---
title: 'In the News: September 23, 2026 (Extra 2)'
description: 'Anthropic prices a Claude Code task on Opus 5.5: on an API key, a pause past five minutes turns a cheap cache read into a full cache write.'
canonical_url: 'https://darkfactory.dev/news/2026-09-23-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-09-23-extra-2.md'
collection: news
date_published: '2026-09-23T16:06:00-04:00'
date_modified: '2026-09-23T16:06:00-04:00'
---

# In the News: September 23, 2026 (Extra 2)


On an API key, Claude Code's prompt cache lasts five minutes by default. Anthropic's cost guide for Opus 5.5 shows what that means at 120K tokens of context: a pause past the lifetime turns the next $0.02 read into a $0.60 write, and changing effort, switching models or connecting an MCP server mid-task can force the same write. Separately, a research paper lets an optimizer rewrite harness code from failure traces and rolls back any change that makes a held-out set worse.

## 1. Anthropic prices a Claude Code task on Opus 5.5, setting by setting

**[What a task costs on Opus 5.5](https://claude.com/blog/what-a-task-costs-on-opus-5-5)** · Addy Osmani, Anthropic · Claude blog, September 22, 2026

Osmani's guide prices a task as a loop in which every turn resends the conversation, so turns, cache reads, output and model choice set the bill. At list prices of $4 per million input tokens, $20 per million output tokens and $0.20 per million cache reads, his illustrative 2.8M-token task costs $11.20 in input with no caching and $1.62 at a 90% cache hit rate. The cache lifetime depends on how you pay: an hour on a subscription, five minutes by default on an API key or cloud provider. "On an API key, a six-minute coffee break turns the next $0.02 read into a $0.60 write," he writes. Changing effort or thinking settings, switching models, connecting or disconnecting an MCP server, or compacting can each force a new cache write.

The guide also names the controls. `CLAUDE_CODE_SUBAGENT_MODEL` puts every subagent on one model, and a model named in a subagent's definition overrides it. Compacting at 150K tokens costs about $0.25 and pays for itself in about ten turns. On one internal 44-ticket customer-support benchmark, moving from Opus 4.8 to Opus 5.5 at low effort cut cost by about 18%, and running `/claude-api prompt-audit` cut a further 9%. Osmani calls that "an example rather than a number to expect." He also says up front that some of the post's figures are list prices and some are illustrations built from them.

**Why it matters:** For teams on API keys, the five-minute cache lifetime turns session habits into a cost line: long pauses, mid-task effort changes and MCP churn each pay a full cache write. The guide's own advice is the right test. Run one real task on each model and compare what `/usage` reports.

## 2. A harness that grows its control code from failure traces

**[Grow the Harness, Not the Context: From Strategy-Free Scaffolds to Reusable Specialist Agents](https://arxiv.org/abs/2609.26760)** · Laizhen Li, Xitong Gao et al., Shenzhen Institutes of Advanced Technology (CAS) and University of Macau · arXiv, September 22, 2026

The paper starts from a scaffold with fixed model and tool interfaces and no task-solving controller. Each round, an optimizer repairs a window of failed tasks at once, and it may edit only the entry function and the functions that appear in those failures' execution traces, within an edit budget. A held-out gate rolls back any sequence of repairs that lowers gate success, restoring the code along with the task cursor and counters. The authors direct parsing, validation, error recovery and stopping into code and keep model calls for semantic reasoning. Against a tool-calling agent on BrowseComp-Plus and WebArena-Verified, with deployment models from 4B to 120B parameters, the abstract reports 76.0 to 91.8% fewer LLM calls and 74.4 to 98.6% lower inference cost. It reports the highest mean success in five of six benchmark and model settings, and a mean 0.7 points behind the best in the sixth. We did not review the full results tables; these figures are the authors' summary in the abstract. Each benchmark's final evaluation uses 50 tasks.

**Why it matters:** For a factory that handles many tasks from one family, the paper's design gives a concrete rule for automated harness edits: scope each change to the code a failure touched, and accept it only if a held-out set does not regress. Both benchmarks are web research and browsing, so whether the approach carries over to coding tasks is untested here.