Daily Vibe - 10-18-26 · Vibe Coders

12d • News

Daily Vibe - 10-18-26

ForgeCode is telling everyone it beats Claude Code and Codex. The benchmark that's actually independently verified says something much smaller, and the real story buried underneath is about scaffolding, not which model is smarter. Also: Jev's two-week numbers landed, and WebWork gave assistants 120 ways to touch your timesheets.

ForgeCode's benchmark win comes with a big asterisk

ForgeCode, a terminal coding harness from Tailcall Inc., is topping TermBench 2.0 with an 81.8% score, and its marketing has run hard with that number as proof it beats Claude Code and Codex. One problem: TermBench 2.0 lives at tbench.ai, and ForgeCode submitted its own result. That's a self-reported score, not an independent one.

Switch to SWE-bench Verified, the benchmark Princeton and UChicago actually run and hand-check, and the gap shrinks to almost nothing. ForgeCode paired with Claude 4 scored 72.7%, a couple points off top Codex and Claude Code setups. That's not the blowout the headlines suggested.

The genuinely useful finding here isn't about ForgeCode winning or losing. Tailcall also showed that wrapping Gemini 3.1 Pro in ForgeCode's own orchestration layer, instead of a generic scaffold, pushed that same model's Terminal-Bench 2.0 score from 55% to 80.2%. Same weights. Same model. A 25-point swing just from changing how the tool plans and retries. Scale AI found the same pattern running three agent systems on identical Claude Opus 4.5 weights: scores ranged from 50.2% to 55.4% on SWE-bench Pro. The orchestration around a model moves the needle about as much as the model itself does.

To be fair, ForgeCode isn't vaporware. Its GitHub repo has around 7,380 stars and 1,438 forks, with commits landing within the last day, and it's model-agnostic enough to run Claude, GPT, or Gemini depending on what you already pay for. That's a real project with a real community behind it. What's missing is any sign that working teams are actually ripping Claude Code or Codex out of production to replace them with it. Stars measure curiosity, not switched workflows.

[YOUR TAKE: have you actually compared ForgeCode against your daily driver, or is this one staying on the "cool, noted" shelf?]

Jev's ecosystem keeps growing, and so does the receipts pile

Jev launched September 15 as a decision model, not a chatbot: you give it a state and typed questions, it returns a label, a score, or a yes/no probability, no generated text to parse. In about four weeks, at least 20 competitors shipped from Cloudflare, OpenAI, Perplexity, Liquid AI, AWS's Strands Labs, and independent devs. Cloudflare's own run has its Clef model beating Jev 94.20 to 79.74 on BANKING77 F1, while Jev still wins GPQA Diamond 78.3 to 48.0. Most of these numbers are self-reported on different benchmark suites, so don't stack them against each other like they're the same test. Full rundown.

The adoption numbers are the part with real teeth. Per Vercel's own data, 35% of paid teams on its platform started using Jev within 14 days of release, nearly double the reach of Sonnet 5.5, which Vercel calls the fastest two-week adoption in its history. Vercel itself swapped Jev in for GPT-6 Luna to classify malicious commands. Same report: open-source models now make up over 70% of tokens running through Vercel's gateway and close to 25% of spend. Details here.

Your AI assistant can now approve timesheets

WebWork Time Tracker shipped an MCP server exposing 120 tools across 21 areas: timesheets, leave requests, tasks, teams, screenshots, the works. Ask Claude, Gemini, or ChatGPT (through developer mode) to approve this week's timesheets or flag who was late, and it happens in the chat window instead of a dashboard. Every action runs under the signed-in user's existing role, so someone who can't approve timesheets in WebWork's app can't do it through the assistant either, and deletions need explicit confirmation. Gartner's projecting 40% of enterprise apps will ship task-specific AI agents by the end of 2026, up from under 5% in 2025, which makes permission design like this worth studying now rather than after something breaks. Read the full piece.

What this means for you

Before you trust any "beats Claude Code" claim, find the benchmark that's independently run and hand-checked, not the one the vendor hosts and grades itself.
If a tool shows a huge score jump on an unchanged model, that's a scaffolding story, not a model story. Worth knowing before you credit the wrong layer for the win.
Decision models like Jev are cheap for a reason: no generated output to bill. If you're running classification at volume, that's the category to shop in right now.
If you're building or wiring up an MCP server that can act (not just read), copy WebWork's homework: role-scoped access, confirmation on deletes, and no shared admin key.

What's the last benchmark claim that made you go "wait, who ran this"? Drop it below.

Sources