---
title: 'In the News: October 31, 2026 (Midday)'
description: "Replit's ViBench 2.0 has the best agents passing every test plan in 49 to 65 percent of builds, and most failures sit where features meet."
canonical_url: 'https://darkfactory.dev/news/2026-10-31-midday'
markdown_url: 'https://darkfactory.dev/news/2026-10-31-midday.md'
collection: news
date_published: '2026-10-31T13:20:00-04:00'
date_modified: '2026-10-31T13:20:00-04:00'
---

# In the News: October 31, 2026 (Midday)


Replit's new long-horizon benchmark finds that the strongest coding agents build a whole product correctly in roughly half to two-thirds of attempts, and that most of the misses come from features interacting. Separately, a Claude Code release puts a removal date on Opus 4.8 fast mode.

## 1. Replit's ViBench 2.0: the best agents pass every test plan in 49 to 65 percent of builds

**[ViBench 2.0: A Benchmark for Long-Horizon Software Building](https://frontier.replit.com/vibench-2-long-horizon-software-benchmark/)** · Preeya Kirani, Daniel Furman, Peter Zhong and Michele Catasta, Replit · Replit, October 28, 2026

In ViBench 2.0 the agent starts from an empty project, builds a first version of a product from a short spec, then adds up to eighteen more features on top of its own code. The benchmark has seventeen products, 241 feature requests and 544 test plans. In Replit's account of Fireworks' official runs, GPT-6.1 Sol passed every test plan in 65% of builds, GPT-6 Astra in 60% and Opus 5.5 in 49%. Sol cost about $12 per build and Astra about $90. DeepSeek V4.1 Flash passed 77% of test plans for about $2.50 per build, but no open-weight model passed a full build.

Replit reports that nearly three-quarters of the top models' failed builds fail only on intersections between features, such as undo taking back a teammate's change. In about one in 40 of those builds the product never starts on an empty database, because schema code references a table before the statement that creates it. Grading is done by an agent running written test plans in a real browser. Replit used Opus 5.5 as the grader and says it never passed a plan with one of 34 planted, visible bugs, and that GPT-6.1 Sol confirmed 99.2% of 477 of its passes. Eight of the seventeen products and the harness are public. All figures are Replit's; the full methodology write-up is listed as coming soon, and we did not rerun the benchmark.

**Why it matters:** If your agents add features to code they wrote earlier, test the seams between features and a clean start on an empty database, not each feature alone. The grading setup, a browser-driving agent working from written test plans and checked against planted bugs, is a pattern you can copy into your own verification loop.

## 2. Claude Code 2.1.316 deprecates Opus 4.8 fast mode, with removal on November 30

**[Claude Code v2.1.316 release notes](https://github.com/anthropics/claude-code/releases/tag/v2.1.316)** · Anthropic · GitHub, released the evening of October 30, 2026

The release notes state that fast mode for Opus 4.8 is deprecated as of October 30, 2026 and will be removed on November 30, 2026. After removal, fast mode requests on Opus 4.8 return an error and do not fall back to standard Opus 4.8. The notes say to migrate to Opus 5.5 to keep the speedup. The same release adds `CLAUDE_CODE_SUBAGENT_EFFORT`, which sets the effort level subagents run at when nothing else names one, instead of inheriting the session's.

**Why it matters:** An unattended job that pins fast mode on Opus 4.8 will stop working on November 30 rather than slow down, so search your configs and scheduled runs for it now. The new subagent effort variable gives you a direct lever on what delegated work costs.

## 3. Martin Alderson on making agents fast: check thinking budgets first

**[Making agents fast - what I've learnt over the last 12 months](https://martinalderson.com/posts/making-agents-fast/)** · Martin Alderson · martinalderson.com, October 29, 2026

Alderson applies web-performance habits to agent workflows. He argues that inference speed is only part of it: with provisioning, tools and suspend and resume unchanged, he says halving inference time takes about 20% off a task, a figure he labels rough. His ranking of fixes puts thinking budgets first: "Check thinking budgets! This is overwhelmingly the lowest hanging fruit." A fast-tier model at maximum thinking, he writes, can be much slower than a slower tier at low or medium thinking.

Other points are practical. A `for` loop with `sleep` in an agent's trace is a red flag that it is polling for something that should notify it. Agents chaining `xargs -I{}` over many files can start hundreds of thousands of processes. Shell startup files such as `.bashrc` matter when an agent starts a new shell per command, and broken IPv6 in sandboxes adds delay to each request. The post is experience, not measurement, and gives no benchmarks.

**Why it matters:** Before paying for a faster tier or a new provider, read your agent's tool-call traces and your thinking settings. They are cheaper to fix and, by this author's account, often the larger share of the wait.