GLM-5.3 Free and Cheap: Every Route and Price (October 2026)

GLM-5.3 has no free hosted tier, and on October 9, 2026 the cheapest way to run the full model is Wafer's US endpoint on OpenRouter at $0.049 input and $3.39 output per million tokens. On a cache-heavy agent session that comes to a little over a third of what Z.ai's own API charges ($1.40/$4.40). If GLM-5.3-Flash is good enough for the task, DeepInfra lists it at $0.075/$0.25 on a 50% promo, and Z.ai sells it first-party at $0.15/$0.50.

Last updated: October 9, 2026. What changed since October 3: both Z.ai Coding Plan campaigns (all-day off-peak billing and unlimited overnight GLM-5.3-Flash) ended on October 7 with no extension posted as of today. The OpenRouter host floor reshuffled: Reka moved from $0.12/$1.14 to $0.14/$4.20, Baidu went to list price, and Wafer is the new cheapest route. Batch is now a DeepInfra-only promo, and Mistral Large 4 arrived at almost exactly GLM-5.3's list price (see "Mistral Large 4 vs GLM-5.3 on price").

The short answer#

Every price above was read from Z.ai's pricing page, the OpenRouter GLM-5.3 endpoint list and the GLM-5.3-Flash endpoint list on October 9, 2026, and the OpenRouter rows were checked against its endpoints API. Several of the cheapest rows are labelled promotions, and the sub-$0.10 input rates move often, so check the endpoints API before you commit traffic.

What GLM-5.3 is#

GLM-5.3 is Z.ai's coding model, built on the same base model as GLM-5.2 with the gains coming from post-training. It exposes three reasoning levels - low, high, and max (the default) - and on the vendor's own benchmark table it scores 66.9 on DeepSWE v1.1 (GLM-5.2: 46.2) and 42.5 on SWE-Marathon v1.1 (GLM-5.2: 19.4). Those are vendor-run numbers, not independent reproductions.

Three siblings matter for cost. GLM-5.3-Flash ($0.15/$0.50 first-party) is the cheapest way into the 5.3 line. GLM-5.3-FlashX ($0.37/$1.25) is a faster Flash that OpenRouter describes as 320B total parameters with 18B active, at up to 200 tokens per second. GLM-5.3-Prime is a throughput variant that OpenRouter lists at 1.5-2x the output throughput of 5.3, still only through Alibaba at $2.80/$8.80 and still absent from Z.ai's own pricing page.

Is anything free?#

Not for GLM-5.3. Z.ai's pricing page lists no free tier for GLM-5.3, 5.3-Flash or 5.3-FlashX, and OpenRouter shows no :free variant for either model. What is free:

• The weights. zai-org/GLM-5.3 is a free download (753B parameters, FP8). Running it is not free; see "Self-host" below.
• Older Z.ai Flash models. Z.ai's pricing page lists GLM-4.7-Flash and GLM-4.5-Flash as free on the API. They are not GLM-5.3, but they cost nothing for light scripting.
• Cached-input storage. Z.ai marks cache storage as "Limited-time Free" across the GLM-5 line. You still pay the cache read rate.

The free GLM-5.3-Flash window that made this page popular in September was a Coding Plan perk, not a free tier, and it ended on October 7 (details below).

Cheapest paid routes for GLM-5.3#

Z.ai charges $1.40 input, $0.26 cached input, and $4.40 output per million tokens for GLM-5.3, unchanged since launch. OpenRouter now lists 33 providers for the model, and the spread is wide. The rows worth knowing on October 9:

The trap in this table is the top row. OpenRouter's own page summary quotes GLM-5.3 at "$0.0392 per million input tokens, $12 per million output tokens", which is Relace's endpoint: the cheapest input on the list paired with the most expensive output. Reasoning tokens bill as output, and max is the default effort, so an agent that thinks a lot pays more on Relace than on Z.ai first-party.

Two changes since our October 1 snapshot: Reka's endpoint moved from $0.12/$1.14 to $0.14/$4.20, and Baidu moved from $0.16/$0.49 to the $1.40/$4.40 list price. OpenRouter's traffic-weighted average for the model is about $0.23 input and $4.15 output, which tells you most traffic already goes to the cheap-input hosts. The sub-$0.10 input rates are the least stable rows: Wafer's US endpoint has shown both $0.039 and $0.049 input on OpenRouter within a single day, so read the endpoints API on the day you route.

The very cheap hosts mostly do not state quantization, and recent uptime on the full list ranges from about 78% (Nebius) to 100%. Test on your own tasks and check recent uptime before moving production traffic. For the general case for spreading traffic across hosts, see Model Routers and the Optionality Advantage.

Worked example: one agent session#

Sticker prices hide how agents spend tokens. Take one coding-agent session of 10 million input tokens, 9 million of them cache reads, plus 500,000 output tokens. That shape (long context re-sent every turn, short replies) is typical of tools like Claude Code and OpenCode. Cost at each route's listed rates:

The math for the Z.ai row is 1M x $1.40 + 9M x $0.26 + 0.5M x $4.40 = $5.94. Two things fall out of it. Wafer's input rate is so low that cache hits barely matter there: if every cache read missed, the session would still cost about $2.19, while the same miss on Z.ai pushes it to $16.20. And full GLM-5.3 batch no longer beats the cheapest interactive host for this shape of work, so batch only earns its place for output-heavy jobs.

Pin a cheap host with a price cap#

OpenRouter load-balances across providers by default. To route on price but stay off the $12-output endpoint, combine sort with max_price in the provider object. The request shape below follows OpenRouter's provider routing docs:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.3",
    "messages": [{"role": "user", "content": "Write a failing test for the date parsing bug"}],
    "provider": {
      "sort": "price",
      "max_price": {"prompt": 1, "completion": 3.5}
    }
  }'

The docs define max_price in dollars per million tokens, so this request only goes to endpoints at or under $1 input and $3.50 output. The :floor model suffix is a superset of sort: "price": per the Floor variant docs it also makes discounted flex service-tier endpoints eligible, billed at whichever tier actually serves the request. If you want price sorting without flex eligibility, use the explicit sort plus max_price, and the cap is what keeps you off Relace. Where it breaks: a cap that is too tight can leave no eligible endpoint, so keep it loose enough to cover two or three hosts.

Flash and batch: the budget routes#

GLM-5.3-Flash is a native multimodal model that Z.ai positions for efficient coding and long-horizon agent work, at $0.15 input, $0.03 cached input and $0.50 output first-party. On OpenRouter it has 29 providers. The cheapest on October 9 are DeepInfra at $0.075/$0.25 (a 50% promo, fp4) and Wafer at $0.07/$0.30. Watch for the same trap as the full model: OpenInference lists $0.032 input but about $1.65 output.

OpenRouter's Batch API runs requests asynchronously: you submit a batch, poll for status, and collect results later. Both batch listings are now served only by DeepInfra, and both are promos:

Batch fits evaluation runs, backfills and overnight jobs, not interactive coding. On Z.ai's Coding Plan, requests for GLM-5.2 and GLM-5.1 route to GLM-5.3 automatically, and requests for GLM-4.7 route to GLM-5.3-Flash, so the cheaper model is often what you get without choosing it. OpenCode Go still includes GLM-5.3-Flash at the $10 tier, with an estimated 6,320 requests per 5 hours on its plan page.

Flat-rate: the GLM Coding Plan after October 7#

The GLM Coding Plan has Lite, Pro, and Max tiers, starting at $18 per month for Lite. All plans support GLM-5.3 and GLM-5.3-Flash, and usage is metered in credits:

What ended. Both campaigns this page covered had October 7 end dates. The all-day off-peak rate ran September 25 to October 7. The GLM-5.3-Flash usage campaign (zero quota use from 23:00 to 09:00 Singapore time in ZCode and AutoClaw, doubled quota in other agents) ran September 3 to October 7, after one extension from September 20. On October 9 both Z.ai pages still show those dates and no new extension.

What still applies. Off-peak usage is billed at 50% of standard credits. Peak is Monday to Friday, 14:00-18:00 Singapore time, which is 08:00-12:00 in Berlin until the clocks change there on October 25 (07:00-11:00 after), and 02:00-06:00 in New York until November 1 (01:00-05:00 after). So the campaign's end barely changes anything for developers in the Americas, whose working day is all off-peak, and halves the effective allowance for European mornings.

What a session costs in credits. Z.ai's formula is (input x multiplier + cached input x multiplier + output x multiplier) / 10,000, with GLM-5.3 multipliers of 6.9, 1.7 and 24, and GLM-5.3-Flash multipliers of 2.3, 0.56 and 8. Run the same 10M-token agent session from the worked example through it:

One catch: at peak, one GLM-5.3 session (3,420 credits) is bigger than Lite's entire 5-hour window (2,000 credits). Z.ai's own estimate puts Lite at 48-97 million GLM-5.3 tokens per week at a 95% cache hit rate, or 146-292 million on Flash, and it claims up to 92% savings against the pay-as-you-go GLM-5.3 API. Treat that as the vendor's best case. Against Wafer's per-token price, the plan still wins if you run several sessions a week: about 5.8 off-peak GLM-5.3 sessions would cost roughly $13 a week on Wafer, against about $4 a week for Lite.

For the subscription route through OpenCode, see the OpenCode developer guide. Z.ai's own coding CLI is covered in the ZCode developer guide.

Mistral Large 4 vs GLM-5.3 on price#

Mistral Large 4, announced October 6, lists at $1.36 input and $4.18 output per million tokens in Mistral's announcement, within a few cents of GLM-5.3's $1.40/$4.40. Mistral's pricing docs currently show a sale price of $0.68 input, $0.07 cached input and $2.09 output, with no end date stated.

On the same 10M-token agent session, Mistral Large 4 costs $2.36 at the sale price and $4.71 at list, against $5.94 for GLM-5.3 on Z.ai and $2.18 on Wafer. Its cache reads ($0.14 at list) are about half of Z.ai's $0.26, which is why it undercuts Z.ai even at list. The difference is availability: GLM-5.3 weights are downloadable today, while Mistral says Large 4 weights arrive by the end of October. One more detail from the endpoint list: Mistral itself serves GLM-5.3 on OpenRouter, at Z.ai's list price.

One month on Flash: what a real team found#

Wagtail, the open source CMS team, set out to spend September on GLM-5.3-Flash alone and published the result on October 2. Their own verdict is that the challenge failed: about half of roughly 2 billion tokens went to Flash, the other billion went to other models. The numbers are useful because they are cost data from daily work, not a benchmark.

• The cheap half was cheap. Flash's share of the month cost about $68, which they put at roughly 4 kWh of energy.
• One bad model choice cost more than that. A vibe-coded prototype burned about 450 million tokens and $150 almost overnight. The author later clarified on Hacker News that this session ran on full GLM-5.3, not Flash.
• Capacity was the real limit. Flash degraded on the inference providers they use, which they blame on its popularity. They switched to DeepSeek V4.1 Flash and Qwen 3.8 Flash, which they call a simple swap.
• Why they still like it. A 1M-token context window, image input, and a wide spread of providers.

Their conclusion matches the routes above: send most day-to-day coding to a flash-tier model, and keep a fallback ready. The pricing side of that fallback is in Where to Run DeepSeek V4.1 Flash Free and Cheap. Where this breaks: Wagtail's work was on Wagtail itself and sites built with it (UI tasks, AI R&D, a bit of documentation writing and a lot of evals), and your capacity and cache behavior depend on which provider you route to.

What people are actually saying#

The Hacker News thread on the Wagtail report reached 233 points and 180 comments. The useful parts are about model choice and cache cost, not the headline:

• Plan with 5.3, execute with Flash. One commenter who uses both says full GLM-5.3 is a much better model, but that Flash carries out a detailed 5.3-written plan "for a fraction of the price". The Wagtail author disagrees in the same thread: 5.3 is only slightly better than Flash and rarely worth the steeper price on usage-based billing.
• Medium effort is the value point. A commenter who ran about 300 agentic and coding tasks reports that Flash's best cost per pass came at medium effort, with max effort scoring about 5% higher at more than twice the cost per pass.
• Cache pricing decides the bill. Two commenters argue cache cost matters more than output cost on long agent sessions, because you pay it every turn. One says Flash was slow on the Coding Plan and moved to Fireworks, and that DeepSeek ends up cheaper because of its cache pricing.
• Speed on the Coding Plan is mixed. A Z.ai subscriber uses Flash for planning and review because it is too slow for execution there, and another reply finds the full model more consistent from Z.ai.

The counter-case: a commenter who called $150 for 450 million tokens excessive suspected bad cache hits, and the author's clarification (it was full GLM-5.3) explains most of the gap. These are individual reports, not measurements.

Setup with Claude Code#

The Coding Plan exposes an Anthropic-compatible endpoint, so pointing Claude Code at it is a settings change. Z.ai's Claude Code guide gives this manual configuration for ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_AUTH_TOKEN": "your_zai_api_key",
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3[1m]",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": 1,
    "API_TIMEOUT_MS": "3000000"
  }
}

That maps the Opus and Sonnet slots to full GLM-5.3 (glm-5.3[1m]) and the Haiku slot to GLM-5.3-Flash (glm-5.3-flash[1m]), all with the 1M-token context. Z.ai's automated setup script and helper default all three slots to GLM-5.3-Flash instead, so check which mapping you ended up with. To keep more of a session on the cheaper model, point ANTHROPIC_DEFAULT_SONNET_MODEL at glm-5.3-flash[1m] as well. Z.ai says the change takes effect in a new terminal window. In OpenCode, log in and run /models to pick the model. To point Codex at an OpenAI-compatible GLM endpoint instead, see custom model providers in Codex.

Self-host#

The weights are on Hugging Face as FP8 safetensors, about 753 billion parameters. The license is a custom glm-5.3 license, not the MIT license GLM-5.2 uses, so read it before commercial use. The model card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth as serving paths. This is datacenter-class hardware, not a laptop; for what fleet-scale serving of the GLM line takes, see Cloudflare's serving playbook. For genuinely local coding on modest hardware, see the best local coding LLMs and the best local models hub.

Which route should you pick?#

• Cheapest per-token production use: Wafer (US) or another cheap-input host on OpenRouter, behind a max_price cap, after testing quality at its quantization.
• Cheapest interactive coding: GLM-5.3-Flash, from DeepInfra while the promo lasts or first-party from Z.ai, with full GLM-5.3 for planning.
• Non-interactive bulk work: Flash batch at $0.06/$0.20. Full 5.3 batch only for output-heavy jobs.
• Daily agentic coding in one tool: the GLM Coding Plan, especially if your working hours fall outside 14:00-18:00 Singapore time.
• Comparing against Mistral Large 4: Mistral's sale price beats Z.ai first-party on agent sessions, but not the cheapest GLM-5.3 host (Wafer US).
• Self-host or air-gapped: the Hugging Face weights, or GLM-5.2 if you need the MIT license.

FAQ#

Is GLM-5.3 free?#

No hosted tier is free. The weights are a free download under the glm-5.3 license, but running a 753B-parameter model costs real compute. The cheapest hosted route on October 9, 2026 is Wafer's US endpoint on OpenRouter at $0.049 input and $3.39 output per million tokens. Z.ai's free API models are older ones: GLM-4.7-Flash and GLM-4.5-Flash.

Did the Z.ai off-peak and free Flash campaigns end?#

Yes. Both had an October 7, 2026 end date: all-day off-peak billing on the Coding Plan, and unlimited overnight GLM-5.3-Flash in ZCode and AutoClaw. As of October 9 Z.ai has not posted an extension. Standard off-peak billing (50% of credits outside Monday to Friday, 14:00-18:00 Singapore time) still applies.

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

For the full model, a cheap-input host on OpenRouter such as Wafer (US). Avoid Relace for reasoning-heavy work: its input is under $0.04 but its output is $12 per million tokens. For the 5.3 family, GLM-5.3-Flash via DeepInfra at $0.075/$0.25 is cheapest while the promo lasts.

Is there a batch discount?#

Yes, on OpenRouter through DeepInfra: GLM-5.3 batch is $0.45 input and $2.00 output, and GLM-5.3-Flash batch is $0.06/$0.20. Both are labelled promotions. For cache-heavy agent work, the cheapest interactive host now matches or beats full-model batch.

Is Mistral Large 4 cheaper than GLM-5.3?#

Than Z.ai's first-party API, yes. Mistral Large 4 lists at $1.36/$4.18 and is on sale at $0.68/$2.09, against GLM-5.3's $1.40/$4.40. On a cache-heavy agent session the cheapest third-party GLM-5.3 host, Wafer (US), is still slightly cheaper than Mistral's sale price ($2.18 against $2.36); the next-cheapest hosts are not.

Is GLM-5.3-Flash good enough to code with every day?#

For many tasks, yes. Wagtail's team spent about $68 on Flash for half of a 2-billion-token month, but hit provider capacity problems and had to fall back to other flash-tier models. Pair it with a fallback route and a stronger model for planning and review.

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

Yes. Z.ai documents an Anthropic-compatible endpoint for Claude Code, Codex can use an OpenAI-compatible GLM endpoint as a custom provider, and in OpenCode you select the model from /models. See the setup section above.

Continue Reading#

• Open-Weight Model Hosting Prices - every provider's GLM-5.3 per-token price, context and quantization side by side, each value quoted
• Mistral Large 4 Is the Open-Weight Frontier Test - the model that just matched GLM-5.3's list price
• How to Use GLM and Other Custom Model Providers in Codex - point Codex at a cheap GLM host
• Anthropic's GLM-5.3 Cyber Report - a competitor's red team measured GLM-5.3's cyber capability
• Where to Run GLM-5.2 Free and Cheap - the previous edition, with the MIT-licensed weights
• Where to Run DeepSeek V4.1 Flash Free and Cheap - the same access math for the other cheap coding model
• Which AI APIs Train on Your Data? - what free tiers and paid APIs do with your prompts, from each provider's own terms
• Budget AI Coding Models Compared - how the cheap tier stacks up per task
• Where to Access AI Models in 2026 - access routes, free tiers, and prices for every major model

Sources#