing, the classifier cut Bedrock memory lookups from eight down to three, reducing injected context from 147 database records to just 25 without degrading final output accuracy. Median evaluation latency came in at roughly 115 ms on a desktop RTX 3090 and 153 ms on an Apple Silicon M3 MacBook. The harnesses and code are up on GitHub, offering a drop-in pattern to stop context-window bloating in agentic loops.

Microsoft routes GitHub Copilot across local RTX Spark silicon

Source: commandline.microsoft.com ↗

Microsoft is restructuring GitHub Copilot to treat local hardware as a primary inference tier. Under an initiative dubbed Project HydraFusion, Copilot will automatically partition workloads between hosted infrastructure and on-device GPUs based on current system load, latency margins, and task complexity.

The on-device component relies on a quantized variant of MAI Code 1.1 Flash, targeted at workstations like the Surface Laptop Ultra packing NVIDIA RTX Spark chips. These systems provide roughly 1 PFLOP of compute alongside up to 128 GB of unified memory, allowing large coding models to stay resident in VRAM without thrashing host memory. To let these local models touch the host safely, Microsoft introduced Microsoft Execution Containers (MXC). MXC acts as an OS-level sandbox that isolates local shell triggers, debugging servers, and workspace disk access, cutting network egress while isolating system credentials.

Windows shifts to hybrid agent orchestration with MXC

Source: blogs.windows.com ↗

Expanding on its developer tooling, Microsoft published its broader hybrid intelligence roadmap, establishing Windows as an execution broker for autonomous agents. General availability of Microsoft Execution Containers (MXC) brings native sandbox isolation directly into the Windows runtime, aimed squarely at securing deployments for Microsoft Agent 365 and local developer assistants.

Instead of treating background tasks as standard user-space binaries, the operating system routes agent tooling through containerized boundaries with explicit permission barriers for file paths, environment keys, and network routes. The runtime layer supports both Windows ML drivers and direct llama.cpp execution paths, letting IT teams drop offline agents onto DGX Stations, mini-PCs, and unified-memory hardware without exposing the underlying host OS to unsafe tool operations.

SAP embeds USDC and EURC settlement into core ERP workflows

Source: forkast.news ↗

SAP announced SAP Pay at SAP Connect, integrating stablecoin settlement directly into SAP Cloud ERP. The rollout lets corporate treasury teams settle vendor invoices in USDC and EURC across 89 global payment corridors and more than 40 fiat currencies, treating digital assets like standard clearinghouse balance lines.

The integration sits on Circle’s Arc blockchain infrastructure, wired into SAP by enterprise connectivity provider Tereina. For enterprise systems engineers, this strips out the custom payment bridges and separate ledger sync jobs usually needed to handle stablecoins. Invoices settle directly against general ledger modules, exposing native APIs for cross-border reconciliation. SAP is also pairing the rails with Joule AI agents to enable automated, policy-gated invoice settlement without requiring manual AP sign-offs.

AWS releases Strands Box to sandbox rogue CLI agents

Source: theregister.com ↗

Amazon Web Services launched Strands Box, an open-source isolation framework designed to intercept unsafe agent actions before they hit host environments. The framework wraps execution runtimes - including the Dogwood Local Engine, Strands Shell, and local Python interpreters - in an ephemeral boundary with state-aware policy enforcement.

Unlike simple permission boundaries, the Strands Box policy engine includes temporal awareness, tracking the sequence of execution rather than isolated actions. It blocks anomalous behavior like an agent generating credentials, modifying a file, and attempting an outbound Slack webhook in the same loop. For infrastructure engineers running unconstrained evaluation loops or tool-using agents locally, it provides an audited security perimeter against catastrophic file deletions and unauthorized API updates.