---
title: "AI engineering, October 15, 2026"
description: "The Forward Pass daily issue, October 15, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-10-15
updated: 2026-10-15
---

# AI engineering, October 15, 2026

The Forward Pass daily issue, October 15, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-15

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_Meta opens Muse Wonder developer access, GlobalFoundries readies New York fabs, Anthropic ships Haiku 5.5 across clouds, Liquid AI releases LFM2.5-24B-A2B_

Haiku 5.5 hides its thinking text by default. Those tokens still count toward your usage.

Meanwhile, Meta opens Muse Wonder to developers through an SDK and pay-as-you-go API. GlobalFoundries was selected to build and operate two new fabs in Malta, New York, as part of a $21 billion US-Japan initiative.

If you read one thing today: Anthropic's Haiku 5.5 rollout. Use its migration details to update your thinking controls and message handling before switching to `claude-haiku-5-5`.

## Meta's Muse Wonder models use 10% fewer output tokens than Muse Spark

**Top News** · Reported · thelec.net

Muse Wonder gives you access to Meta's latest agent-focused model through an SDK or pay-as-you-go API. Its Active Memory retains and retrieves context from earlier conversations.

Here's what you get:
- Muse Wonder costs $4 per million input tokens and $20 per million output tokens.
- Active Focus tries alternative approaches when a task runs into obstacles.
- Artificial Analysis scored Muse Wonder at 52 and Muse Wonder Full at 56 on its Intelligence Index at the highest reasoning setting. Gemini 4 Argon scored 53, while Muse Wonder Full tied Claude Sonnet 5.5 for second place at evaluation time.
- Artificial Analysis found both versions used about 10% fewer output tokens than Muse Spark.

One catch: Muse Wonder Full has no standalone API or SDK for external access. It is restricted to Meta's dedicated environment, with access through a paid Muse Code subscription.

Sources: [thelec.net](https://www.thelec.net/news/articleView.html?idxno=14550)

## GlobalFoundries plans New York fabs targeting 900,000 wafers a year

**Top News** · Reported · datacenterdynamics.com

GlobalFoundries is preparing to build and operate two new chip fabs at its Malta, New York, site under a $21 billion US-Japan initiative. The first site is permitted and shovel-ready, with construction expected to begin later this year.

Here's what is planned:
- The two facilities are expected to produce around 900,000 300mm wafers annually.
- Japan will fund the expansion, with an initial $10 billion for the first fab and a further $10 billion investment package planned for the second.
- GlobalFoundries will reallocate $995 million from its existing CHIPS and Science Act award to silicon photonics, silicon-germanium, and advanced packaging in Malta and Burlington, Vermont.

One catch: The first fab is expected to come online in 2029.

Sources: [datacenterdynamics.com](https://www.datacenterdynamics.com/en/news/globalfoundries-to-build-and-operate-two-new-chip-fabs-in-ny-as-part-of-21bn-us-japan-deal/)

## Anthropic ships Haiku 5.5 through Claude API and cloud platforms

**Top News**

Haiku 5.5 is Anthropic's model for high-volume, latency-sensitive work. It is available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

Here's what changed:
- Haiku 5.5 supports a 1M-token context window and up to 128k output tokens.
- Adaptive thinking is enabled by default, with effort controls. Set `thinking.display` to `"summarized"` to receive thinking text, which counts toward token usage.
- OpenTelemetry export is in beta for tracing Messages API requests and Managed Agents sessions.
- Scheduled Claude Managed Agents deployments are available on Claude Platform on AWS.

One catch: Migration requires checking existing calls. Manual extended thinking, some sampling parameters, assistant message prefill, and the older computer-use tool can return 400 errors. Haiku 5.5 also lacks structured outputs on Bedrock.

Try it: Use `claude-haiku-5-5` through the Claude API.

Sources: [platform.claude.com](https://platform.claude.com/docs/en/release-notes/overview)

## Signals

1. [Liquid AI releases LFM2.5-24B-A2B reasoning model running at 184 tokens/s on Apple M5 Max](https://www.liquid.ai/blog/lfm2-5-24b-a2b) · 137 HN points
2. [LangChain ships Deep Agents Deploy with one-command deployment for model-agnostic agents](https://www.langchain.com/blog/deep-agents-deploy-an-open-alternative-to-claude-managed-agents)
3. [HiRO-ACEv2 simulates extreme weather at 3 km resolution for about 1/300th of X-SHiELD’s cost](https://allenai.org/blog/hiro-acev2)
4. [CODE-KIT cuts coding agent boundary violations by 52% in smolagents](https://huggingface.co/papers/2610.18206) · 9 Upvotes
5. [Databricks cuts Qwen 3.8 27B fine-tuning time from 7h 44m to 1h 17m](https://www.databricks.com/blog/reducing-qwen-training-time-6x-ai-runtime)
6. [diagram-design gives coding agents 45 diagram types with self-contained HTML and SVG output](https://github.com/cathrynlavery/diagram-design) · 183 Stars today