Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Claude Code Review

Claude Code Review artwork

Opus 5.5 arrives at lower cost, Planner and reviewer agents vet payments, OpenAI Codex plugin ships for Claude

Sep 23, 20264 minAudio
ShareXLinkedInFacebookWhatsAppEmail

Your daily Claude Code briefing. What shipped, what changed, and what it means for the things you are building. Edition date: 2026-09-23 Chapters: 00:00 Opus 5.5 lands in Claude Code, matching rival models at a lower price 01:15 Everything Claude Code's planner and reviewer agents can vet payment code before it ships 02:08 OpenAI ships a Codex plugin that reviews code inside Claude Code 02:59 A hook stops Claude Code from installing exploited Python packages 03:50 CodeGraph beat Graphify and cut Claude Code's token use 35% in a head-to-head build test Claude Code Review · claudecodereview.ai

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

Maya: Opus 5.5 went live inside Claude Code today through the v2.1.280 update, and the headline is price. On a same-day benchmark walkthrough it performs roughly on par with Fable 5.1 while costing forty percent less to run.

James: That price gap is real on raw tokens too. Opus 5.5 runs four dollars per million input and twenty per million output, well under Fable 5.1 and GPT6 Astra's shared ten-input, fifty-output rate. Cached reads and writes got cheaper as well.

Maya: And task-level testing backs it up. On Terminal Bench 4.0 at high effort it scored 64.2 percent for three eighty-eight per task, roughly half Astra's price and clearly ahead of Fable, which scored 49.4 percent for ten fifty.

James: That pricing matters for the weekly allowance. Paid plans split quota between Fable and Opus, and Opus 5 was the weaker half, described as effectively wasted. Getting close to Fable output makes the whole allowance useful.

Maya: Anthropic also fixed the dense, jargon-heavy replies people struggled with, saying Opus 5.5 follows explicit writing-style instructions more closely. It carries Fable's guardrails too, rerouting risky cybersecurity or biology questions to Opus 4.8. It's available everywhere today.

James: Shifting to a workflow one. A builder demo shows Everything Claude Code's planner and reviewer agents pointed at payment flows. One agent maps the flow, then deliberately tests a failed charge or a double-click before checkout code ships.

Maya: That's the gap that bites in production. A checkout page looks finished in a demo, then breaks the moment a real payment misfires. This is a free, open-source setup that splits one session into specialized agents.

James: The builder, Affaan Mustafa, spent ten months on a private version before winning one of Anthropic's hackathons, then open-sourced it on GitHub. Reports put it at sixty-eight subagents with a library of nearly three hundred skills, installed with two commands.

Maya: And it isn't locked to Anthropic. A demo shows it working with Codex, Cursor, and Gemini too. The demonstrator's caveat is explicit: the output still needs a human check, and model access is billed separately.

James: Speaking of Codex, OpenAI shipped a plugin that installs into Claude Code through Anthropic's marketplace. It surfaced in a demo posted September 20, letting Codex critique code Claude just wrote without leaving the session.

Maya: Installing it takes four typed commands: add the marketplace source, install the plugin, reload, then run setup. After that, typing slash codex review hands whatever Claude produced over to Codex for a second look.

James: The pitch is that the two models have different blind spots from how each was trained, so one reviewing the other catches mistakes a single tool would miss. It reframes the Claude-versus-Codex debate as run both.

Maya: With Codex cast as the critic. And because it uses Claude Code's existing add-on system, there's no separate download or account to manage, just the same handful of commands as any other plugin.

James: Now a safety one. Claude Code will happily run pip install or uv add for whatever package it decides it needs. A new open-source hook called Package Doctor intercepts install commands and blocks the risky ones.

Maya: And the filtering is narrower than a typical scanner. It ignores plain age and raw advisory counts. What triggers a block is a flaw appearing on CISA's list of vulnerabilities known to be actively exploited.

James: That precision is the point. When it blocks pillow ten zero zero, it names a fixed version, pillow twelve three zero, so Claude can act instead of retrying the same failed command. It runs as a PreToolUse hook, no API keys needed.

Maya: And it fails open. If PyPI or the OSV database is unreachable, the install still goes through and the hook reports the package went unchecked, rather than blocking you on someone else's outage.

James: Last one is a head-to-head test posted September 16. CodeGraph and Graphify went against plain Claude Code on the same feature build. CodeGraph indexed the project in under two seconds using zero model tokens, then let Claude skip repeated file rereads.

Maya: And it showed. CodeGraph's full run finished in eight forty-five, planning forty percent faster than baseline, using 4.23 million tokens, thirty-five percent below plain Claude. It also passed both timing checks cleanly, while Graphify's export ended one frame early.

Maya: That's your read on where Claude Code is heading this week, here at Claude Code Review. We'll be back with more as the tooling keeps moving.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Claude Code folds a simulated iPhone Duo, Anthropic's project helper answers in the chat

    Nov 18, 20264 minAudio

    Claude Code news and practical tips: what shipped, what changed, and how to use it in your own work. Edition date: 2026-11-18 Chapters: 00:00 Claude Code Folds and Unfolds a Simulated iPhone Duo to Test App Layouts...

    Credit the people on this episode

  2. Make Every Claude Code Change Pass Automated Tests

    Nov 11, 20264 minAudio

    Claude Code news and practical tips: what shipped, what changed, and how to use it in your own work. Edition date: 2026-11-11 Chapters: 00:00 JPMorganChase Runs Claude Code on Its Own Systems, With Thousands of...

    Credit the people on this episode

  3. Claude Code cuts long-session lag 35%, Cursor users told to check for a leaked key, edits now locked to your project

    Nov 4, 20263 minAudio

    Claude Code news and practical tips: what shipped, what changed, and how to use it in your own work. Edition date: 2026-11-04 Chapters: 00:00 Claude Code Starts Replies 35% Sooner in Very Long Chats 01:01 Cursor Users...

    Credit the people on this episode

  4. Self-hosted runners stay reserved longer, CLI adds Tasks and Elicitations, Scripts answer multiple choice prompts

    Oct 28, 20265 minAudio

    Claude Code news and practical tips: what shipped, what changed, and how to use it in your own work. Edition date: 2026-10-28 Chapters: 00:00 Self-hosted Claude Code runners can stay reserved until background work...

    Credit the people on this episode

  5. Hand a long coding job to Claude from your phone and let it keep running with your laptop closed

    Oct 21, 20264 minAudio

    Your daily Claude Code briefing. What shipped, what changed, and what it means for the things you are building. Edition date: 2026-10-21 In this episode Hand a long coding job to Claude from your phone and let it keep...

    Credit the people on this episode