Zhipu AI Review (2026): Features, Pricing & Verdict

What Is Zhipu AI?

Zhipu AI is the Beijing frontier lab — founded in 2019 as a spinout from Tsinghua University, and internationally rebranded as Z.ai — whose GLM (General Language Model) family completed the most symbolically loaded arc in the 2026 model market: open weights catching, then passing, the closed coding frontier. The receipts are specific. In May 2026, GLM-5.1 scored 58.4% on SWE-bench Pro — the harder, contamination-resistant successor to SWE-bench Verified — making it the first open-weight model to top that leaderboard, ahead of GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%), while its SWE-bench Verified score (77.8%) sat within four points of Opus. Then on 13 June 2026, GLM-5.2 arrived: a Mixture-of-Experts flagship of roughly 753 billion total parameters with ~40 billion active per token, a 1-million-token context window with up to ~131K output tokens, and a vendor-published benchmark table headlined by 62.1 on SWE-bench Pro (ahead of the GPT-5.5 figure Zhipu cites at 58.6), Terminal-Bench 2.1 at 81.0 (82.7 with the best harness — within a few points of Claude Opus 4.8’s ~85), AIME 99.2 and GPQA-Diamond 91.2 — all self-reported, all pending independent verification, and all backed by the one fact nobody disputes: the weights landed on Hugging Face days later under a clean MIT licence with no regional or revenue restrictions. That licence choice matters doubly here because Zhipu is also a geopolitical curiosity — the lab trains without Nvidia hardware, a differentiator that has only grown sharper as export restrictions tightened through 2026 — and because it makes GLM-5.2 the strongest coding model anyone can legally self-host commercially, full stop. The platform around the models is the most Western-developer-friendly in the Chinese cohort: an OpenAI-compatible endpoint, and — uniquely — a true Anthropic-compatible endpoint (api.z.ai/api/anthropic), which means the same environment variables that point Claude Code at Anthropic can be redirected to GLM with a base-URL swap and zero code changes. That plumbing is what made the GLM Coding Plan the viral budget story of the year: a flat-rate subscription (Lite/Pro/Max at $18/$72/$160 per month list, with a 30% intro promo and yearly discounts to ~$151/$605/$1,344) bundling GLM-5.2, GLM-5-Turbo, GLM-4.7 and GLM-4.5-Air inside Claude Code, Cline, Roo Code and 20-plus other tools — 5-hour and weekly prompt quotas of roughly 80/400/1,600 prompts per five hours by tier, with quota equivalent to an estimated 15–30x the fee at API prices. Metered access prices the flagship at $1.40/$4.40 per million tokens (cached input $0.26, storage currently free), GLM-4.7 at a value-tier $0.60/$2.20, and three Flash models — including a vision model — at literally $0. Within our Model Providers & AI Infrastructure category, Zhipu is the coding specialist with the cleanest exit: not the price floor (DeepSeek), not the breadth king (Qwen), but the lab whose best model is open, whose endpoints speak both Western dialects, and whose subscription undercut the closed coding incumbents hard enough to become a phenomenon.

Core Features

GLM-5.2 and the open coding crown

Zhipu’s claim to this category is concentrated in one model line and one discipline, and the concentration is the strategy: GLM is built to be the best coding and agentic model you can actually own. The architecture follows the cohort’s playbook with bigger numbers — ~753B total parameters, ~40B active per token (a larger activation than DeepSeek’s 32B or MiniMax’s 10B, which buys capability at some serving cost), a 1M-token context window that matches DeepSeek and Qwen at the top of the market, and output headroom around 131K tokens that matters for exactly the whole-file-rewrite and long-plan workloads agentic coding generates. The benchmark story deserves both its trumpet and its asterisk. The trumpet: GLM-5.1’s SWE-bench Pro leaderboard win in May was independently ranked — the first time open weights led the hardest mainstream software-engineering benchmark — and GLM-5.2’s vendor table extends it (62.1 SWE-bench Pro, 81.0–82.7 Terminal-Bench 2.1, 76.8 MCP-Atlas for tool-protocol competence, 54.7 on Humanity’s Last Exam with tools), placing it within a few points of Claude Opus 4.8 on terminal work at roughly a sixth of the price. The asterisk: the 5.2 table is self-reported pending neutral-harness confirmation, and this category’s recent history (Moonshot’s contested Terminal-Bench harness) says wait for third-party numbers before treating decimals as truth — though GLM’s prior, verified leaderboard run makes these claims more credible than most. What the benchmarks translate to in practice, per the now-substantial community record: genuinely strong multi-file refactors, debugging and agentic tool use (the MCP-Atlas score reflects real Model Context Protocol competence), reasoning depth that holds up on math and science probes, and — the differentiator nobody else in the Chinese cohort matches — behaviour tuned to feel native inside Claude Code, because that is transparently the workload Zhipu optimised for. The honest capability boundaries: this is a coding-and-agents specialist, not a multimodal platform — vision exists (the free GLM-4.6V-Flash), but there’s no speech, video or image generation business here (MiniMax’s turf), no swarm orchestration layer (Moonshot’s), and general-assistant polish trails the Western flagships; output speed is respectable but unremarkable; and at $1.40/$4.40 the flagship is the most expensive of the Chinese big four per token — the pitch is frontier coding at a sixth of Western prices, not the absolute floor. For the workload it targets, though, the combination is unique in this review: leaderboard-credible coding, a million tokens of context, and weights you can take home.

The Coding Plan and the Anthropic-endpoint play

Zhipu’s commercial masterstroke isn’t a model — it’s a socket. The Anthropic-compatible endpoint (api.z.ai/api/anthropic) implements the Anthropic messages API contract, tool use and streaming included, which makes Z.ai the only provider anywhere that Claude Code treats as native: swap two environment variables and the world’s most popular agentic coding tool is running on GLM-5.2, no forks, no shims, no code changes. The GLM Coding Plan monetises that socket, and its economics are why it went viral as the budget Claude alternative. The structure: three tiers — Lite at $18/month, Pro at $72, Max at $160 list (a 30% intro promo currently prices them at $12.60/$50.40/$112, and yearly billing lands at roughly $151/$605/$1,344) — all bundling the same model lineup (GLM-5.2, GLM-5-Turbo, GLM-4.7, GLM-4.5-Air) inside 20-plus supported tools: Claude Code, Cline, Roo Code, OpenClaw, Kilo Code, OpenCode, Goose and friends. Quotas are prompt-based, not token-based — roughly 80/400/1,600 prompts per rolling 5 hours and 400/2,000/8,000 per week by tier, plus bundled MCP web-search/reader calls (100/1,000/4,000 monthly) — and Zhipu estimates the bundled quota at 15–30x the subscription fee in API-price terms, which is why heavy daily users almost always beat metered billing on it. Max at $160 undercuts ChatGPT Pro ($200) and Claude Max’s upper tiers while buying the highest prompt quota of any flat-rate agentic-IDE plan currently tracked. Now the fine print, which is where Zhipu’s character shows. Quota drains dynamically: GLM-5.2 and GLM-5-Turbo consume 3x quota during peak hours and 2x off-peak (a promo runs 1x off-peak through September 2026), so a “400 prompts per 5 hours” Pro tier is really ~133 flagship prompts at peak — a mechanism functionally similar to surge pricing that the marketing doesn’t lead with. Hit the cap and the plan simply stops — no overage, no auto-upgrade — and the plan is contractually restricted to supported coding tools: call the API from your own script or production app and you’re on metered billing, full stop. Add the history — the $3/month early-adopter plan withdrawn in February 2026, the ~30% per-token increase that accompanied GLM-5’s launch, tier framing that has shifted between quarterly and monthly — and the pattern is a lab that prices aggressively to acquire developers, then ratchets. None of it changes the bottom line that the Coding Plan is the cheapest credible frontier-coding subscription in the market; all of it says: take the promo, bank the savings, and re-run the maths at every renewal.

MIT weights, free tiers and the trust equation

Zhipu’s trust story is the strongest in the Chinese cohort on the axis that matters most — exit rights — and middling on the axes this cohort always struggles with. Start with the exit: GLM-5.2’s weights ship under a genuine, unmodified MIT licence — commercial use, modification, redistribution, self-hosting, no regional restrictions, no revenue thresholds, no authorisation clauses — which makes it categorically different from MiniMax’s commercially-gated modified-MIT and even a notch cleaner than Moonshot’s lightly-modified terms, and equal to DeepSeek’s gold standard. The consequence compounds with the capability claim: if GLM-5.2 is (or is near) the best open coding model, then the best coding model you can legally run on your own GPUs, for commercial purposes, anywhere on earth, is a Zhipu model — and 25 providers on OpenRouter (from ~$0.93/$3.00, undercutting first-party rates) plus hosts like DeepInfra and Fireworks mean Western-infrastructure inference is a routing decision, not a project. The free layer is unusually generous and worth naming precisely: GLM-4.7-Flash, GLM-4.5-Flash and the vision GLM-4.6V-Flash are listed at $0 for input, cached input and output on the API — genuinely free registered-user models, not trials — and chat.z.ai offers the flagship conversationally for nothing, which together make Z.ai arguably the cheapest serious prototyping platform in this review. Residency and jurisdiction follow the cohort’s standard analysis: Zhipu is a Beijing company, the first-party endpoints are its platform, and organisations that bar Chinese-parent hosted services will bar these too — with the mitigation being the best available anywhere: MIT weights on your infrastructure or a Western host’s, covering the actual flagship rather than a lesser sibling. Two Zhipu-specific trust notes round the picture. First, the Nvidia-free training stack: strategically it insulates GLM’s roadmap from export-control shocks better than any peer’s (a live concern in 2026), though it also means performance-per-watt and serving-cost claims ride on less-familiar silicon. Second, verification culture: Zhipu publishes benchmark tables promptly and in detail, but the June 5.2 numbers remain vendor-reported — the right posture is the one GLM-5.1 earned: credible claims from a lab with a verified leaderboard win, awaiting the neutral harness. The operational hygiene is the cohort’s usual, with emphasis on commercial terms rather than licence terms: prices here move (upward) with little ceremony, promos expire on calendar dates, and quota mechanics get retuned — pin model IDs, budget at list, and let the MIT weights be the insurance policy they genuinely are.

Scored Categories

Pricing

Frequently Asked Questions