Grok for Startups is live: 6–12 months of credits for funded teams
The catch: venture backing and a live product. Plus a 2.6B edge model, Copilot’s November model cull, and OpenAI’s unverified math-proof claims.
Copy markdown
Real credits, a rumored 1:1 match
xAI opened Grok for Startups: venture-backed teams with a live product get roughly 6–12 months of credits spanning Grok, Grok Build, Grok Bot and the API from one shared balance, released 20% on approval and 80% after onboarding (six-month expiry). The viral “every dollar matched 1:1” line is all over launch posts but isn’t in xAI’s official terms — apply for the base credits, treat the match as unconfirmed, and note bootstrappers and personal-email signups don’t qualify.
A 2.6B model small enough to run on your laptop
Liquid AI’s LFM 2.5 2.6B (Oct 24, per release trackers) pairs a 66K context with hosted pricing around $0.35/$0.50 per million input/output tokens — and at 2.6B parameters it runs on-device. Weekend build: a private, offline-capable assistant, router or classifier that never sends user data to a server.
Copilot culls more models Nov 16 — repin now
GitHub Copilot’s next deprecation round hits November 16: Opus 4.8 and Opus 5 fold into Opus 5.5, and GPT-5.6 Terra moves to GPT-6.1 Sol or GPT-6 Luna. If any of those are pinned in an extension, CI job or saved workflow, switch the pin before mid-November so your automations don’t silently fail.
Claude Code now refuses rm in bypass mode
Claude Code 2.1.309 denies rm/rmdir immediately (with a rewrite hint) when you’re in bypass-permissions mode, so unattended runs keep moving instead of hanging on a blocked delete, and sed -i now gets the normal Bash prompt rather than slipping through. If you run Claude Code headless, expect deletions to be refused outright — move files to a trash folder instead.
OpenAI claims three AI-made proofs — nothing ships
OpenAI says unnamed internal models produced three results — the local smoothing conjecture in all dimensions (also resolving Kakeya), a disproof of the Exponential Time Hypothesis, and an existence proof for 4D Yang–Mills, a Clay Millennium problem — formalized in Lean on Oct 16–19. The math and CS writeups and Lean files are public to inspect; the physics artifacts are withheld and no model is callable. What this means for you: nothing to use today, but if your crypto or algorithms lean on ETH-style hardness, track the ongoing verification by Impagliazzo, Paturi and Williams.