AI Coding Plan Limits: Claude Code, Codex, Kimi, GLM and ClinePass Compared

Last checked: August 29, 2026 at 06:19 UTC

AI coding limits are designed to resist a clean comparison. One company publishes message ranges, another publishes only multipliers, another exposes exact credits but changes the burn rate by model and time of day, and another shows the real allowance only after sign-in.

The result is a market full of numbers that look comparable but are not. A “five-hour limit” might be a rolling token budget, a dynamically estimated message range, or one layer beneath a separate weekly or monthly cap. A subscription may also share its allowance with chat, document agents, voice, spreadsheets, or every device and API key on the account.

This guide separates the clocks, pools, and billing paths. It does not turn undisclosed quotas into fake precision.

Evidence labels used below

• Officially fixed: The provider publishes a number or formula.
• Officially described but dynamic: The provider describes the limit, but workload, demand, model choice, or risk controls can change it.
• Observed or estimated: A disclosed estimate or a calculation with stated assumptions, not a guarantee.
• Unknown: The provider does not publish enough information to calculate the limit.

What changed in this edition

• OpenAI now publishes Plus message ranges for GPT-5.6 Sol, Terra, and Luna, while local messages and cloud tasks share a five-hour window. Weekly limits may still apply.
• Kimi Code now sits inside a membership-wide monthly credit pool and also has its own rolling five-hour and seven-day controls. Model access differs sharply by tier.
• Z.ai’s current GLM Coding Plan publishes exact five-hour and weekly credits plus token multipliers. Legacy plans use different rules and can remain active.
• ClinePass has expanded its open-model lineup, but its numerical five-hour, weekly, and monthly caps remain unpublished.

Limits at a glance

Scroll or swipe to compare.

Prices exclude taxes, app-store differences, promotions, and regional billing. “Current price” means the public web price visible at the evidence cutoff.

Reset clocks and shared pools in plain English

A rolling five-hour window does not usually reset at midnight. Capacity used at 2:15 p.m. returns according to the provider’s rolling-window logic around five hours later. A fixed weekly reset is different: the entire weekly allowance returns at an assigned day and time. A subscription-anchored seven-day window starts from purchase or activation and repeats every seven days.

The second distinction is the pool. If Claude chat and Claude Code share usage, an afternoon of document work can reduce the coding capacity available that evening. If Kimi Code has coding-specific limits inside a membership-wide monthly pool, either layer can stop the next request. OpenAI’s agentic pool can extend beyond Codex to Work and supported productivity agents. Z.ai’s supported coding tools draw from the same plan quota, even when the client changes.

The practical rule is to check every active clock and every shared surface. The largest headline multiplier does not help if a smaller weekly pool is already empty.

Claude Code limits

Plans, clocks, and pools

Claude Code is included with paid Claude plans. Individual pricing is $20 monthly or $200 annually for Pro, $100 monthly for Max 5x, and $200 monthly for Max 20x.

Anthropic describes the allowance as dynamic. Every plan has a rolling five-hour window; paid plans add weekly limits. Max 5x provides five times Pro capacity per five-hour session, while Max 20x provides 20 times Pro. Anthropic does not promise a fixed message, token, or task count because conversation length, repository context, model choice, attachments, and features change consumption.

Claude on web, desktop, mobile, and Claude Code draw from the same pool. Anthropic may also impose model-specific, feature-specific, weekly, or monthly controls. The reset day and time for the weekly allowance are assigned to the account and shown under Settings → Usage.

Classification: officially described but dynamic. The five-hour and weekly clocks are official; a universal task count is unknown.

Model eligibility changes the effective limit

Current Claude Code documentation lists Fable 5, Opus 5, Sonnet 5, earlier Opus releases, and Sonnet 4.6 as supported models. Access and the default depend on the plan, account, provider route, and client version.

One expensive edge case is long context. Anthropic says current paid plans can use one-million-token-capable models in Claude Code, but Pro users need usage credits for one-million-token Opus access. A larger context window also makes larger requests possible; it does not make them free. Thinking tokens count, even when the interface collapses or redacts them.

When the included allowance ends, Claude Code blocks further subscription requests until the displayed reset. Paid users can enable extra usage where available, upgrade, or use a Console/API route billed by tokens. An ANTHROPIC_API_KEY in the environment makes Claude Code use API billing rather than the subscription, so authentication mistakes can create charges instead of consuming the plan.

OpenAI Codex limits

Published ranges, variable tasks

Codex is included in ChatGPT Free, Go, Plus, Pro, Business, and Enterprise paths, but eligibility and capacity differ. Plus is $20 monthly. OpenAI’s two Pro tiers are $100 and $200 monthly and advertise five times or 20 times Plus usage.

For the Plus plan, OpenAI’s public table showed these local-message ranges per shared five-hour window when checked:

The range is wide because a message is not a fixed unit of work. Repository size, prompt length, output, reasoning level, tools, delegation, fast mode, and task duration all change consumption. Local messages and cloud tasks share the five-hour window. OpenAI says additional weekly limits may apply but does not publish one universal weekly number.

Codex is no longer an isolated pool

OpenAI says Codex, ChatGPT Work, ChatGPT for Excel, and Workspace Agents draw from the same agentic allowance and credit pool when those products are available. Ordinary ChatGPT voice has separate caps, while Voice used inside supported Work/Codex flows can touch the agentic pool under the current rules.

After included usage runs out, eligible Plus and Pro accounts can buy credits. Business, Edu, and Enterprise workspaces with flexible pricing can use workspace credits. Users can also switch to a cheaper model such as Luna or run local Codex tasks with their own API key at API rates.

OpenAI is migrating credit billing from per-message estimates to input, cached-input, and output-token rates. Migration state can differ by plan and workspace. The usage dashboard and /status are more authoritative for an individual account than any static article.

Classification: mixed. The five-hour ranges are official estimates; actual task burn and weekly limits remain dynamic.

Kimi Code limits

Kimi has four monthly membership tiers: Andante at ¥49, Moderato at ¥99, Allegretto at ¥199, and Allegro at ¥699. Annual subscriptions are available, but the public help page tells readers to view the live checkout for the exact discounted amount.

Kimi Code has three layers of control:

1. A rolling five-hour coding limit.
2. A coding quota that refreshes every seven days from the subscription date; unused quota does not roll over.
3. A membership-wide monthly credit pool that refreshes each billing cycle.

All membership features, including Kimi Code, Kimi Work, research, PPT, Kimi Claw, K3, and Agent Swarm, draw from the monthly pool. Kimi Code’s five-hour and weekly controls apply only to coding, but the monthly pool sits underneath them. Emptying the monthly pool freezes Kimi Code even when the coding-specific dashboard still shows capacity.

Kimi’s current product overview advertises approximately 300–1,200 requests per five-hour window and up to 30 concurrent requests. It does not map that broad range to individual tiers or define a standard request, so it is an official estimate rather than a transferable entitlement. Kimi tells users to inspect the signed-in console for their current percentage and reset time. All logged-in devices and membership API keys share the coding quota.

Kimi model eligibility by tier

HighSpeed delivers roughly five to six times the output speed at about three times the quota consumption. Kimi also says the one-million-token K3 route consumes about twice as much quota as K3-256K. Switching models or reasoning effort invalidates the existing context cache, so frequent switching can make the quota fall faster.

For model-level benchmark evidence and public API economics, see Kingy’s Kimi K3 benchmarks and API pricing.

When a time-limited quota is exhausted, a member can wait, upgrade, or enable Extra Usage. Extra Usage is a shared pay-as-you-go wallet for Kimi and Kimi Code. It is deducted only after the subscription quota and can bypass the monthly, weekly, or five-hour membership stop. The user should set a monthly spending cap; Kimi says no cap applies by default.

Classification: officially estimated and dynamic. The broad five-hour request range, windows, sharing rules, and model entitlements are official; exact per-tier coding allowances remain account-visible rather than publicly fixed.

GLM Coding Plan limits

Z.ai is the most transparent provider in this group for new individual plans. Its documentation publishes the credits, reset rules, model multipliers, and formula.

The live subscription page showed standard/list monthly prices of $18 for Lite, $80 for Pro, and $168 for Max when checked. It was also displaying temporary discounted monthly prices of $12.60, $56, and $117.60. Z.ai’s older migration notice still contains $72 and $160 examples for Pro and Max, but labels that table “for reference only”; the live plan page controls current checkout pricing.

Five-hour credits are restored five hours after consumption. The weekly allowance starts with the subscription and resets every seven days. Every supported coding tool shares the subscription quota.

All new individual tiers support GLM-5.3 and GLM-5.3-Flash. Requests for GLM-5.2 or GLM-5.1 route to GLM-5.3, while GLM-4.7 routes to GLM-5.3-Flash under the current mapping.

Kingy’s GLM-5.3 specs, API and access guide covers the underlying model separately from these subscription-credit rules.

Z.ai calculates model credits as:

credits = (fresh input tokens × input multiplier + cached input tokens × cached multiplier + output tokens × output multiplier) ÷ 10,000

Usage outside the weekday peak period is charged at half the standard credit rate. Z.ai defines peak as Monday through Friday, 14:00–18:00 Singapore time (UTC+8). Web Search, Web Reader, Zread, and visual-understanding MCP calls also consume credits.

Two qualifications matter. First, concurrency remains dynamic: Z.ai can change it with resource availability, although Max is designed for more simultaneous projects than Pro or Lite. Second, legacy subscriptions use different prompt-based rules and can remain active. Readers must identify their plan version before applying the table above.

When the coding-plan quota ends, the supported coding endpoint does not automatically take money from the general API balance. The user waits for reset or upgrades. General pay-as-you-go API use is a separate route and balance.

Classification: officially fixed credits and formula, with dynamic concurrency.

Where ClinePass fits

ClinePass is a $9.99 monthly subscription for a curated set of open models inside Cline and through the Cline API. A separate promotional landing page advertised $4.99 for the first month when checked, followed by the standard recurring price.

Cline publishes three clocks:

• A rolling five-hour window.
• A calendar-week limit.
• A calendar-month limit.

It does not publish the numerical allowance for any window. It also does not state the calendar reset timezone, unused-capacity treatment, or a reproducible conversion from reference token prices to quota. Cline says the subscription supplies two to five times the usage available at standard API rates, but that is a marketing range rather than a user-verifiable token budget.

The plan currently includes models from Z.ai, Moonshot, DeepSeek, MiniMax, MiMo, and Qwen. Model-specific reference prices differ widely, so switching models can change how fast the same hidden quota falls. ClinePass usage through an API key remains part of the subscription; it is not a separate pay-as-you-go API balance.

ClinePass is relevant here because it is cheap, has a strong Kingy search footprint, and exposes the disclosure problem clearly. Readers who want the full model list and product verdict should use Kingy’s ClinePass review.

Classification: unknown numerical caps. The windows and standard price are official; the amount inside each window is not public.

What these plans are worth at API prices

Subscription quota is not redeemable API credit. “API-equivalent value” is a comparison of token cost for the same illustrative workload, not cash value, guaranteed output, or proof that two products run the same model configuration.

Reference workload

The calculations below use one million billed model tokens split as follows:

• 200,000 fresh input tokens.
• 700,000 cached input tokens.
• 100,000 output tokens, including billable reasoning output where the API counts it.

Formula:

API-equivalent cost = fresh input × fresh-input rate + cached input × cached-input rate + output × output rate

Tool charges, cache writes, long-context premiums, fast modes, retries, images, regional uplifts, taxes, and multi-agent multiplication are excluded.

Z.ai’s static API pricing table did not include a GLM-5.3 row at the evidence cutoff. The GLM comparison below applies Z.ai’s published GLM-5.1 rates to the GLM-5.3 Coding Plan token mix as a proxy. It is not a confirmed GLM-5.3 API price.

These figures compare token rates, not task completion. A cheaper model that retries, reasons longer, or fails more often can cost more per accepted change.

A reproducible GLM plan example

GLM is the only plan here that exposes enough subscription information for a bounded whole-plan estimate. Because the static API table lacks GLM-5.3, the dollar column uses the disclosed GLM-5.1 API rates as a comparison proxy.

Under the reference mix, one million GLM-5.3 tokens consume 497 plan credits at the standard rate:

(200,000 × 6.9 + 700,000 × 1.7 + 100,000 × 24) ÷ 10,000 = 497 credits

Off-peak usage consumes 248.5 credits for the same token mix. Applying the weekly caps and 4.348 weeks per average month produces this range:

This is an illustrative upper-bound-style use case, not a forecast. Real coding sessions have a different input/output mix, cache rate, tool-call count, model split, concurrency profile, and idle time. A user who does not exhaust every weekly credit receives less value. A legacy plan does not follow this formula.

Why the other plans do not get a monthly dollar figure

• Claude: Anthropic does not publish a token allowance or fixed message count for the subscription pool.
• Codex: OpenAI publishes message ranges, but one message can consume radically different tokens and agent work. Purchased-credit prices can also vary by account and migration state.
• Kimi: The public page publishes a broad request range but not each tier’s numerical coding or monthly token pool.
• ClinePass: Cline publishes a two-to-five-times value claim and reference rates, but not the quota formula or caps needed to reproduce it.

Any precise monthly API-equivalent number for those plans would be an estimate built on private account observations, not a current public entitlement.

Which plan fits your usage pattern?

Light coding use

Choose the product you already pay for and favor the cheaper eligible model. Plus with Luna or Terra, Claude Pro with Sonnet, Kimi Andante with K2.7 Code, or GLM Lite can all cover intermittent work. The critical question is whether the coding pool is shared with other features you use heavily.

Moderate daily use

Favor visibility over a large marketing multiplier. GLM Pro publishes a formula and credits. Kimi Moderato unlocks K3 but still hides the numerical coding quota. Claude Max 5x and OpenAI Pro 5x buy more capacity while keeping dynamic limits. Track a normal week before upgrading.

Heavy or parallel agent use

Use a metered fallback with a spending cap. Parallel agents, long context, high reasoning, fast modes, and repeated cache invalidation can consume quota far faster than interactive chat. A plan with paid overage is less disruptive than one that stops completely, but uncapped auto-reload can turn convenience into a surprise bill.

For product capabilities, safety defaults, and workflow fit beyond limits, see Kingy’s Grok Build vs Codex vs Claude Code comparison and the deeper Codex vs Claude Code guide.

FAQ

Change log

Official sources

Testing and disclosure

Kingy did not purchase every subscription or run controlled quota-exhaustion tests for this reference. The article reports current public documentation and labels private/account-visible quantities as unknown. Live dashboards, regional checkout, promotions, grandfathered plans, risk controls, and rollout state may differ.

No provider supplied or reviewed this article. No affiliate claim is required for the research package; add the site’s standard disclosure if production links become commercial.