Where to Run GLM-5.3 Free and Cheap: Every Provider Compared (2026)

Start here: OpenCode Go already serves GLM-5.3 with the full 1M context at the same price as GLM-5.2, and referral code M6HEHM4JM5 gives you $5 in credits plus $5 off the first month. It is the lowest-friction way to try GLM-5.3 through an agent-first interface before committing to a full plan.

Official Sources#

GLM-5.3 launched on August 14, 2026, and the access story is different from every prior GLM release in one important way: the weights are not open yet. Z.ai says the checkpoint ships in roughly two weeks, after a safety evaluation. That means the usual open-weights land rush - a dozen hosts undercutting each other within days - has not started. Today there are exactly two ways to run it: through Z.ai directly, or through OpenCode Go. Both are cheap, and one is effectively free to start.

This post maps what is live now, what it costs, and what to expect when the weights drop. Prices were verified on August 14, 2026. Pricing pages are the only source of truth and they move, so treat the numbers as a snapshot, not a contract.

Last verified: August 14, 2026.

What GLM-5.3 is, in one paragraph#

GLM-5.3 is Z.ai's newest coding model, built on the same base as GLM-5.2 with every capability gain coming from scaled-up post-training. It keeps the 1M-token context window and adds three selectable reasoning levels (low, high, max - Z.ai recommends max for coding, and it is the default). On Z.ai's launch benchmarks it improves on GLM-5.2 across the board: 66.9 on DeepSWE v1.1, 42.5 on SWE-Marathon v1.1, and 31.4% on Z.ai's Code Bench at high effort, edging out Claude Opus 4.8's 29.5% on that last one while still trailing Claude Fable 5. Those are vendor-run numbers, not independent reproductions. Open weights are promised about two weeks after launch; until then there is no license to inspect and no self-hosting.

Setup with Claude Code#

The GLM Coding Plan exposes an Anthropic-compatible endpoint, so pointing Claude Code at GLM-5.3 is an environment-variable swap, not a code change. Add these to your shell config (.bashrc, .zshrc, or equivalent):

export ANTHROPIC_BASE_URL="https://open.z.ai/api/paas/v4/"
export ANTHROPIC_API_KEY="your-glm-coding-plan-key"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.3[1m]"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.3[1m]"
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000

The [1m] suffix enables the 1M-context variant; without it you get the standard window. Alternatively, set the same values under env in ~/.claude/settings.json. Because the endpoint is Anthropic-compatible, existing MCP servers, skills, and hooks keep working without modification.

In OpenCode, log in to your Z.AI Coding Plan (or OpenCode Go) account and run /models to select glm-5.3 - no config edits needed.

The free routes (right now)#

Three paths will run GLM-5.3 with little or no upfront cost today. None is unlimited, so read the terms before you wire a production agent to them.

• OpenCode Go referral credits. OpenCode Go added GLM-5.3 on day one, with the full 1M context and the same per-token pricing as GLM-5.2 ($1.40 input, $4.40 output, $0.26 cached input per million tokens). Use referral code M6HEHM4JM5 and you get $5 in credits plus $5 off the first month. Because GLM inference is cheap, that promo stretches to a real amount of agentic coding compared with paying frontier-model rates for the same workload. If you want to try 5.3 today without touching a Z.ai account, this is the route.
• Already on a GLM Coding Plan? You have it. Z.ai rolled GLM-5.3 out to all existing Coding Plan subscribers at no extra cost. If you subscribed for GLM-5.2, switch your model id to glm-5.3[1m] and you are done - the marginal cost of the upgrade is zero.
• Z.ai ZCode CLI free quota. Z.ai continues to seed its own coding CLI with a large free token allowance to pull developers onto the GLM line, and ZCode picked up 5.3 at launch. Quotas and eligibility change, so confirm on z.ai before relying on it.

Two free routes from the GLM-5.2 era are not here yet: there is no Hugging Face Inference Providers window and no third-party host bundling, because both depend on the open weights that have not shipped. Expect that section of the map to fill in fast once the checkpoint lands. If your goal is genuinely free and local today, GLM-5.2's weights are still on Hugging Face under MIT, and our best local models hub covers laptop-class options.

The cheapest paid routes#

The competitive multi-host pricing table that makes open-weights models cheap does not exist for GLM-5.3 yet. Until the weights ship, every paid route resolves to Z.ai's infrastructure. Here is the live picture.

A few things worth knowing before you pick a row:

• Z.ai has not published a separate GLM-5.3 API price. Its official pricing table still lists GLM-5.2 at $1.40 input and $4.40 output, and OpenCode Go is serving 5.3 at exactly those rates, so that is the working number - but do not assume it is final until Z.ai's table updates.
• The API requires thinking enabled. Direct API calls to glm-5.3 must set "thinking": {"type": "enabled"} with a reasoning_effort of low, high, or max. Reasoning tokens are output tokens, so max costs more per task than the sticker price suggests. low is the lever if you are cost-sensitive.
• OpenRouter, DeepInfra, and Fireworks are absent for now. There is no z-ai/glm-5.3 on OpenRouter yet. When the weights drop, expect the same pattern as 5.2: a dozen-plus hosts within days, fp4 routes undercutting the first-party fp8 price, and a blended rate well under Z.ai's list. See the OpenRouter profile for why the router route matters once it exists.

For the worked cost-per-task math on the GLM line versus closed models, the GLM-5.2 cost math post runs the numbers - the shape carries over to 5.3.

Direct from Z.ai: API vs the Coding Plan#

If you go first-party, Z.ai sells two things, and they suit different usage shapes.

• Per-token API at roughly $1.40 input and $4.40 output per million tokens, with cached input near $0.26 (the GLM-5.2 rates, pending a 5.3-specific table). Right for variable or bursty usage where you pay for exactly what you run.
• GLM Coding Plan flat-rate subscriptions, which bundle GLM-5.3 access into agentic coding tools (Claude Code, OpenCode, Cline, ZCode, and 20-plus others). As verified on z.ai/subscribe on August 14, 2026:

The plan meters usage in credits with per-token multipliers (6.9 input, 1.7 cached input, 24 output), and off-peak usage - outside Monday to Friday 14:00-18:00 UTC+8 - is charged at 50% of the standard rate, which matters if your agents run overnight US time. Note the Pro and Max monthly prices rose versus the GLM-5.2-era tiers ($72 and $160 then). The subscription wins when you code with it daily; the API wins for spiky or automated workloads. Quotas are from Z.ai's devpack docs, so check the current terms.

Local and self-host: not yet#

This is the section that normally makes a GLM release interesting, and for GLM-5.3 it is a waiting room.

• Open weights are promised roughly two weeks after the August 14 launch, once Z.ai completes a safety evaluation. Until the checkpoint and license are published, there is no Hugging Face download, no Ollama tag, no vLLM path, and no license to review. Z.ai's prior releases shipped MIT, but do not assume 5.3 matches until the card is up.
• In the meantime, GLM-5.2 is the self-host option. Same base model, MIT license, weights on Hugging Face, first-class vLLM and SGLang support. Everything in the GLM-5.2 access guide still applies, including the honest caveat: at roughly 756B total parameters this is a datacenter-class model, not a laptop one. For genuinely local coding on modest hardware, see the best local coding LLMs.

Which route should you pick?#

• Just trying it: OpenCode Go with referral code M6HEHM4JM5, or the Z.ai ZCode free quota.
• Already a GLM Coding Plan subscriber: switch your model id to glm-5.3[1m]; the upgrade costs nothing.
• Cheapest production tokens: the Z.ai API (or OpenCode Go) at ~$1.40/$4.40 today; revisit in two weeks when third-party hosts and routers come online and undercut it.
• Daily agentic coding in a tool you live in: the GLM Coding Plan (Lite or Pro), so cost is predictable.
• Self-host or air-gapped: wait for the weights, or run GLM-5.2 now.

FAQ#

Is GLM-5.3 free?#

GLM-5.3 is free or nearly free in two places right now: OpenCode Go referral credits (code M6HEHM4JM5 gives $5 in credits plus $5 off the first month) and Z.ai's ZCode CLI free token quota. Existing GLM Coding Plan subscribers get it at no extra cost. There is no free hosted API window yet and no self-hosting, because the open weights have not shipped.

What is the cheapest way to use GLM-5.3?#

Today, OpenCode Go or the Z.ai API at roughly $1.40 input and $4.40 output per million tokens - they are the only hosts. Once the open weights land (promised about two weeks after the August 14, 2026 launch), expect OpenRouter, DeepInfra, and others to undercut that, as they did with GLM-5.2.

Can I run GLM-5.3 with Claude Code, Cursor, or OpenCode?#

Yes. The Z.ai GLM Coding Plan supports Claude Code, Cursor, Cline, and 20-plus tools; set the model id to glm-5.3[1m] for the 1M context. In OpenCode, run /models and select glm-5.3 after logging in to OpenCode Go or a Z.AI Coding Plan.

Can I run GLM-5.3 locally?#

Not yet. The weights are promised roughly two weeks after the August 14, 2026 launch, after a safety review. Until then GLM-5.2 is the self-host option in the family - same base model, MIT-licensed, on Hugging Face - though it needs datacenter-class hardware either way.

How is GLM-5.3 different from GLM-5.2?#

Same base model, scaled-up post-training. Z.ai's launch numbers improve on GLM-5.2 across every reported benchmark, and it adds selectable reasoning levels (low, high, max). The other practical differences are the reasoning-required API and, for now, the closed weights.

Continue Reading#

• Where to Run GLM-5.2 Free and Cheap - the previous edition of this guide, still the map for self-hosting and third-party hosts
• Where to Access AI Models in 2026 - the hub covering access routes, free tiers, and prices for every major model
• GLM-5.2 Cost Math for Open-Weight Coding Models - the worked cost-per-task numbers; the shape carries over to 5.3
• The Best Local Coding LLMs of 2026 - smaller dense models for genuinely local, laptop-class inference
• Model Routers and the Optionality Advantage - why the router route will matter for 5.3 the moment the weights drop

Sources#

• Z.ai: GLM-5.3 research blog
• Z.ai subscribe (GLM Coding Plan tiers) and devpack overview (quotas, credits, off-peak)
• OpenCode Go referral offer and GLM-5.3 availability announcement
• Unite.AI: Z.ai launches GLM-5.3
• Emergent: GLM-5.3 pricing breakdown
• Kingy AI: GLM-5.3 specs, benchmarks, API
• Hugging Face: zai-org/GLM-5.2 (the currently available open weights)