すべての号
ニュースレター·November 8, 2026·読了 4 分
Haiku now narrates your subagents (99% cheaper) — and Inkling 2 Small drops open weights next week
Quiet week for flashy launches, loud week for the plumbing: coding agents got cheaper to run, open-weight models kept stacking up, and the money side of the industry (hello, Anthropic IPO chatter) got weird. Here's what I'd actually pay attention to if you ship code with AI.
## This Week in AI Coding
- **Claude Code cuts the cost of watching its own subagents.** Per the [Latent Space AINews recap](https://latent.space/p/ainews-it-s-swapping-season/), Haiku 5.5 now writes the 30-second status updates for background subagents. Anthropic's claim: about 99% cheaper per update and roughly 5% less of your usage limit per session, on v2.1.316+. Small change, but if you run long agent sessions this is where the quiet token burn was.
- **Revolut reportedly self-hosts a model to cut AI coding costs ~90%.** Same recap, relaying The Register: GLM-5.3 on 56 B200 GPUs for about 2,000 engineers. Second-hand, so treat the number as a claim. The direction is real though: token bills are now a line item people engineer around.
- **Cognition's SWE-2 Flash.** Reportedly 20% of the cost of SWE-2 and about 3x faster, post-trained on Thinking Machines' Inkling-Small, and free through November for Devin Pro, Max and Teams. Weights aren't public, so benchmark it on your own repo before believing the speed-up.
- **Inkling 2 Small is the open-weights story of the week.** A 594B MoE with 17B active; [AINews](https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/) reports ARC Prize numbers of 81.7% on ARC-AGI-2 at $0.27 per task, with Apache 2.0 weights expected next week. Also in that write-up: John Schulman said early checkpoints took unexpected actions to circumvent guardrails during training and evals. Worth remembering whenever someone says "just let the agent run."
- **Talent and consolidation.** Cat Wu left the Claude Code team to lead Core Products PM at OpenAI; Factory is acquiring HumanLayer and Meta is acquiring Conductor (all relayed by AINews, so wait for the primary announcements before quoting them).
- **Can a local model read a neural net's weights?** Simon Willison [ran a paper's weight-interpretation prompts on Qwen 3.8 27B](https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/) on his laptop. It got Breakout (88% confidence, ~52 minutes), Pong and Pac-Man roughly right, and failed Freeway. His takeaway is the sane one: interesting, not "the age of source code is over."
- **The Anthropic IPO guessing game.** AINews [rounds up](https://latent.space/p/ainews-anthropic-ipo-whither-now/) a reported $80B run rate against a $100B year-end target, plus falling prediction-market odds on an IPO. Mostly speculation, but it matters if your stack depends on one vendor's pricing staying friendly.
## הדופק / The Pulse
**What changed this week**
- [Harness] Claude Code shipped three releases, v2.1.317 to v2.1.319 (2026-11-03 to 2026-11-06). **Verified** (official release notes): https://github.com/anthropics/claude-code/releases/tag/v2.1.319
- [Harness] Codex CLI 0.167.0 went stable (2026-11-06), with 0.169 alphas already out. **Verified** (official release): https://github.com/openai/codex/releases/tag/rust-v0.167.0
- [Harness] Haiku 5.5 now writes Claude Code's subagent status updates, ~99% cheaper per update (2026-11-05). **Unverified**: Anthropic's claim as relayed by a newsletter, not checked against the changelog: https://latent.space/p/ainews-it-s-swapping-season/
- [Models] Inkling 2 Small (594B MoE, 17B active), Apache 2.0 weights expected next week (2026-11-06). **Unverified**: benchmark figures are third-party posts relayed by AINews: https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/
- [Models] Qwen 3.8 27B tested locally on reading model weights (2026-11-03). **Verified** as his own experiment, not a benchmark: https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/
**From our channel:** the video feed was unreachable this week, so no list here. Everything is on the [Vibe Coding Academy channel](https://www.youtube.com/@thevibecodingacademy).
**AI jobs this week**
- Member of Technical Staff (Software Engineer, AI Policy and Strategic Initiatives), Perplexity (San Francisco, onsite), $220,000 to $405,000: https://llmhire.com/jobs/member-of-technical-staff-software-engineer-ai-policy-and-strategic-initiatives-perplexity-3cd1c9?utm_source=vca&utm_medium=feed
- Staff+ Software Engineer, STEM, Anthropic (San Francisco, New York, Seattle; onsite): https://llmhire.com/jobs/staff-software-engineer-stem-anthropic-901008?utm_source=vca&utm_medium=feed
- Applied AI Engineer, Enterprise - EDU, OpenAI (San Francisco, onsite), $251,000 to $278,000: https://llmhire.com/jobs/applied-ai-engineer-enterprise-edu-openai-cf90e5?utm_source=vca&utm_medium=feed
- Forward Deployed Engineer - ML (London), Modal (onsite): https://llmhire.com/jobs/forward-deployed-engineer-ml-london-modal-f870d6?utm_source=vca&utm_medium=feed
- Sr. Model Engineer, Imagine, xAI (Palo Alto, onsite): https://llmhire.com/jobs/sr-model-engineer-imagine-xai-975007?utm_source=vca&utm_medium=feed
403 new roles this week, 5,648 tracked (LLMHire API, 2026-11-08). More at [llmhire.com](https://llmhire.com).
The Pulse, all week: https://www.vibe-coding.academy/he/pulse
New to this? Start with the free first lesson: https://www.vibe-coding.academy/he/start
## Tool of the Week
**Claude Code itself, specifically the subagent status change.** If you run background agents, update to v2.1.316 or later and watch your usage meter for a few sessions. If Anthropic's numbers hold, it's a free win. If they don't, tell me.
## Quick Links
- [AINews: It's Swapping Season](https://latent.space/p/ainews-it-s-swapping-season/)
- [AINews: Inkling 2 Small](https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/)
- [Simon Willison: Qwen 3.8 27B on weights](https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/)
- [Claude Code v2.1.319 release](https://github.com/anthropics/claude-code/releases/tag/v2.1.319)
— Guy
[endofcoding.com](https://endofcoding.com) · [vibecodingebook.com](https://vibecodingebook.com) · [llmhire.com](https://llmhire.com) · [Vibe Coding Academy](https://www.vibe-coding.academy/he/) · [YouTube @thevibecodingacademy](https://www.youtube.com/@thevibecodingacademy)
すべての号に戻るEnd of Coding ニュースレター