Z.ai (GLM) Alternatives in 2026: Coding Plans, APIs and Older GLM Models

The GLM Coding Plan from Z.ai costs from $18 monthly and now routes every GLM-5.2 and GLM-4.7 request to newer models. Its API lists GLM-5.3 at $1.40 per million input tokens.

Those two facts explain most searches for a Z.ai alternative. Some people want a different coding subscription for Claude Code. Some find GLM-5.3's API rate high next to cheaper Chinese models. And some simply used an older GLM version that the plan no longer serves. This guide takes each group in turn, with prices read from every vendor's own page and the benchmark numbers Z.ai itself publishes.

What Z.ai sells, and what changed this year

Z.ai, the company behind the GLM models, sells three things people look for replacements for:

• The GLM Coding Plan: a monthly subscription for agent tools such as Claude Code, Codex, OpenCode and Cline.
• The API: pay-as-you-go access to GLM-5.3, its cheaper Flash variants and older GLM-5 and GLM-4 models.
• A chat app: chat.z.ai, whose page title says it runs on GLM-5.3-Flash.

Two changes drive a lot of the current searching. First, the Coding Plan's documentation says requests for GLM-5.2 and GLM-5.1 are now sent to GLM-5.3, and requests for GLM-4.7 go to GLM-5.3-Flash. If you tuned prompts on GLM-5.2, you are no longer running it inside the plan.

Second, GLM-5.3 kept the price of its predecessor. That is fair for a better model, but it leaves room for rivals that charge less per token.

How to pick a replacement for Z.ai

Answer these before reading any table:

1. Are you paying for the Coding Plan or the API? They have different competitors. Subscriptions compete on usage limits; APIs compete on cost per million tokens.
2. Do you need GLM specifically? GLM-5.3's weights are public, so you can run the same model elsewhere, or keep the Coding Plan and add a second provider for overflow.
3. When does your work happen? Two of the options below charge less at certain hours, and the hours differ.
4. How heavy is your agent use? Coding agents send far more input than output, because they reread files. Price input first.

If you would rather describe your setup than score it yourself, Smart Match shortlists tools from the directory based on a short brief.

The shortlist: Z.ai alternatives by use

We took these figures from the vendors' own pages on the morning of 28 October, and they move often.

Our picks, briefly:

• Most Coding Plan users should stay. It is the cheapest entry plan we found, and its limits are published in tokens.
• Switch to Kimi Code if you want Moonshot's K3 model in the same tools.
• For the API, add DeepSeek as a second supplier and send batch work to it outside peak hours.
• For bulk, low-stakes calls, use MiniMax M3 or stay on GLM-5.3-Flash.

Coding subscriptions compared

The GLM Coding Plan has three tiers: Lite, Pro and Max. Z.ai's developer docs give the entry price as 18 US dollars a month and publish each tier's limits as credits per five hours and per week.

The token ranges assume a 95% cache hit rate. Off-peak use costs half the credits, and the peak window is weekday afternoons, 14:00 to 18:00 in Singapore, which is why each range doubles.

Kimi Code and the M Plan against the GLM plan

Two rivals sell the same kind of subscription. Their entry prices sit within $4 of each other, and both work in agent tools that GLM users already know, so the decision comes down to which model you want behind the key.

Here are the two closest rival plans:

Kimi Code's entry Plus plan costs $19 a month, or $15 a month on yearly billing. Moonshot's page says K3 is available on Plus and above, and the newer K3.1 Pro runs at standard speed on Plus.

Kimi Code's full ladder, from Moonshot's plan page:

Yearly billing saves about a fifth on every tier. Moonshot does not publish token allowances for these tiers, which makes a like-for-like comparison with the GLM plan's published ranges impossible; the only fair test is a month of your own work on each.

Kimi's faster "Highspeed" mode is reserved for Pro and above. Moonshot says it is about two to three times faster and uses credits twice as fast. Kimi Code runs on desktop, in a CLI and as IDE plugins.

MiniMax's M Plan uses M3.1 Flash for text on every tier, with Explore at $55 and Build adding 7.5 times Go's usage. Its edge is the H3 video model bundled into the two upper tiers.

Our verdict for coding: keep the GLM plan unless a specific model is the point. If your agent struggles on long, multi-step tasks, a month of Kimi Code Plus is the cheapest way to test K3 inside the tools you already use. Tools such as Cursor, OpenCode and Cline accept any of these keys, so the test costs one config change. Our AI coding assistants guide compares the agents themselves.

What GLM-5.3 costs on the API

Z.ai's pricing page lists the whole family per million tokens. Storing cached input is free for a limited time, which matters for agents that resend the same files.

Two things in that table are easy to miss. The cached rate is roughly a fifth of the full input rate on every row, so an agent that rereads the same repository pays far less than the headline price. And the two previous flagship versions cost exactly what GLM-5.3 costs, so nobody saves money by pinning an older model.

Web search through the API costs one cent per use.

API alternatives: DeepSeek, Kimi and MiniMax

DeepSeek V4-Pro: close on benchmarks, half price at night

DeepSeek V4-Pro is the nearest like-for-like rival. DeepSeek's pricing page lists it at $1.32 per million input tokens and $3.96 per million output tokens at peak, slightly under GLM-5.3.

Off-peak, those fall to $0.66 and $1.98. DeepSeek charges full price on weekdays between 01:00 and 04:00 UTC and again between 06:00 and 10:00 UTC; weekends count as off-peak.

Both DeepSeek models offer a 1M-token context and Anthropic-format endpoints, which means Claude Code can point at them directly. The cheaper DeepSeek V4.1 Flash, still reachable under the old V4 Flash name, competes with GLM-5.3-Flash rather than the full model. DeepSeek also starts billing server-side Search and Fetch calls on 1 November 2026.

Kimi K3: stronger on some tasks, at a premium

On the Moonshot AI API, Kimi K3 costs $3 for each million input tokens and $15 for each million output tokens.

The newer K3.1 Pro costs $2 and $10, which makes it the cheaper and more recent of the two.

Both carry a context window of 1,048,576 tokens. Moonshot's older K2 series, including Kimi K2.7 Code, closes to new registrations on 31 October 2026 and retires at the end of December, so start any new project on K3. The Kimi app record covers the consumer side.

MiniMax M3: the budget alternative

MiniMax M3 is billed at $0.30 for a million input tokens and $1.20 for a million output tokens, after what MiniMax calls a permanent 50% cut.

That is roughly a fifth of GLM-5.3's input price. Its smaller sibling, M3.1 Flash, is $0.15 in and $0.60 out, nearly the same as GLM-5.3-Flash.

Monthly bill for an agent-heavy workload

To make the rates concrete, here is one coding-agent month: 20 million input tokens and 5 million output tokens, with no caching. Input dominates because agents reread files on every step.

We multiplied each vendor's listed rate ourselves. Real agent traffic caches heavily, which would cut every row, and GLM's cache storage is currently free.

At these volumes the Coding Plan wins outright for most people. On Z.ai's own figures, even Lite allows 48M to 97M GLM-5.3 tokens a week at a 95% cache hit rate, more than this whole month. The API only makes sense when you need to call GLM from your own software. HokAI's model leaderboard puts price, speed and benchmarks for these models on one page, and our cheapest LLM API guide widens the comparison.

Peak hours: the overflow trap

Three of the providers here price by time of day, and their busy windows overlap more than you might expect.

The GLM plan's peak sits exactly inside one of DeepSeek's two peak windows. So a developer who hits the GLM five-hour limit in a Singapore afternoon and fails over to DeepSeek pays DeepSeek's full rate at that moment. Neither discount helps you then.

Two practical consequences follow:

• For overflow during Asian afternoons, MiniMax is the better fallback. Its standard rate does not change by hour, and M3 costs less than DeepSeek's off-peak V4-Pro rate anyway.
• Run batch jobs outside 01:00 to 10:00 UTC. That clears both peaks and earns both discounts. For a team in Europe it means scheduling heavy runs after mid-morning.

Kimi's API page lists no time-of-day pricing, and Kimi Code describes its quotas only as regularly refreshed.

Pointing Claude Code at a different model

Claude Code is the agent most often paired with the GLM Coding Plan, so the switching cost question usually becomes: which rivals can it talk to without extra software?

• GLM Coding Plan: supported directly, according to Z.ai's docs, along with Codex, OpenCode and Cline.
• MiniMax M Plan: MiniMax's docs list Claude Code, Codex, Cursor and OpenCode as supported tools for the plan's API key.
• DeepSeek API: DeepSeek publishes an Anthropic-format base URL next to its OpenAI-format one, and its pricing page refers to Claude Code use explicitly.
• Kimi Code: sold as its own desktop app, CLI and IDE plugins, so the natural path is Moonshot's client rather than Claude Code.

The DeepSeek endpoint, for reference, as listed on its pricing page:

OpenAI format:     https://api.deepseek.com
Anthropic format:  https://api.deepseek.com/anthropic

In practice this means three of the four alternatives drop into an existing Claude Code setup by changing a key and a base URL. Prompts, project files and habits stay the same, which keeps a trial cheap.

How GLM-5.3 compares, according to Z.ai

Z.ai's model card for GLM-5.3 publishes a benchmark table against its main Chinese rivals. These are the vendor's own numbers, run with its own settings, so read them as a claim rather than an independent test.

On Z.ai's own figures, the three are close on coding. Kimi K3 edges ahead on tool use, and GLM-5.3 leads on CyberGym. The same table puts Qwen3.8-Max behind GLM-5.3 on all four of these rows.

The practical reading: no rival is clearly stronger at coding, so cost and tooling should decide. Z.ai also says GLM-5.3 improved 50% over GLM-5.2 on its in-house code benchmark, which is a reason to accept the automatic upgrade rather than hunt for the old model.

If you used an older GLM model

Many searches name a specific version: GLM-4.5, GLM-4.7, GLM-5.1 or GLM-5.2. The answer depends on where you used it.

• Inside the Coding Plan: you no longer can. The plan forwards those requests to GLM-5.3 or GLM-5.3-Flash automatically.
• On the API: older models are still listed, at lower prices. GLM-5.2 and GLM-5.1 cost the same as GLM-5.3, so there is no saving in staying on them.
• Self-hosted: keep running whatever weights you have. Nothing forces an upgrade.

If you moved to GLM-4.7 for its price, GLM-5.3-Flash is now the closer like-for-like, and it is cheaper on output.

Running GLM yourself instead

Z.ai publishes GLM-5.3's weights on Hugging Face. The model card lists 753B parameters and support for SGLang, vLLM, Transformers, KTransformers and Unsloth, plus Huawei Ascend hardware.

That size needs a multi-GPU server. A smaller open model through Ollama or LM Studio is the realistic route on one machine, and our local coding models guide lists what fits.

In between, hosted inference companies such as Fireworks AI and Groq serve open models by the token; our Fireworks vs Together comparison explains how those platforms price.

One practical setting: GLM-5.3 defaults to maximum reasoning effort unless you pass low or high. On a self-hosted or third-party endpoint, setting it lower is the fastest way to cut latency and cost.

Data handling before you move code

Coding agents send your source files to whichever model you choose, so the data question is sharper here than for chat. We did not audit each provider's terms for this guide. Before you switch, check three things:

• The API terms that apply to your key. Subscription plans and pay-as-you-go APIs can carry different rules.
• Whether prompts are used for training, and how to opt out if they are.
• Where requests are served from, if your company has data-residency rules.

For proprietary code, the cautious order is: self-hosted open weights first, then a provider whose terms your legal team has read. For review workflows on top of these models, our code review tools guide compares the options.

When staying on Z.ai is the right call

Here is the best case against switching. GLM-5.3 is an open-weights model that Z.ai's own table puts level with Kimi K3 and ahead of DeepSeek V4-Pro on several coding benchmarks, and the GLM Coding Plan is the cheapest entry subscription we priced. Few alternatives beat both.

That case holds for most solo developers. The exceptions are narrow: you want K3 specifically, you want video in the same plan, or you call the model from your own product and want a lower per-token rate.

If none of those applies, the better move is redundancy rather than replacement. Keep the plan, and configure DeepSeek or MiniMax as a fallback for when the five-hour limit runs out. Our guide to choosing the right LLM sets out a simple way to test models side by side on your own tasks.

How to switch a coding agent in an afternoon

1. List the tasks you actually give the agent. Five real tickets beat a benchmark.
2. Add the new provider's key alongside Z.ai's. Many agent tools let you define several model endpoints and choose per session.
3. Run the same five tickets on both. Count failed runs, not just speed.
4. Check usage after a week. Compare credits or dollars spent per completed ticket.
5. Read the refund terms before cancelling. Time the switch to your renewal date if partial months are not refunded.

More options to consider

• Qwen, from Alibaba, comes in hosted and open-weight versions, and our explainer on what Qwen is for maps the range. Alibaba Cloud's pricing page would not load in either of our browsers, so we have not priced it here.
• For a free chatbot rather than a coding tool, our AI chatbots roundup compares the consumer apps, and our ChatGPT vs Kimi comparison covers Moonshot's app.

What would change these picks

• Z.ai cuts GLM-5.3's API price. At DeepSeek's off-peak rate, the API case for switching would disappear.
• The free cache storage ends. Z.ai labels it limited-time, and agent workloads lean on caching.
• Kimi Code adds a cheaper tier. Kimi already sits only $1 above the GLM plan at entry.
• DeepSeek changes its off-peak windows. The half-price argument depends on them.

DeepSeek begins charging for Search and Fetch on 1 November, a day after Kimi K2 closes to newcomers. We will reprice the agent-month table once both changes have landed.

Frequently asked questions

What is the best alternative to the GLM Coding Plan?

Kimi Code is the closest rival, at $19 a month on monthly billing against $18 for the GLM Coding Plan, and it includes Moonshot's K3 model on every plan. MiniMax's M Plan starts at $22 and adds H3 video on its upper tiers. For most solo developers the GLM plan remains the cheapest entry point.

Is DeepSeek cheaper than GLM-5.3?

Yes, slightly at peak and by half off-peak. DeepSeek V4-Pro costs $1.32 per million input tokens and $3.96 per million output tokens at peak, against $1.40 and $4.40 for GLM-5.3. Outside DeepSeek's peak windows its rates halve to $0.66 and $1.98.

Can I still use GLM-4.7 or GLM-5.2 in the Coding Plan?

No. Z.ai's documentation says requests for GLM-5.2 and GLM-5.1 are now routed to GLM-5.3, and requests for GLM-4.7 go to GLM-5.3-Flash. The older models are still listed on the pay-as-you-go API.

Is GLM-5.3 better than Kimi K3 for coding?

On Z.ai's own benchmark table the two are close: GLM-5.3 scores 88.2 on Terminal Bench 2.1 against 88.3 for Kimi K3, and 66.9 on DeepSWE against 67.5. Those are vendor-reported figures, so test both on your own tasks before paying for either.

Can I run GLM-5.3 myself instead of paying Z.ai?

Yes, the weights are published on Hugging Face, and the model card lists support for SGLang, vLLM, Transformers, KTransformers and Unsloth. At 753B parameters it needs a multi-GPU server, so most individuals will use a smaller open model locally instead.

Covered in this guide

• Z.ai: Z.ai (formerly Zhipu AI) is a Beijing lab behind the GLM model family, founded 2019 and listed on the Hong Kong Stock Exchange (2513) since January 2026.
• Cline: Free, open source AI coding agent for VS Code, JetBrains, and the terminal with 11M+ installs; bring your own Claude, GPT, or Gemini key.
• Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
• DeepSeek: DeepSeek is a Chinese AI research company developing frontier language and reasoning models, including DeepSeek V4-Pro. Founded by High-Flyer hedge fund CEO Liang Wenfeng in 2023, the company is known for achieving GPT-level performance at dramatically lower compute and API cost.
• V4 Flash: DeepSeek-V4 Flash: 284B-param MoE with 13B active (April 2026), 1M context, 79% SWE-bench, $0.14/M input. Open-source MIT license.
• DeepSeek V4-Pro: DeepSeek V4 Pro: an MIT-licensed frontier-scale MoE, generally available since August 2026, with 1M-token context and text-only input.
• DeepSeek V4.1 Flash: The MIT-licensed vision-and-text model, launched in September 2026 and refreshed with an October snapshot.
• Fireworks AI: Enterprise inference platform for open models: serverless tokens from $0.10 per 1M, managed fine-tuning per training token, and dedicated NVIDIA GPUs from $8.00/hr. $1 free credits, self-serve, no cold starts.
• GLM-5.2: GLM-5.2 is Z.ai's June 2026 open-source 744B MoE model with a 1M-token context window, 85.6% GPQA Diamond (Vals AI), and MIT license at $1.40/1M input tokens.
• GLM-5.3: GLM-5.3 is Z.ai's text-only flagship for agentic coding and security work, with a 1M-token context, a live API and open weights.
• GLM-5.3-Flash: Open-weight, natively multimodal MoE model from Z.ai with a 1M-token context window and MIT license, released in August 2026.
• Groq: Fast open-model inference, now built on NVIDIA-licensed LPU hardware
• Kimi app record: Native multimodal AI assistant with Agent Swarm capabilities and 256K context window
• Kimi K2.7 Code: Open-weight 1T/32B-active MoE coding model from Moonshot AI, released June 2026 under a Modified MIT license and built for long-horizon agentic coding.
• Kimi K3: 2.8T-parameter open-weight MoE model from Moonshot AI (July 2026) with a 1M-token context window and the top published open-weight result on GPQA Diamond.
• LM Studio: Free desktop app that runs open LLMs locally on llama.cpp and MLX, with an OpenAI-compatible local server and an optional $20 cloud plan.
• MiniMax: Listed in Hong Kong since January 2026 and controlled by founder Yan Junjie, MiniMax builds the MiniMax M3 language model and the open-weight H3 video model from its Shanghai base.
• MiniMax M3: MiniMax's June 2026 open-weight flagship, built around a 1M-token context window, native multimodal input, and frontier-level agentic coding performance.
• M3.1 Flash: MiniMax's October 2026 Flash model: 1M-token context, text, image and video input, and always-on reasoning with five effort levels.
• Moonshot AI: Beijing AI lab founded in 2023 by Yang Zhilin, maker of Kimi and the open-weight Kimi K3 models, valued at $20B at its May 2026 round.
• Ollama: MIT-licensed runtime that runs open-weight LLMs on macOS, Windows and Linux with a local API, plus an optional paid cloud for larger models.
• OpenCode: MIT-licensed coding agent for terminal, desktop and IDE that works with any model provider, from 208K GitHub stars.
• Qwen: Alibaba's AI assistant with 203M monthly active users, supporting 119+ languages, multimodal input, and models up to 1M token context windows.
• Qwen3.8-Max: Qwen3.8-Max is Alibaba's flagship MoE model, launched Aug 2026, with a 1M-token context, native multimodal input, and a text-only open-weight checkpoint released weeks later.

Sources