Qu'est-ce que GLM-5.3 ? Le lancement discret du plan de codage de Z.ai, tarifs et améliorations confirmées

What Is GLM-5.3?

GLM-5.3 is the newest coding-focused model available through Z.ai's subscription-based GLM Coding Plan. Z.ai, formerly known as Zhipu AI, develops the Chinese GLM family of language, vision, image, video, and agent models.

The current official documentation positions GLM-5.3 for coding agents rather than ordinary browser chat. Developers can select it in tools that support custom models, including Claude Code, Goose, Codex, and Cline. Z.ai also says that Coding Plan requests made to the previous glm-5.2 and glm-5.1 routes are now automatically sent to GLM-5.3.

Here is what can be verified today:

This table is deliberately shorter than the speculative GLM-5.3 specification sheets circulating online. Architecture, parameter count, benchmark scores, and an open-weight license remain unknown. Copying those details from GLM-5.2 would create a more impressive table, but it would also make the table wrong.

Is GLM-5.3 Released? Release Date and Preview Status

GLM-5.3 has been released inside one product: the GLM Coding Plan. Z.ai's model-switching guide now displays “GLM-5.3 Now Live” and says that Lite, Pro, and Max subscribers can switch to it.

There is still no exact GLM-5.3 release date. Z.ai's general model release notes do not yet contain a dated GLM-5.3 entry, while its standard model overview and per-token pricing pages still lead with GLM-5.2. The safest description is therefore a quiet Coding Plan launch, not a complete public rollout.

This also changes how the term “GLM-5.3 preview” should be understood. Before the documentation update, GLM-5.3 was effectively a release-watch keyword supported by community discussion. It is now a callable Coding Plan model with an official ID. What remains incomplete is the documentation around the model—not its existence inside the plan.

For production teams, that gap is important. A route can be callable before its architecture, billing outside a subscription, capacity limits, independent quality, and self-hosting terms are known. Availability answers “Can I test it?” It does not yet answer “Should I replace a production model with it?”

Which GLM-5.3 Features Are Confirmed?

An optional 1-million-token context

Z.ai documents glm-5.3[1m] for developers who need a 1-million-token context. In Claude Code, the setup also requires CLAUDE_CODE_AUTO_COMPACT_WINDOW to be set to 1000000.

The suffix matters. Developers should not assume that every request using the plain glm-5.3 string automatically receives the same context configuration. The long window is relevant for large repositories, long agent histories, migration plans, and multi-file debugging tasks. The cost is higher context consumption and potentially longer reasoning cycles.

One million tokens also does not guarantee that a model will remember every dependency equally well. Context capacity measures how much can be submitted. Retrieval accuracy, constraint retention, and tool-loop stability still need task-level testing.

Low, high, and max reasoning effort

The GLM-5.3 Coding Plan maps different tool settings into three actual effort levels:

Low effort is the sensible starting point for small edits and simple questions. High is the middle ground. Max is intended for complex coding and agent work where deeper reasoning matters more than response cost or latency.

No official GLM-5.3 benchmark currently measures the difference between these settings. Treat effort as a test variable, not a quality guarantee. The more useful production metric is cost per accepted task after retries—not whether every request ran at max.

Coding-agent compatibility

The official switching guide lists three access patterns:

• Claude Code and Goose use the Anthropic-compatible endpoint.

• Codex uses Z.ai's /api/v1 endpoint.

• Cline and other compatible tools can use the OpenAI-style coding endpoint.

This is the clearest signal about GLM-5.3's current purpose. It is being distributed as a model for code generation, debugging, repository questions, and long-running agent tasks—not as a new multimodal consumer assistant.

Text-first access, with vision kept separate

Z.ai tells Cline users to disable “Support Images” when configuring glm-5.3. Its Coding Plan separately offers Vision Understanding through a vision MCP service.

In plain language: the current direct GLM-5.3 route should be treated as text and code input. A screenshot may still enter a broader Z.ai workflow through another service, but that is not evidence that the underlying glm-5.3 model natively accepts images.

This is one of the largest gaps in early GLM-5.3 coverage. Community demand for vision is real. A feature request, however, is not a shipped modality.

GLM-5.3 vs GLM-5.2: What Was Actually Upgraded?

The most visible confirmed upgrade is routing: GLM-5.3 is now the default current model inside the Coding Plan, and earlier GLM-5.2/5.1 requests are automatically routed to it. Z.ai has not yet published enough technical evidence to quantify the underlying quality improvement.

Z.ai's GLM-5 repository reports that GLM-5.2 scored 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro. It also reports that IndexShare reduced per-token FLOPs by 2.9 times at a 1-million-token context and that changes to speculative decoding increased acceptance length by up to 20%.

Those are vendor-reported GLM-5.2 results. They provide a baseline, not GLM-5.3 evidence. We do not yet know whether 5.3 changes the architecture, adds new training data, improves tool calling, shortens reasoning traces, or simply represents a newer post-training revision served through the same family.

Another correction is worth making: some early articles call GLM-5.2 a 753-billion-parameter model. Z.ai's current repository and download table list it as 744B-A40B. The official figure is the safer one to carry into a predecessor comparison.

Developers who need a documented and currently accessible member of the same family can use the GLM-5.2 API on GPT Proto while waiting for a verified GLM-5.3 route.

GLM-5.3 Pricing: What Does It Cost?

There are two different answers to GLM-5.3 pricing, depending on whether “pricing” means the Coding Plan subscription or a standard pay-per-token API.

GLM Coding Plan subscription

Z.ai's plan documentation lists subscription access starting at $18 per month. At the time of checking, its public subscription page displayed promotional prices of $12.60 for Lite, $56 for Pro, and $117.60 for Max, compared with list prices of $18, $80, and $168. Promotions can change, so teams should confirm the checkout price before budgeting.

All three plans support GLM-5.3. Max does not buy a different “GLM-5.3 Max” model; it buys a larger usage allowance.

GLM-5.3 credit consumption

The Coding Plan overview gives GLM-5.3 these credit multipliers:

Z.ai charges half the standard credit rate during its documented off-peak periods. That can materially change how much work a subscription supports, but credits are not dollars. Converting these multipliers into a made-up per-million-token price would mix two billing systems.

Standard API token price

Z.ai has not listed GLM-5.3 on its standard API pricing page. The page currently prices GLM-5.2 at $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens.

Do not apply those prices to GLM-5.3. They are useful only as a predecessor baseline.

Is “GLM-5.3 Max” a Separate Model?

No official source currently identifies a separate model named GLM-5.3 Max.

The word “Max” appears in two different places. First, Max is the highest GLM Coding Plan subscription tier, with 28,000 credits per five-hour period and 140,000 credits per week. Second, max is the deepest documented reasoning-effort setting for glm-5.3.

Therefore, a search such as “GLM-5.3 Max vs GLM-5.2 Max” usually means comparing both versions at maximum reasoning effort. It does not establish two separate Max-branded models.

That comparison cannot be completed reliably yet. GLM-5.2 has published vendor and independent results; GLM-5.3 does not. A fair test would need the same agent setup, task set, context, effort setting, retry policy, and acceptance criteria. Comparing one polished demo from 5.3 with an old 5.2 benchmark would tell us very little.

GLM-5.3 vs Kimi K3 vs Claude Fable 5

GLM-5.3 is already being discussed beside Kimi K3 and Claude Fable 5 because all three target demanding coding or long-context work. The evidence available for them is not equal.

That is also why searches for “GLM-5.3 vs Kimi K3” or “GLM-5.3 vs Fable 5” do not yet have an evidence-based winner.

Independent testing currently gives Kimi K3 a score of 60 on the Artificial Analysis Intelligence Index. The same source records text and image input, 1-million-token context, and a 2.8-trillion-parameter MoE architecture with 104 billion active parameters. Anthropic's official documentation lists Claude Fable 5 pricing at $10 per million input tokens and $50 per million output tokens, with a 1-million-token context.

The current LMArena code leaderboard also includes both Kimi K3 and Fable 5. GLM-5.3 has not yet been added. Until it appears in reproducible evaluations, claims that GLM-5.3 beats either model are predictions, not comparisons.

For developers choosing today:

• Choose GLM-5.2 when open weights and lower documented token cost matter most.

• Choose Kimi K3 when you need a 1M context, image input, and strong independently measured coding performance.

• Choose Fable 5 when the highest documented Anthropic tier matters more than token cost.

• Test GLM-5.3 if you already have a Z.ai Coding Plan, but keep a fallback until its behavior and billing fit your workload.

GPT Proto currently provides Kimi K3 API access at $2.70 input and $13.50 output per million tokens. Claude Fable 5 is also available at $8 input and $40 output per million tokens. These are deployable alternatives while GLM-5.3 is not yet in the GPT Proto catalog.

Can You Access GLM-5.3 on GPT Proto?

No. GLM-5.3 is not currently available on GPT Proto.

Do not send glm-5.3 to the GPT Proto endpoint or describe GPT Proto as a GLM-5.3 provider. Developers who need a GLM route today can call GLM-5.2 instead. The following request uses GPT Proto's existing OpenAI-compatible API and the real glm-5.2 model ID:

curl "https://gptproto.com/v1/chat/completions" \
  --header "Authorization: Bearer $GPTPROTO_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "glm-5.2",
    "messages": [
      {
        "role": "user",
        "content": "Review this implementation plan and identify the three highest-risk dependencies."
      }
    ]
  }'

When GLM-5.3 becomes available, its live model page—not a rumor post—should be the source of truth for the GPT Proto model string, modalities, context, and price.

You can also browse the GPT Proto AI Model Gallery to compare currently available text, reasoning, image, video, and multimodal routes using one API key and one balance.

Should Developers Switch to GLM-5.3 Now?

If you already subscribe to the Z.ai Coding Plan, GLM-5.3 is worth testing now. Automatic routing means you may already be receiving it when requesting GLM-5.2 or GLM-5.1 through that product. Record the resolved model, effort level, context setting, token usage, latency, retries, and whether the result passed review.

Do not treat the routing update as proof that every workload improved. A newer coding model may follow instructions better while consuming more output credits. It may finish harder tasks while responding more slowly at max effort. The useful question is not “Did the version number increase?” It is “Did accepted work become cheaper or more reliable?”

Wait before making GLM-5.3 a production dependency if you need any of the following:

• A stable pay-per-token API price

• Public rate limits and capacity guarantees

• Downloadable weights and a confirmed license

• Native image input on the same model route

• Independent coding, agent, latency, and throughput results

• A verified GPT Proto model page and endpoint

For a project that must ship today, use a documented model and make the model ID configurable. That lets you evaluate GLM-5.3 later without rewriting the surrounding application.

Final Verdict

GLM-5.3 is real and callable, but its release is narrower than the model name suggests. It is currently the newest model inside Z.ai's GLM Coding Plan, with a documented 1M-context option, low/high/max effort controls, and compatibility with several coding agents.

What we do not have is just as important: no dated launch post, no GLM-5.3 model card, no standard per-token price, no open weights, and no benchmark report. That makes GLM-5.3 a credible evaluation candidate—not yet a model that can be ranked honestly against Kimi K3, Fable 5, or even GLM-5.2 on cost per successful task.

If you already pay for the Coding Plan, test it. If you need a documented API route now, start with GLM-5.2, Kimi K3, or Claude Fable 5, and keep the model switchable.