Compare
Ox Alpha vs the frontier
A single, sourced master table. Ox Alpha next to the 8 models it gets measured against, on the axes that actually decide a build: context, price, coding ability, speed, and whether you can run the weights yourself.
| Model | Released | Context | Input /1M | Output /1M | Coding benchmark | Speed | Reasoning | Open weights |
|---|---|---|---|---|---|---|---|---|
Ox Alpha Z.ai (Zhipu AI), revealed as GLM-5.3 Flash | Aug 20, 2026 | 1M | $0.15 | $0.50 | DeepSWE full 113-task ~63% (community) | ~25-50 tok/s | Yes | Yes |
GLM-5.3 Z.ai (Zhipu AI) | Aug 14, 2026 | 1M | $1.40 | $4.40 | DeepSWE v1.1 66.9% (vendor-reported) | ~102 tok/s | Yes | Yes |
Claude Opus 4.8 Anthropic | May 28, 2026 | 1M | $5.00 | $25.00 | SWE-bench Verified 88.6% | ~61 tok/s | Yes | No |
GPT-5.6 Sol OpenAI | Jul 9, 2026 | 1.05M | $4.00 | $20.00 | SWE-bench Verified 82.2% | ~73 tok/s | Yes | No |
Claude Fable 5 Anthropic | June 9, 2026 | 1M | $10.00 | $50.00 | SWE-bench Verified 95% | ~50-75 tok/s | Yes | No |
Qwen 3.8 Max Alibaba (Qwen team / Alibaba Cloud) | Aug 3, 2026 | 1M | $2.00 | $6.00 | SWE-bench Pro 67.7% (vendor-reported) | ~68.9 tok/s | Yes | Yes |
Grok 4.6 xAI | Aug 12, 2026 | 500K | $2.00 | $6.00 | Terminal-Bench v2.1 88.4% | ~62 tok/s | Yes | No |
Gemini 3.7 Flash Google (Google DeepMind) | Aug 13, 2026 | 1M | $0.75 | $3.75 | DeepSWE v1.1 65.3% | ~348 tok/s | Yes | No |
DeepSeek V4 Pro DeepSeek | V4-Pro-0813, Aug 13, 2026 | 1M | $0.66 | $1.98 | SWE-bench Verified 80.6% (vendor-reported) | ~70 tok/s | Yes | Yes |
Context and price are shortened for scanning (full figures and exact terms are in the notes). Ox Alpha is highlighted.
The Ox Alpha price is Z.ai's own GLM-5.3 Flash list price, read live from OpenRouter (fetched October 8, 2026 at 23:13 UTC). Other hosts charge more or less: see GLM-5.3 Flash on every host.
These figures are a dated snapshot. Run the same models yourself in GenMagic, a multi-model studio we also run, to compare their live output side by side.
What the table says
Where Ox Alpha actually stands
It wins on price and context
Now revealed as GLM-5.3 Flash, it is the cheapest model here (Z.ai lists it at $0.15 per 1M input against $0.66 to $10.00 for the others, and some hosts sell it for $0.075), with a full 1M-token window matched only by the largest models. On those two axes almost nothing beats it.
It trails on measured coding
Its community DeepSWE run (~63%) sits near GLM-5.3 and Gemini 3.7 Flash, but well behind the SWE-bench Verified leaders (Claude Fable 5 at 95%, GPT-5.6 Sol at 82.2%, and the prior Anthropic flagship Opus 4.8 at 88.6%, since succeeded by Opus 5).
It is a fast, cheap mid-tier model
At ~25 to 50 tok/s it lags Gemini 3.7 Flash (~348) and GLM-5.3 (~102), and its measured coding sits mid-pack. Now that the operator is known (Z.ai), the anonymity caveat is gone, but the throughput and coding gaps are real: strong for high-volume agentic work, not a frontier coder.
Notes and sources
Every caveat, cited
The detail behind each row: what is vendor-reported vs independent, pricing nuances, and where the numbers come from.
Ox Alpha
Z.ai (Zhipu AI), revealed as GLM-5.3 Flash ConfirmedRan as an anonymous stealth model on OpenRouter, then revealed: when the preview closed in late August 2026, OpenRouter (the host that ran it) retired the 'stealth/ox-alpha' slug and identified the model as Z.ai's GLM-5.3 Flash (z-ai/glm-5.3-flash), confirming the community's leading fingerprinting theory. Z.ai has not issued a separate formal statement of its own. The stealth preview was free ($0/$0); the revealed model is cheap but paid (Z.ai lists it at $0.15/1M in and $0.50/1M out after launching at half that, and some third-party hosts charge less). Multimodal input (text, image, video), 1M context, 131K max output, adjustable reasoning effort. Benchmark figures are community-run and independent, not vendor-published, so treat as reported.
GLM-5.3
Z.ai (Zhipu AI) ReportedReleased Aug 14, 2026 (Aug 13 PT). Same base model as GLM-5.2; gains from extended post-training. Pricing ($1.40 in / $4.40 out, $0.26 cached read per 1M) and 1M context confirmed across OpenRouter and Artificial Analysis. Coding/agent benchmarks are ALL vendor-reported by Z.ai and not independently re-run: Terminal-Bench 3.0 28.3, DeepSWE v1.1 66.9, SWE-Marathon v1.1 42.5, Toolathlon Verified 73.0. No standard SWE-bench Verified or LiveCodeBench figure published. Standout result: CyberGym 84.5% (found 2,436 vulns across 269 projects). Marketed as the strongest open-weights coding model. Weights published on Hugging Face (zai-org/GLM-5.3) on Aug 28, 2026 under the GLM-5.3 License: MIT-style rights plus a Z.ai security review for model-as-a-service companies with more than $10B in annual revenue. Output speed varies by provider.
Go deeper
Comparison compiled 9-model wide from 42 primary sources. Last verified September 4, 2026.
Build something real on Founden
You have seen the data and the live model. When you want a working product rather than an API call, Founden builds it: a company, a product, a 3D world, whatever you have in mind, without wiring up any of the infrastructure yourself.