The one thing
01GitHub Copilot Business and Enterprise admins have until 22 October to choose what new Copilot features do by default.
A new global policy under AI Controls → Copilot sets whether eligible generally available features are Enabled, Disabled, or left to organizations. It also covers the Copilot Code Review policy and the MCP servers in Copilot policy.1 It has no effect until 22 October; from then, every eligible feature still marked Unconfigured follows the global default. Explicit choices are kept and previews stay opt-in.1
The changelog does not say what an enterprise that selects nothing will get.1
Shipped
01Claude Code 2.1.282 ignores telemetry-export variables in project settings
Project and local settings files no longer apply OpenTelemetry variables that turn on export, set its endpoint or capture content, such as CLAUDE_CODE_ENABLE_TELEMETRY and OTEL_LOG_*.2 A repo that switched on export in .claude/settings.json now exports nothing. A new startup notice, /status and claude doctor list the ignored variables.2 Move them to user or managed settings or the environment.
The other telemetry-related line in the same release
With telemetry off on a direct Anthropic API connection, auto mode now uses the server-side classifier by default; CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out.2 The release notes do not mention AGENTS.md.
The conversation
01Hacker News argued that price per token is the wrong number for choosing a model.
- arshxyz, Hacker Newsreader-posted, unverified
Wrote that a cheaper model can spend more tokens on the same task, and that per-token tables ignore cached-token pricing, which can make a "cheaper" model cost more.
- ford, Hacker Newsreader-posted, unverified
Said some models need two to three times the tokens for the same result, and pointed to cost-per-task figures instead.
Anthropic moved the same way on 24 September: refusals that arrive before any output are now billed at the model's normal rates when stop_details.category is bio, frontier_llm or reasoning_extraction.3 If your traffic touches those categories, the rate card no longer tells you the whole bill.
Hacker News was swept in full: 29 stories in the window, 8 above 80 points, with comments. The X layer ran on background only: the newest roster digest (102 accounts) and Reddit digest were collected 24 September 05:30 UTC, about 24.5 hours old, covering the window of No. 38. The newest ai-dev-x keyword file holds a placeholder instead of posts.
Skip this
05Best LLM for every budget, updated daily 169 points for a per-token price table; its own thread lists why per-token is misleading.
Google Project Suncatcher, ML infrastructure in space A research programme with no date or product a developer can use.
LFM2.5-VL-DSpark on the Hugging Face blog A vendor speed post whose benchmark the vendor wins; nothing to pin.
Hackers steering ChatGPT and Gemini to scam numbers A Medium write-up with no vendor confirmation; a poisoning risk for search answers, not for code.
SWE-Prometheus and SWE-PolyVision Two new benchmarks from the same group; neither changes how anyone works this week.
Everything we saw
4141 candidates scanned · 3 used in this issue — the rest, with the reason each one was left out
| Item | Source | Signal | Call |
|---|---|---|---|
| Default enablement of Copilot features for Business and Enterprise | GitHub Changelog | Policy takes effect 22 October | Led the issue |
| anthropics/claude-code v2.1.282 | GitHub Releases | OTel variables in project settings now ignored | Printed as Breaking |
| Claude Platform: pre-output refusals billed in three categories | Claude Platform release notes | Dated 24 September, all platforms | Used in Our take |
| Require proof of presence for high-impact actions | GitHub Changelog | Public preview, EMU with Entra ID only | Narrow scope; watching for the PR-merge version |
| openai/codex rust-v0.157.0 | GitHub Releases | Adds GPT-6 Sol and Luna; removes the ultrafast tier from gpt-5.6-sol | No breaking section; no decision changes |
| pydantic/pydantic-ai v2.50.0 | GitHub Releases | Compatibility notes limited to the new DecisionModel | Affects only early DecisionModel users |
| strands-agents/harness-sdk mcp/v0.3.0, harness-cli/v0.1.3 | GitHub Releases | Deprecates cache_tools and the LlamaAPI provider; new to the watchlist today | Deprecations, nothing removed; watching |
| cline v4.1.21 and desktop v0.0.35; vercel/ai 7.0.114, 6.0.291, 5.0.266; langchain-core 1.6.5, langchain-openai 1.6.6; agent-native v0.1.265 | GitHub Releases | Patch cadence; cline raises js-yaml to 4.3.2 for a parser fix | No decision changes |
| OpenAI: elevated errors on GPT-6 Astra Pro | status.openai.com | 0.8h, impact none, resolved | Weather |
| GitHub: billing information updates; Cursor: Grok Bot degradation | status pages | 3.8h minor, resolved; Cursor minor, open at 06:00 | Weather |
| Vercel Bug Bounty now public | Vercel Changelog | Programme announcement | Off the beat |
| Where Does Exactly-Once Live?; Denial-of-Wallet in tool-calling agents | arXiv | 12 of 132 papers read in the 48h sweep | Useful framing, no measurement that changes practice yet |
| ai-dev-x keyword digest, 2026-09-24 | X digest | File present, contents a placeholder | Source unreachable |
End of feed. That is everything from the window worth your time.
Explore more news, analysis and practical guides on NO FEED.