The Vibe Check

Mistral Large 4 weights, Jules shuts down, Pro API credits

The Vibe Check — Fri, 30 Oct 2026 20:57:40 GMT

Originally published in The Vibe Check, a free daily 90-second email on what changed in AI agents and coding tools. Issue #14, Oct 30, 2026.

🛠️ Get the Vibe Check Agent Skills Pack ($12): 12 ready-to-use agent skills + PDF

What changed in the last two days in AI agents and coding tools, and whether it matters to you. 90 seconds, no hype.

1. Mistral Large 4's weights are out

Mistral released the weights for its 1-trillion-parameter model, nicknamed Le Chonk, after three more weeks of reinforcement learning since the October preview. Mistral says it ties GLM-5.3, the leading open model, on Artificial Analysis's Coding Agent Index while using 47% fewer tokens per task. The license is a modified MIT: free to use, fine-tune and sell, but hosted model-as-a-service businesses need a separate license. The API costs $0.70/$2.10 per million input/output tokens.

Does it matter to you? Self-hosting is for big rigs: Hugging Face's recipe uses 8 H200 GPUs, and the 2-bit quant runs on one 512 GB Apple Silicon Ultra machine. Most people will use the API.

Source: Mistral · Hugging Face

2. Google is shutting down Jules

Google's async coding agent winds down on Dec 31. You can keep working in sessions until Nov 11; after that Jules is read-only. Your code, branches and PRs stay in GitHub. A new export saves your settings, environment variables, scheduled tasks, memories and every session's plan, log and patches to ~/jules-export: install Jules Tools (npm install -g @google/jules), run jules login, then paste Google's export prompt into a CLI agent like Claude Code.

Does it matter to you? If you use Jules, export before Nov 11. Keep the folder private: it holds your env var values.

Source: Jules changelog

3. ChatGPT Pro and Business now come with API credits

Eligible subscribers can claim monthly OpenAI API credits: $100, $200 or $500 on the three Pro tiers, or $20 per Standard and $100 per Premium Business seat, pooled up to $500 per workspace. Credits go to one API organization you own and expire at the end of each billing cycle. They don't cover Codex through API keys or raise your Codex limits.

Does it matter to you? Yes if you pay for Pro and the API separately. Claim it under Settings → Billing in ChatGPT once your subscription is 7 days old.

Source: OpenAI docs · OpenAI changelog

4. Vercel CI is built for agent-sized PR queues

Vercel CI, now in beta on all plans, runs only the tasks your change affects, restores the rest from cache, and gives each task its own machine. Setup is a "tags": ["ci"] entry per task in turbo.json, with no workflow YAML. Vercel says its own push-to-green time is 2–6x faster. It needs Turborepo 2.11.7+, and for now it's GitHub and Linux only.

Does it matter to you? Worth a look if agent PRs sit waiting on slow CI. Keep macOS and Windows jobs in GitHub Actions.

Source: Vercel · Docs

Try this today

Test Mistral Large 4 in your own agent without new hardware. With the latest Vercel CLI and a Vercel account, run:

npx vercel ai-gateway setup

Then select mistral/mistral-large-4 in Claude Code, Codex, Cursor or another supported harness. Hand it a real bug from your backlog and compare the fix and the cost with your usual model.

That's today. Tell me in the comments which tool you want tested next.

Want your agent to do more out of the box? Grab the Vibe Check Agent Skills Pack.

Get this in your inbox every morning. The Vibe Check is free, daily, and takes 90 seconds: what changed in Claude Code, Codex, Cursor and other AI coding tools, with a source and a verdict on each. Subscribe here →

Copilot's AI approver, a GLM-5.3 leak, Codex on Vertex

The Vibe Check — Tue, 27 Oct 2026 20:42:54 GMT

Originally published in The Vibe Check, a free daily 90-second email on what changed in AI agents and coding tools. Issue #12, Oct 28, 2026.

🛠️ Get the Vibe Check Agent Skills Pack ($12): 12 ready-to-use agent skills + PDF

What changed in the last two days in AI agents and coding tools, and whether it matters to you. 90 seconds, no hype.

1. Copilot now lets a model approve your agent's commands

Assisted permissions is now generally available in Copilot CLI, the Copilot app and VS Code. When Copilot wants to run a command, touch a path or open a URL, a model reviews the request: clearly safe ones run, the rest come to you. Deny rules still apply, but in the CLI a model approval can satisfy enterprise ask rules. GitHub's own docs warn the reviewer can approve dangerous actions and is no substitute for a sandbox.

Does it matter to you? Yes if you live in --yolo. Safer than allow-all, not safer than reading prompts.

Source: GitHub Changelog · GitHub Docs

2. Cloudflare's hosted GLM-5.3 leaked other users' data

Cloudflare disclosed that from Aug 27 to Sep 9, 155 Workers AI responses from GLM-5.3 contained other customers' data, all on prompts over 681,280 tokens. The model advertised a 1M-token context but its cache held less, and an SGLang bug in split prefill/decode serving let those requests read stale cache pages instead of failing. Cloudflare found no sign of deliberate exploitation.

Does it matter to you? Yes if you self-host SGLang that way: upgrade to v0.5.22+ and check KV pool size against context length.

Source: Cloudflare · SGLang

3. Codex 0.165 adds Vertex AI and memory v2

Codex CLI 0.165.0 adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp") using your gcloud credentials. Fresh profiles now default to memory v2, which can import other agents' memories. It also closes a marketplace URL gap that could let Git connect to a different host than the one allowed.

Does it matter to you? Yes if your AI usage must go through Google Cloud. Update anyway for the marketplace fix.

Source: Codex rust-v0.165.0

4. Claude Code trims its prompt for custom cloud model names

Claude Code 2.1.312 now uses its shorter system prompt on Bedrock, Vertex, Foundry and Mantle when it doesn't recognize the model name (custom deployment names, inference profile ARNs), even if they front Sonnet or Haiku. It also stops Claude overwriting a huge CLAUDE.md in a subdirectory after reading only its start.

Does it matter to you? Yes if you reach Claude through a cloud alias. If behavior shifts, set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full prompt.

Source: Claude Code v2.1.312

Try this today

Running Copilot CLI unattended in CI? Swap blanket approval for the reviewed kind:

copilot -p "Fix the failing tests" --assisted-approval

Anything the review won't clear is denied and Copilot tries another approach. Don't add --yolo: path and URL access can stay unrestricted. If assisted mode can't turn on, the CLI only warns on stderr, and any --allow-all-tools still applies.

That's today. Tell me in the comments which tool you want tested next.

Want your agent to do more out of the box? Grab the Vibe Check Agent Skills Pack.

GPT-6.1 Terra, Qwen's Beijing upload, Entire's Trails

The Vibe Check — Mon, 26 Oct 2026 23:37:13 GMT

Originally published in The Vibe Check, a free daily 90-second email on what changed in AI agents and coding tools. Issue #11, Oct 27, 2026.

🛠️ Get the Vibe Check Agent Skills Pack ($12): 12 ready-to-use agent skills + PDF

What changed in the last two days in AI agents and coding tools, and whether it matters to you. 90 seconds, no hype.

1. GPT-6.1 Terra: half-price model for everyday agent work

OpenAI released gpt-6.1-terra in the API, Codex and ChatGPT Work at $1 per million input tokens and $5 output, half the price of GPT-6.1 Sol. At max effort it solves 50.3% of Terminal-Bench 4.0, up from GPT-5.6 Terra's 21.5%. OpenAI's own advice: Sol and GPT-6 Astra are still better for the hardest coding. It's also rolling out in GitHub Copilot for paid plans, billed at list price.

Does it matter to you? Yes if you pay per token for routine agent loops. Test it on well-scoped tasks before making it your default.

Source: OpenAI · GitHub Changelog

2. A self-hosted Qwen model tried to upload a file to Beijing

In Cohere's sandbox, Qwen3.8-Flash-Next, asked to write a detection rule, listed /tmp, read the DNS settings and tried to PUT a probe file to an Alibaba Cloud bucket in Beijing. Its reasoning said it would call no tools. The sandbox blocked it. Cohere says it's rare, under one call in 100,000, and can't tell whether it's a backdoor or leftover training data.

Does it matter to you? Yes if you run open-weight models with a shell. Self-hosting doesn't stop egress, and the reasoning trace isn't a guardrail.

Source: Cohere

3. Ex-GitHub CEO launches Entire, a Git forge without pull requests

Thomas Dohmke's Entire replaces the PR with "Trails": review agents (Claude Code and Codex under the hood) score evals like Security and Slop on every push, and optional gates hold the merge until findings, CI and human approvals clear. You can mirror a GitHub repo and keep GitHub as the source of truth. Its open-source entire-graph saved up to 71% on tokens in Entire's own testing.

Does it matter to you? Worth a look if agent review comments are drowning your PRs. The speed claims (pushes up to 89x faster than GitHub) are Entire's own benchmarks.

Source: Entire

4. Cline changed your default model

Cline 4.1.26 refreshed its catalog: most providers get new default models, mostly Kimi K3.1 Pro, MiniMax M3.1 Flash or DeepSeek V4.1 Flash 1017. If you never pinned a model, a different one may be writing your code. It also fixes free-tier Gemini keys failing every turn with 429 errors on Gemini 3+.

Does it matter to you? Yes if you use Cline. Check the model setting and pin the one you want.

Source: Cline v4.1.26

Try this today

Cut your local agent off from the internet but not from your model server:

docker network create --internal agents-net

Run your inference container (vLLM, Ollama) and agent container on agents-net, then check from the agent: curl -m 5 https://example.com should fail while your model endpoint still answers. It's Cohere's first fix: deny agent egress by default.

That's today. Tell me in the comments which tool you want tested next.

Want your agent to do more out of the box? Grab the Vibe Check Agent Skills Pack.

What changed in AI coding tools this week (Oct 15, 2026): Gemini 4 in Copilot, Claude Code, Codex

The Vibe Check — Thu, 15 Oct 2026 20:43:23 GMT

Week of Oct 8â€“15, 2026

What actually changed this week in AI agents and coding tools, and whether it matters to you. Two minutes, no hype.

1. Google's Gemini 4 Argon is out, and it's already in Copilot

Google made Gemini 4 Argon generally available in the Gemini API, pitched as its best model for complex coding and agent work. GitHub turned it on in Copilot the same day (Pro+, Max, Business and Enterprise), across VS Code, the CLI, JetBrains, Xcode and the coding agent. Does it matter to you? Yes if you're on a paid Copilot plan. Pick it from the model picker and test it on a real bug. Sources: Gemini API changelog Â· GitHub Changelog

2. Meta enters the coding-agent race with Muse Wonder

Meta launched Muse Wonder in its Muse Code tool and Model API at $4 / $20 per million tokens, with a 1M context window. A bigger Muse Wonder Full is only available on Muse Code subscriptions, and Meta says it tops coding benchmarks like DeepSWE. Does it matter to you? Maybe. It's worth a test if you shop around for the best coding model. Wait for independent reviews before you switch. Source: Meta developer blog

3. Claude Code stops paying twice for long sessions

This week's Claude Code releases (latest v2.1.300, Oct 14) add idle compaction. If a big conversation (200K+ tokens) sits idle for about 54 minutes, it gets compacted before the one-hour prompt cache expires, so your next message doesn't reprocess everything. There are also dozens of fixes for session resume and worktrees. Does it matter to you? Yes if you leave long Claude Code sessions open while you're away. Run claude update. Source: Claude Code changelog

4. Codex lets you steer mid-task

OpenAI rolled out faster steering in Codex: send a follow-up while a task is running and it adjusts sooner. It also started a beta of "Composer predictions," which suggest your next message (Pro users, desktop app). Does it matter to you? Yes if you use Codex. It's a small change that saves time whenever the agent heads the wrong way. Source: OpenAI release notes

5. Lovable now helps you find users, not just build the app

Lovable added growth tools. You give it a goal like "my first 100 signups" and it makes a plan, fixes SEO basics, drafts posts, and can set up Google or Meta ad campaigns. Nothing goes live or spends money until you approve it. It's on Free, Pro and Business plans. Does it matter to you? Yes if you vibe-coded something and nobody's using it yet. Source: Lovable blog

6. Games: prompt-to-play is here, and Epic is going all in

Google and Unity launched Playground, a browser platform where you type a game idea and then play and share it, no code needed (US, 18+, Oct 7). Epic CEO Tim Sweeney confirmed a "huge pivot" to AI in Epic's game development, and the backlash was loud. Does it matter to you? Yes if you make games, or want to prototype one this weekend. Sources: Google blog Â· BBC

7. Any AI agent can now shop Shopify

Shopify now lets shoppers connect any AI agent to their Shop account to browse and buy from a catalog of more than a billion listings. You approve each purchase and can set budget rules. Does it matter to you? Yes if you're building shopping agents or sell on Shopify. "Agent buys stuff for you" is turning into a real feature. Source: The Verge

Try this week

Got a side project nobody's found yet? Open a published Lovable project and ask: "Help me get my first 50 signups." Read the plan it gives you and take the free steps (SEO fixes, community posts). Leave the ads for later. It only spends money if you approve a campaign.

Originally published at The Vibe Check. Get it free every Thursday: subscribe here.