Skip to content
NewsModel release

Stealth model Ox Alpha outscores Claude Fable 5 on coding, draws 221K users as forensics point to Zhipu AI

· by Pondero Newsdesk

The short version

An anonymous model called Ox Alpha appeared on OpenRouter on August 20, 2026, outpaced Claude Fable 5 on community coding benchmarks, and drew 221,000 unique users before independent researchers traced its serving infrastructure to Zhipu AI. No official attribution has appeared.

Stealth model Ox Alpha outscores Claude Fable 5 on coding, draws 221K users as forensics point to Zhipu AI

An anonymous model called Ox Alpha appeared on OpenRouter on August 20, 2026 without a company name or a press release. It offered a 1,048,576-token context window for free, scored above Claude Fable 5 on a community coding benchmark, and drew roughly 221,000 unique users in three days, per Bloomberg. By August 22, independent researchers had published serving-layer forensics pointing to Zhipu AI's infrastructure with 0.98 operator-layer confidence. Zhipu has not confirmed or denied anything.

What happened

OpenRouter listed stealth/ox-alpha on August 20, describing it as "a reasoning model designed for coding, sustained agentic work, and production workloads" from "a third-party provider who has chosen to remain anonymous during this preview," per the model page. The context window of 1,048,576 tokens supports text, image, and video. OpenCode, an open-source coding agent, made it available on its Zen plan through approximately August 27. The anonymous provider claimed capacity of 100 trillion tokens per day, per TechTimes.

Community benchmarks followed immediately. Developer Ben Davis ran Ox Alpha against DeepSWE, a coding-agent evaluation testing AI performance on real software engineering problems, and reported 80% on a 10-task run. Claude Fable 5 scored 65% and GPT-5.6-Sol scored 52% on the same tasks, per TechTimes citing aiwiki.ai. These are community-run results on a 10-task sample, not scores from an audited public leaderboard; a single task's outcome shifts the pass rate by 10 percentage points. Stripe CEO Patrick Collison, whose company announced an agreement to acquire OpenRouter on August 19, called the model "very impressive" in a post on X, per TechCrunch.

The forensic case for Zhipu AI

Developer research moved from speculation to serving-layer evidence within 48 hours of the model's appearance.

Researcher Chetaslua sent a malformed request to OpenCode's Ox Alpha route, setting the top_p parameter to a string instead of a number. The server returned a Java validation error exposing its own internal class: com.wd.paas.api.domain.v4.chat.ChatCompletionRequest. That package path maps to Zhipu's documented API route at open.bigmodel.cn and api.z.ai, per the explainx.ai forensic writeup. Chetaslua rated operator-layer confidence at 0.98.

Two independent signals corroborated the stack trace. Malformed role-field requests to Ox Alpha returned error code 1214 in the same format as Z.ai-hosted GLM models on OpenRouter. Running the identical test against a different host's instance of GLM-5.2 returned a different error format entirely, meaning the dialect identified the serving operator, not just the model weights. Separately, researcher @aitrackerbot ran 25 diverse prompts through Ox Alpha and recorded token counts; Chetaslua extended that to 30 probe strings covering 14 writing systems, code, and SQL. All 30 matched GLM-5.3's tokenizer with a constant 75-token offset per request, the signature of a system prompt prepended invisibly to every query, per explainx.ai.

Explainx.ai described the stack trace as "implementation detail leaking at the routing layer" and distinguished it from earlier behavioral fingerprinting attempts. As of August 23, Zhipu AI, OpenRouter, and OpenCode had all made no official statement.

Why it matters for AI-tool operators

The benchmark numbers, if they hold under independent replication, require a reconsideration of the standard tier list. A model scoring 80% where Claude Fable 5 scored 65% on the same community coding tasks is significant enough to warrant a controlled evaluation before the free window closes August 27. That much is straightforward.

The complications start with attribution. Operators who don't know who built the model can't assess its training data, fine-tuning approach, or IP posture. A 100-trillion-token daily capacity claim from an anonymous provider carries no SLA and no identifiable entity to hold accountable if the service disappears.

Data retention presents the sharpest near-term concern. The OpenRouter listing for Ox Alpha states that prompts and completions are retained by the provider but not used for training. The platform's governing Stealth Model EULA contradicts that notice; it grants OpenRouter and the unnamed provider the right to use user content for training, evaluation, and improvement, and instructs users who object to that use to avoid stealth models entirely, per TechTimes citing the EULA terms. A separate OpenCode route claims zero days of retention and no training use. Two routes to the same model, two different stated policies, both applying to an anonymous provider with no named legal jurisdiction. Security and compliance teams sending production code through either route have no clean answer for the question of who holds that data.

Context: a documented launch pattern

Ox Alpha is the fifth stealth model in six months on OpenRouter, per TechTimes. The prior four followed the same arc: anonymous launch, free preview, community forensics, then official claim after the preview closed. Z.ai's GLM-5 was one. Xiaomi's MiMo-V2-Pro appeared under the names Hunter Alpha and Healer Alpha. Ant Group's Lingxi Ling-2.6-flash and Meituan's LongCat-2.0 completed that prior sequence.

The economics clarify the motive. Serving frontier inference free for a week at 100 trillion tokens per day is a substantial infrastructure outlay. Labs that have run this play before converted a portion of preview users into paying customers at reveal time, and they collected real production-workload performance data without lab-reputation bias shaping the results. OpenRouter, in the process of being acquired by Stripe, provides a Western developer-facing distribution channel that Chinese AI companies lack on their own. The pattern has repeated enough times that "stealth model on OpenRouter" now reads as a recognized launch category rather than a genuine mystery.

What to watch next

The free window closes around August 27. Whether Zhipu AI officially claims the model before or shortly after that date will determine whether the forensic case becomes a confirmed story or remains a well-evidenced open question. The benchmark claim is a larger unknown. The 10-task DeepSWE community run is the only performance data on record; an independent evaluation at the scale of public leaderboards is the number that would require Anthropic and OpenAI to respond publicly. If the 80% figure holds on a controlled run, the competitive picture changes. If it doesn't, Ox Alpha becomes an instructive case study in how anonymity amplifies community benchmark credibility.

Sources