Z.AI remains a low-cost coding subscription candidate to test when supported tools fit. New GLM Coding subscriptions use token-based credits; active legacy plans can retain their existing calculation during the applicable cycle. The plan overview, checked October 7, 2026, lists GLM-5.3 and GLM-5.3-Flash; requests for GLM-5.2 or GLM-5.1 route to GLM-5.3, and GLM-4.7 requests route to Flash. The GLM-5.2 evidence below remains historical, not an evaluation of those backends.

It is not a universal replacement for GPT or Claude. The useful claim is narrower: test the current GLM lanes on your own repository before relying on subscription savings. Keep Claude Opus 5.5 or OpenAI Sol available for final arbitration until that lane passes your tasks.

Start here: Subscribe through Z.AI

What You Get

The GLM Coding Plan is a subscription package for AI-powered coding, not a general unlimited API plan. Z.AI documents support for coding tools including ZCode, Claude Code, Codex, Pi, OpenCode, Cursor, Cline, TRAE, Qoder, Droid, Kilo Code, Roo Code, Crush, Goose, and Eigent. OpenClaw is documented separately as a supported general-agent path with best-effort scheduling. Subscription usage is limited to officially supported tools, products, and endpoints.

Supported plan models currently include:

ModelPractical role
GLM-5.3Flagship lane to evaluate for harder work
GLM-5.3-FlashLower-credit lane to evaluate for routine work

The important operational detail: Z.AI says GLM Coding Plan quota only applies in supported tools and products. API or SDK usage outside the Coding Plan path can be restricted or billed differently.

Why This Plan Is On The Value Shortlist

The plan is useful because it gives you a paid coding lane that is cheaper than keeping every agent loop on GPT or Claude API pricing. It is not free, and it is not unrestricted. The value case is:

Buyer questionZ.AI answerCaveat
“Can I run Claude Code-style workflows cheaper?”Yes, if your tool is supported and the current GLM lane is good enough for your repo.Unsupported tools and direct SDK/API usage do not get Coding Plan quota.
“Which current lane should I test?”Compare GLM-5.3 with Flash on accepted repo tasks.The flagship deducts more credits; lower deduction is not proof of lower cost per accepted result.
“Is this a replacement for Opus or GPT?”It can replace many low-risk coding loops if it passes your evals.Keep Opus 5.5 or Sol for high-risk changes and final arbitration.
“Is the invite discount guaranteed?”Official campaign rules currently list 10% for eligible first GLM Coding orders.Checkout controls the real price before payment.

Pricing And Metering

Z.AI’s July 30 plan update separates two systems:

  • New plans: token usage is translated into visible credits, including input, cached-input, and output contributions. Previous plans are no longer sold to new users.
  • Legacy Plan V1 and Team plans: current price, benefits, limits, and calculation remain unchanged through the active billing cycle; users can subscribe to a current plan after expiry; the notice says plans are not switched automatically.
  • Legacy Plan V2: the current calculation remains during the active cycle, and eligible users can retain auto-renew or upgrade. A switch to new credit metering is manual under the published rules.
  • Weekend treatment: usage on Saturday and Sunday is deducted at off-peak rates all day.

This is a metering relaunch, not a hard-reset program. Do not use the legacy estimates below as the allowance for a newly purchased credit-based plan.

Separate legacy migration support

Z.AI also publishes a different, account-relative migration program for people who were on a legacy plan with auto-renew enabled as of April 30, 2026 Singapore Time. The notice says eligible accounts receive two complimentary months of the equivalent online tier after the existing paid period and can buy the latest plan at 50% off the latest discounted price until three months after that account’s legacy period ends.

This is not the public 10% first-order invite and not a universal “50% off Z.AI” deal. Current-plan users are outside the notice’s stated scope, each eligible account has its own end date, and the signed-in account remains the authority for whether support is still available. Keep the migration benefit out of generic deal headlines and do not combine it with the referral discount.

Historical August 20, 2026 API comparison: Z.AI’s then-checked docs listed GLM-5.2 API pricing at $1.40 per 1M input tokens, $0.26 per 1M cached input tokens, and $4.40 per 1M output tokens. Against Claude Opus 4.8 API list pricing, that is about 72% lower input and 82.4% lower output. The Coding Plan is different: it bundles supported-tool usage behind subscription quotas.

Legacy Plan V25-hour estimateWeekly estimateBest fit
LiteUp to ~80 promptsUp to ~400 promptsOne project at a time
ProUp to ~400 promptsUp to ~2,000 prompts1-2 active projects
MaxUp to ~1,600 promptsUp to ~8,000 promptsHeavy or parallel coding work

For Legacy Plan V2, Z.AI says each prompt is estimated to invoke the model 15-20 times, and actual usage varies by repository size, task complexity, and auto-accept behavior. New credit-based plans should be evaluated using their published credit allocation and deduction rules instead.

For a first test, Lite is the cleanest sanity check: wire it into one supported coding tool, run it against real repo tasks, and decide whether the model saves enough Claude/GPT spend to justify upgrading. Pro and Max only make sense after you have evidence that GLM handles your codebase well.

Current model routing and quota

Treat GLM-5.3 as a quota-sensitive lane. The current plan overview gives Flash lower token-credit multipliers; compare total tokens, retries and accepted results before selecting a daily lane. Legacy prompt estimates above are not new-plan credit allowances.

The overview also advertises all-day off-peak charging for September 25–October 7, 2026. That is a time-limited campaign, not ordinary savings; the page does not establish its exact cutoff time.

Tool Setup

Use the Coding Plan Chat Completions endpoint for tools using that protocol:

1

https://api.z.ai/api/coding/paas/v4

Use the Anthropic-compatible endpoint for Claude Code and Goose:

1

https://api.z.ai/api/anthropic

For Claude Code, Z.AI’s manual configuration example, checked October 7, maps:

1
2
3
4
5
6
7
8
9

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "your_Z.ai_api_key",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash[1m]"
  }
}

These environment names choose GLM backends; they do not confer Claude capabilities or subscriptions. The same guide separately describes an all-Flash default mapping. Choose explicit routing deliberately and verify the returned model and account meter; this example was source-checked, not runtime-tested.

Deployment Eval

Do not buy the plan from a benchmark headline alone. Use public numbers as a shortlist signal, then run the current GLM lane against one real repository task before making it your daily lane.

Historical GLM-5.2 evidence, checked August 20, 2026: these retained signals do not evaluate GLM-5.3 or Flash.

SourceSignalCaveat
Artificial AnalysisIntelligence Index 51, GDPval-AA v2 1524, leading open-weights placementIndependent benchmark signal, not your repo
Z.AISWE-Bench Pro 62.1 and Terminal-Bench 2.1 81.0Vendor-published coding signal
AIHackersNo owned GLM-5.2 eval yetRun the deployment eval below
Eval stepWhat to comparePass signal
Real bugAsk the selected current GLM model and your premium fallback to fix the same failing testGLM produces the smaller correct patch or a clearly usable diagnosis
RefactorGive both models a 2-4 file cleanup with existing project conventionsGLM preserves behavior and avoids unrelated churn
Test behaviorRequire the coding tool to run or explain the relevant testsGLM uses the repo’s test path instead of guessing
ReviewHave GLM review the premium model’s patch, then reverse the orderFindings are concrete, file-grounded, and not invented

If the tested current GLM lane handles those tasks well enough, use it for routine coding loops. Keep GPT or Claude for risky migrations, final arbitration, and places where a wrong patch costs more than the subscription savings.

When To Choose Z.AI

Choose Z.AI when:

  • You want a cheaper supported-tool lane for daily coding.
  • You use Claude Code, OpenCode, Cursor, Cline, TRAE, Qoder, Droid, Kilo Code, Roo Code, Crush, Goose, Eigent, or the documented OpenClaw path and can stay inside supported tool paths.
  • You want subscription-style budgeting instead of every coding loop hitting a full-price API key.
  • You are comfortable with quota management and model routing.

Skip Z.AI when:

  • You need the single strongest coding model regardless of cost. Evaluate Claude Opus 5.5 or OpenAI Sol for that lane.
  • You need an unrestricted general API subscription. The Coding Plan is restricted to supported tools.
  • You cannot tolerate non-refundable subscription terms or automatic renewal.
  • You do not want to manage peak-hour and weekly quota behavior.

Historical August 20, 2026 comparison: GLM-5.2 in context

This retained table describes the earlier comparison, not current model selection or a fresh price check.

LaneAPI input/outputContextWhy you would choose it
GLM-5.2 / Z.AI$1.40 / $4.401MLow-cost supported-tool coding subscription, good candidate for long-context repo loops
Kimi K3$0.30 cache-hit / $3.00 cache-miss input, $15.00 output1MNewest Kimi flagship to test when Kimi is already in your stack
Kimi K2.7 Code$0.19 cache-hit / $0.95 cache-miss input, $4.00 output256K-classCheaper routine Kimi coding API path
OpenAI GPT-5.xPremium API lane1M-classOpenAI-native workflows, broad tool ecosystem, premium arbitration
Claude Opus 4.8 / SonnetOpus $5 / $25; Sonnet $3 / $151M-classHighest-confidence coding reviews, architecture, and complex agent work

GLM-5.2’s own docs publish stronger long-horizon coding claims and position it for project-scale engineering context. Treat that as vendor-published directional evidence, not purchase proof or a guarantee that it beats Opus or GPT in your repo.

Buyer Checklist

Before you subscribe:

  • Confirm the checkout page still shows the expected first-order discount.
  • Confirm your target tool is on Z.AI’s supported-tool list.
  • Compare GLM-5.3 and Flash on your own tasks before deciding their roles.
  • Run one deployment eval against a real repo task before changing your primary coding lane.
  • Keep one premium fallback, such as Claude Opus 5.5 or Sol, available until the selected GLM lane has passed your repo eval.
  • Recheck renewal, refund, and weekly quota behavior before choosing annual billing.

Bottom Line

Z.AI remains a low-cost coding subscription candidate when your workflow fits its supported tools. The trade is straightforward: lower cost and broad tool compatibility in exchange for supported-tool restrictions, plan-specific metering, and subscription rules. Official invite rules currently list a 10% first-order GLM Coding discount for eligible invited users, but verify the final amount at checkout before paying. The promotion change does not alter this conditional recommendation.

CTA: Try the Z.AI GLM Coding Plan

Sources

Broad historical review: August 20, 2026. Current plan routing, tool listing and Claude Code manual mapping checked October 7, 2026; current-source archives are archive-pending. GLM-5.2 pricing and benchmarks retain their dated scope; no GLM-5.3/Flash paid evaluation or CAR result was measured. Recheck credits, legacy-plan treatment, account-relative migration eligibility, quota multipliers, supported tools, benchmark positions, and referral discount text at checkout before purchasing.