GLM-5.2 API Access Compared: Z.ai vs OpenRouter vs the hosts

GLM-5.2 API access is no longer a one-vendor decision. Since Z.ai published the model’s weights under an MIT license on June 16, 2026, roughly 20 independent inference hosts have stood up their own endpoints — and on July 3, 2026, OpenRouter alone routed 29 live GLM-5.2 listings with input prices spanning $0.93 to $3.00 per million tokens.

That spread is the whole story. Pick well and you pay about a third less than Z.ai’s own sticker. Pick carelessly and you pay a 2x premium — or worse, your 128K-token agent outputs get silently truncated to 32K because the cheapest route carries the tightest output cap on the market. The cheapest listing and the right listing are not the same thing.

This guide prices every route worth considering. We pulled OpenRouter’s live endpoints API, cross-checked each figure against the provider’s own model page, and layered in Artificial Analysis’s independent speed benchmarks — then built the scoreboard table and decision rows no single source publishes. Subscription plans, the free Flash tier, and the self-hosting reality check round it out.

GLM-5.2 was announced on June 13, 2026 — GLM Coding Plan subscribers first, API and chatbot access days later — and the open weights landed on Hugging Face and ModelScope under an MIT license on June 16. That license is why this comparison exists at all: anyone with datacenter GPUs can legally serve the model, so within weeks the market filled with independent hosts competing on price, speed, and infrastructure story.

The demand side is simple economics. Anthropic’s Claude Opus 4.8 lists at $5 per million input tokens and $25 per million output. Z.ai’s GLM-5.2 list price is roughly 3.6x cheaper on input and 5.7x cheaper on output. The honest framing — the one our own GLM-5.2 benchmark breakdown lands on — is that GLM-5.2 is near-frontier on many single-shot coding benchmarks at a fraction of the cost, but trails Opus 4.8 on sustained long-horizon agent work. Most teams we work with treat it as a pairing, not a replacement: route high-volume, well-scoped tasks to the cheap model and keep the frontier model for the work that justifies frontier pricing.

Z.ai’s own API is the reference implementation: $1.40 per million input tokens, $0.26 for cached input, $4.40 for output, with cached-input storage listed as “Limited-time Free” on the official pricing page. The model card states a 1M-token context window and a 128K-token maximum output — the most generous output ceiling among the mainstream routes, and the cap most marketplace hosts mirror.

Interestingly, the sticker is not a vendor premium. DeepInfra’s own market analysis, citing Artificial Analysis data, puts the median tracked provider price at exactly Z.ai’s list — $1.40 / $4.40 / $0.26. Fireworks and Novita list the identical figures on their own model pages. In practice, $1.40 in / $4.40 out is the market’s consensus price, and everything else is a deliberate deviation: discounts to win traffic, or premiums to sell speed.

Two more things distinguish the native route. First, the GLM Coding Plan — Z.ai’s subscription — only works against Z.ai’s own endpoint, so plan subscribers who follow our Claude Code setup guide are on this route by definition. Second, data handling: Z.ai’s privacy policy states that services are generally provided from Singapore and that API request content is not stored — “The Company do not store any of the content the Customer or its End Users provide or generate while using our Services.” Its parent company is registered in Beijing, a separate, secondary-sourced fact about legal domicile rather than data flow — we unpack both in the decision rows below.

If you ran the predecessor model through a gateway, note that the playbook from our GLM-4.6 API deployment guide still applies structurally — same OpenAI-compatible surface, same Anthropic-compatible proxy path — but every price in it is now stale. GLM-5.2 repriced the whole market.

OpenRouter is where the price competition becomes visible. On July 3, 2026, its live endpoints API for z-ai/glm-5.2 returned 29 distinct listings across roughly 20 provider brands — DeepInfra, GMI Cloud, Novita, Fireworks, Wafer, Decart, Baidu, SiliconFlow, Together, Cloudflare, and more — several of them listing both a standard and a “fast” tier. One API key, one request format, twenty backends bidding for your traffic.

Because it is a marketplace, OpenRouter has no single GLM-5.2 price — treat every number as a range with a date on it. At our July 3 snapshot, the sub-sticker band ran $0.93–$1.20 per Mtok input across the discounted routes, while the full span stretched from $0.93 to $3.00 on input and $3.00 to $10.25 on output once premium latency tiers are included. These rates fluctuate: Novita’s OpenRouter route, for example, showed $1.09 in / $3.43 out — a live promotional rate roughly 22% below Novita’s own $1.40 / $4.40 sticker, not a list price. At the top end, Wafer’s “fast” tier listed $3.00 in / $10.25 out — about 2.1x Z.ai’s list input and 2.3x its output — sold on lowest end-to-end latency.

One caution on provider counts, because every source publishes a different number. OpenRouter’s raw JSON returned 29 endpoints across ~20 brands; DeepInfra’s July 1 blog counts 25 listed providers from a different snapshot; Artificial Analysis benchmarks 15 high-uptime routes; and models.dev shows 42 rows because it counts subscription SKUs and regional variants as separate entries. None of these is wrong — they measure different scopes — but they are not interchangeable, and any post quoting a single “X providers” headline without a source and date is already stale.

The table below is the asset this post exists for. Every cell cross-references OpenRouter’s live endpoint JSON against the provider’s own pricing or model page, as of July 3, 2026 — no single published source combines both surfaces, which is exactly how promotional rates get miscited as list prices. Context is 1,048,576 tokens nearly everywhere; the columns that actually separate the routes are price, output cap, and quantization.

One correction worth making explicit, because several published comparisons repeat it: DeepInfra is often placed mid-pack at around $1.20 in / $4.20 out. That figure actually matches Decart’s route ($1.20 / $4.20) and roughly Wafer’s standard rate ($1.20 / $4.10) — not DeepInfra’s. Three independent surfaces — DeepInfra’s own July 1 blog post, its model page, and OpenRouter’s live endpoint JSON — all put DeepInfra Standard at $0.93–0.95 in / $3.00 out. DeepInfra is the market floor, not a mid-tier option, and any comparison that says otherwise is working from a stale snapshot.

The structural read: the market has settled into three bands. A consensus sticker at $1.40 / $4.40 that Z.ai, Fireworks, and Novita all defend; a discount band between $0.93 and $1.20 where hosts buy traffic with promos and aggressive quantization; and a premium band at 1.5x-plus multiples where the product being sold is latency and guaranteed throughput, not tokens. Where you should sit in that stack depends entirely on workload shape — which is what the decision rows below are for.

Price-per-token is the visible axis; the two invisible ones bite harder. First, output caps. Z.ai’s native route allows 128K output tokens, and most marketplace hosts mirror the 131,072 ceiling — Novita, Baidu, GMI Cloud, AtlasCloud, Alibaba, Venice, Wafer, and others. SiliconFlow lists 262,144. A handful — Morph, Friendli, Decart, Phala — list the full 1,048,576-token context as their completion ceiling. And DeepInfra, the cheapest route on the board, caps completions at 32,768 tokens: the lowest of any active listing found.

Second, quantization. OpenRouter’s metadata shows the fleet split roughly three ways: fp8 hosts (GMI Cloud, Novita, Baidu, SiliconFlow, Venice, and Z.ai’s own route as tagged), fp4 hosts (DeepInfra, Wafer, Decart, Parasail, Inceptron), and a group that discloses nothing (Together, Cloudflare, Friendli, Morph, DigitalOcean, among others). NVIDIA publishes an official NVFP4-quantized GLM-5.2 variant for datacenter deployment, so a vendor-blessed fp4 recipe exists — but that is not the same as a quality guarantee for any particular host’s implementation.

On quality impact, the honest answer is qualitative: general LLM literature suggests fp8 typically lands very close to the full-precision baseline, while fp4 is more implementation-dependent — and as of July 3, 2026, no GLM-5.2-specific fp8-vs-fp4 head-to-head benchmark had been published. Nobody can currently tell you what the fp4 discount costs in output quality, including the hosts selling it. If your workload is quality-sensitive, run your own evals on the exact route you plan to buy, not on Z.ai’s reference endpoint.

Five buyer profiles cover nearly every team we have priced this market for. Match yours, then verify the live rate on the route before committing — every marketplace number in this post is a July 3, 2026 snapshot.

The residency question is not hypothetical — it is being asked in public. A Cursor community-forum thread titled “Do GLM 5.2 requests go to China?” drew an official reply from a Cursor team member: “the provider for GLM 5.2 and also other Cursor models is Fireworks Ai” — pointing enterprise users at Fireworks’ own trust center rather than Z.ai. That is the pattern we expect more downstream products to follow: open weights served on domestic infrastructure, with the model’s origin and its data path cleanly separated. If your team needs help pressure-testing that kind of routing decision — model, host, and compliance story together — that comparative eval is exactly where our AI transformation engagements start.

Metered tokens are not the only way in. For steady daily usage — especially agentic coding through Claude Code or similar tools — Z.ai’s subscription can undercut every per-token route on this board, and at the other extreme, the free Flash tier costs nothing at all.

The break-even math between plan and metered billing depends on your usage shape — we run the full calculation in our Coding Plan value analysis. The short version: spiky or low-volume API usage favors metered tokens; daily agentic coding favors the plan. The plan also feeds Z.ai’s wider tooling — ZCode, the free desktop agentic development environment Z.ai shipped the week of July 1, 2026, runs on the same quotas, with a 1.5x-quota launch promotion Z.ai states expires July 31, 2026 (our ZCode guide covers it in full).

If the subscription math fits your usage, you can start a GLM Coding Plan directly on Z.ai — every tier runs against the same native endpoint this comparison prices. Referral link: we earn Z.ai platform credits if you subscribe, and new Z.ai accounts get 10% off their first subscription order.

And if you are tempted by the MIT license to run the model yourself, read why self-hosting GLM-5.2 is unrealistic for most teams first — the hardware bill dwarfs years of API spend at these per-token prices.