How to Cut Claude Code Costs With GLM 4.5

TL;DR

Z.ai's GLM coding plan lowers Claude Code costs by redirecting the tool to GLM 4.5 through a custom API base URL. The $3 plan offers up to 120 prompts per five-hour cycle, while the $15 plan offers up to 600, but model selection, configuration reliability, output quality, and private-code security require careful evaluation.

Transcript

It's clear you guys really like clawed code, but if I've learned anything from my comment section is that you guys don't like the new pricing, especially after moving from cursor. I know so many of you guys made the move because cursor got more expensive just to be rugpulled again with Claude. What if I told you there was a way to save 85% of your ... Read More

Key Insights

• Z.ai's coding plan works directly through Claude Code by overriding its API base URL, so users can retain Claude Code's familiar interface while sending requests to GLM 4.5 instead of adopting a separate command-line tool.
• The base plan costs $3 per month and permits up to 120 prompts during each five-hour cycle, while the $15 plan permits up to 600 prompts in the same period, according to the plan details shown during testing.
• The pricing comparison presented claims that the $3 plan provides three times the usage of Claude Pro, while the $15 plan provides three times the usage of Anthropic's $200 Claude Max plan.
• The published CCbench comparison reports that GLM beats DeepSeek V3.1 in 52 percent of comparisons and loses 36 percent, while its comparison against Claude Sonnet 4 shows 40 percent wins, 50 percent losses, and 10 percent ties.
• The central tradeoff is lower cost versus lower benchmark performance, because the presented comparison places GLM about 10 percentage points behind Claude Sonnet 4 while offering savings described as three to seven times the cost.
• The interface-generation tests produced usable results, with the second GLM attempt judged better than the first and somewhat similar to an Opus result, although the reviewer still disliked several sizing and visual-design decisions.
• The configuration can silently select the wrong model when an older explicit model setting remains active, as one test used the configured Claude model until the reviewer changed the setting to default and confirmed GLM 4.5 through a direct question.
• Private-code handling is a significant concern because using the integration sends project context to a different endpoint, and the reviewer was unwilling to test it freely on large, non-open-source codebases without greater trust in that endpoint.

Questions & Answers

Q: How can GLM 4.5 be used with Claude Code?

GLM 4.5 can be used through Z.ai's coding plan by changing Claude Code's API base URL and API timeout, then supplying the Z.ai API key. Claude Code supports base URL overrides, so the integration preserves its normal interface while routing requests elsewhere. Existing explicit model settings should be changed to default, because an older model selection may override the GLM configuration.

Q: How much does the Z.ai GLM coding plan cost?

The plan details shown include a $3 monthly tier with up to 120 prompts per five-hour cycle and a $15 monthly tier with up to 600 prompts per five-hour cycle. The reviewer says the base tier provides three times the usage of Claude Pro, while the $15 tier provides three times the usage of Anthropic's $200 Claude Max plan.

Q: Is GLM 4.5 as capable as Claude Sonnet 4 for coding?

The published CCbench figures shown do not place GLM 4.5 ahead of Claude Sonnet 4 overall. GLM wins 40 percent of their comparisons, Claude wins 50 percent, and 10 percent are ties. The reviewer frames this as roughly a 10-point performance tradeoff, but finds GLM's generated interfaces usable enough to make its much lower price attractive.

Q: What coding quality did GLM 4.5 produce in the tests?

GLM 4.5 produced a usable image-studio interface inside a basic Next.js application, and the reviewer considered the first result one of the better outputs seen from similar tools. A second generation looked even better and somewhat resembled an Opus result. However, later revisions still had questionable sizing, visual hierarchy, and file-editing behavior, so the experience was not consistently polished.

Q: How do you verify that Claude Code is actually using GLM 4.5?

Verification requires more than trusting Claude Code's displayed model label, because the interface may still show a Claude name while requests are routed to GLM. The reviewer checked provider usage, inspected settings, changed an explicit model configuration to default, and asked the running model to identify itself. After the configuration change, it answered that it was powered by GLM 4.5.

Q: Why might Claude Code ignore the GLM 4.5 configuration?

An existing model setting can take precedence over the new integration. During testing, an older Claude Code configuration caused one run to use the explicitly selected Claude model even though the API override was present. Switching the model setting to default allowed the routing configuration to choose GLM 4.5 consistently. Users with older or customized installations should therefore inspect their settings carefully.

Q: What problems occurred while testing GLM through Claude Code?

The testing exposed several workflow problems. The agent repeatedly launched development commands, left a previous development server running, and once reported an error while editing files. Model identification was also confusing because Claude Code displayed a Claude label even when GLM was active, and an old configuration caused one comparison to use a different model than intended.

Q: Is it safe to use the GLM integration with private codebases?

The reviewer does not establish that the integration is safe for confidential projects. Instead, he expresses reluctance to send large, real, non-open-source codebases to an unfamiliar Chinese endpoint. The installation script was inspected and found unobfuscated with nothing suspicious noticed, but checking that script alone does not resolve broader concerns about where private source code and prompts are processed.

Summary & Key Takeaways

• Z.ai integrates GLM 4.5 directly with Claude Code instead of requiring users to adopt a separate command-line tool. Its setup changes the API base URL and timeout, allowing the familiar Claude Code interface to communicate with another model provider while retaining the workflow users already know.

• The pricing presented is $3 monthly for up to 120 prompts per five-hour cycle or $15 monthly for up to 600 prompts. The reviewer considers this substantially cheaper than Anthropic's plans, especially when GLM can produce usable interfaces, although the value depends on acceptable quality and reliable configuration.

• Testing produced mixed but promising results. GLM generated usable interfaces and improved on a second attempt, yet it sometimes ran unwanted development commands, encountered file-editing trouble, and created confusion about which model was active. The reviewer also hesitated to send private, production code to an unfamiliar endpoint.