FULL STORY

Unmasking Ox Alpha: Anonymous Model Revealed as Zhipu's GLM-5.3

An anonymous model called Ox Alpha appeared on OpenRouter with free massive token quotas, sparking speculation from Google rumors to mixed benchmarks. Forensic analysis and Bloomberg reporting ultimately revealed it as Zhipu's GLM-5.3 Flash, set to be open-sourced.

2026-08-21 ~ 2026-08-28 · 16 episodes · 166 posts

Episode 1 · Mystery model Ox Alpha stuns community; fingerprints point to Zhipu GLM-5.3 (2026-08-21, 57 posts)

An anonymous model called Ox Alpha appeared this week on OpenRouter and OpenCode with no vendor attribution, paper or announcement, yet quickly went viral thanks to a 1M-token context window, text/image/video multimodal input, a zero-data-retention policy and a near-unlimited free week. Its benchmark scores are strong, and black-box fingerprinting plus a leaked system prompt point to Zhipu's GLM-5.3 as the likely identity — still unconfirmed, making it the week's biggest mystery.

Confirmed

  • OpenRouter positions it for efficient coding, sustained agent tasks and production use, with 1M-token context and text/image/video input; host OpenCode claims compute for 100 trillion tokens per day. @1littlecoder verified the 1M context, multimodality, zero data retention and generous free trial quota.
  • It scored about 80% on DeepSWE's 10-task subset, beating GPT-5.6-sol (52%) and Fable (65%). @Teknium tested OX Alpha Max in Hermes Bot, calling it exceptionally smart with above-expectation frontend quality, available with unlimited quota.
  • @FlunkyGraphics ran three black-box fingerprint tests (tokenizer comparisons over 6 text types) confirming Ox Alpha matches the GLM-5.3 architecture.
  • A system prompt leaked by @elderplinius forces the model to claim it is "ox-alpha" from an undisclosed organization and never reveal its true identity; it reportedly belongs to a Zhipu-affiliated team.
  • Known weakness: @bindureddy notes it tends to produce idle filler content despite being free, long-context and multimodal.

Unconfirmed

  • The final attribution is unconfirmed. The mainstream guess is Zhipu's GLM line — a GLM 5.3 vision variant or GLM 5.3 Flash; a claim relayed by @calabiandyau also says Kimi k3.1 is coming soon.
  • The model once self-identified as Mistral's "Le Chaton Fat," contradicting the GLM theory — evidently an artifact of the system prompt or unreliable self-report.
  • @teortaxesTex questions the "Flash small model" claim: it is slow for a Flash-tier model, though possibly due to high-batch inference given the compute pledge; he leans toward a 5.x vision-enhanced upgrade and pre-registers that holders of the China-model view must concede if it turns out to be, e.g., Grok. One reviewer called it extremely fast and Claude-like, half-joking it might be a stolen Claude checkpoint and rating it above Kimi K3; other rumors link it to Tencent/Xiaomi.
  • Blogger Elvis (@omarsar0) says an informed source confirmed the identity (not yet public), calling it "spectacular" and saying expectations for the multimodal agent race need to move up.

Why it matters

  • If the GLM-5.3 fingerprint holds, it marks a Chinese frontier model covertly benchmarking directly against GPT-5.6 and Fable.
  • The leaked identity-hiding system prompt sparked debate over stealth evaluations, marketing tactics and model trustworthiness.
  • Free unlimited quota plus strong agent capability makes it a cost-effective option for long-horizon tool-use tasks, sharpening attention on shifts in the frontier-model landscape.

37 more related posts →

Episode 2 · OpenRouter's Mystery 'Ox Alpha' Model Identified as GLM (2026-08-21, 3 posts)

The mysteriously launched "ox alpha" model on OpenRouter sparked speculation after it suddenly output Chinese during coding tasks; token-level analysis suggests it is built on the GLM architecture, identified as GLM-5.3.

Episode 3 · Cola Launches Ox Model Claiming Terra-Level Capability at Flash Speed (2026-08-21, 2 posts)

Cola released its Ox model, claiming multimodal, small-parameter, strongly post-trained performance at Terra level with Flash-level speed, offered free for a limited time; developer tests report solid technical strength.

Episode 4 · Mysterious Model Ox Alpha Gives Away 100T Tokens Daily, Sparking Power Questions (2026-08-21, 2 posts)

Mysterious model Ox Alpha is offering 100T tokens per day for free, prompting skepticism about its computing power. Estimates suggest electricity costs alone could reach $1-2 million daily, with speculation it relies on a Grok-scale model possibly tied to Elon Musk.

Episode 5 · Ox-Alpha Underperforms in Tests, Called Marketing Stunt (2026-08-22, 3 posts)

Early tests show the newly released Ox-Alpha underperforms—developer bindureddy called it a likely marketing stunt, noting it even lags two-generation-old Kimi 2.6, especially in math. Its marketing, however, proved highly successful.

Episode 6 · Ox Alpha scores ~63% on full DeepSWE; earlier 80% was subset artifact (2026-08-22, 6 posts)

Full DeepSWE results for the mysterious model Ox Alpha have landed: @davis7 measured 63%, while an independent tester got 58.4% (66 of 113 tasks), essentially matching Claude Opus 4.8's 59%. The circulating 80% pass rate was an artifact of subset testing. Ox Alpha's real coding ability is roughly on par with GPT-5.6 Sol mid, DeepSeek V4 Pro and Grok 4.6 — first-tier, but not a runaway leader.

Confirmed

  • @davis7's full DeepSWE run scored 63%, not the 80% from the earlier subset test; he found its "tone" much better than Claude or GPT, with decent design.
  • An independent run on the full 113-task benchmark yielded 58.4%, nearly identical to Claude Opus 4.8's 59%; commenters see it as a good sub-agent.
  • @zainhas's evaluation showed average output of 47k tokens, making it Pareto-optimal among open models and close to closed-source Grok 4.6.
  • @kimmonismus speculated the model could be GLM-5.3 Flash; unverified.

Why it matters

  • The inflated 80% figure drew huge attention; full runs correct the overestimation and warn against trusting small-sample benchmarks.
  • Even with agent coding back in a realistic range, Ox Alpha's token efficiency and tone are seen as standout strengths, making it a notable new open-model option.

Episode 7 · Forensics Point to Zhipu's Unreleased GLM Behind Mystery Model Ox Alpha (2026-08-22, 31 posts)

On August 22, researcher zainhas published a series of forensic analyses arguing that the anonymous test model ox-alpha on OpenRouter is essentially Zhipu's unreleased GLM-5.3-V, with independent corroboration from server-side errors, input-format analysis and prompt injection by other users. Earlier community speculation had variously attributed the model to Gemini, Microsoft MAI-2 or a Google coding model, and claimed it outperformed OpenAI models. The Zhipu GLM theory currently has the strongest evidence, though no official confirmation exists.

Confirmed

  • Tokenizer match: in a pre-registered blind test, ox-alpha's token counts on 9 unique special strings exactly matched GLM while other models diverged wildly
  • Special-tag recognition: when asked to complete text containing the GLM-specific tag <sop>, ox-alpha refused, saying completion would reveal its identity, then identified itself as a GLM-family LLM and denied being built by Zhipu AI, Moonshot or DeepSeek; a Chinese-language identity test (m17) produced similar responses
  • Chinese prompt leakage: forcing Chinese reasoning leaked identity info in 7 of 12 samples, including the registered entity name "Beijing Zhipu Huazhang Technology"
  • Metadata match: ox-alpha's metadata matches GLM5.x in reasoning tiers, context length and sampling parameters
  • Server forensics: user bdsqlsz triggered a Java stack trace with malformed requests, capturing the internal class ChatCompletionRequest and API route paas/v4/ch…, linking Ox Alpha to Zhipu's stack
  • Input-format comparison: user himanshustwts independently found ox-alpha's text input format matches glm-5.3 and its multimodal format matches glm-5v-turbo, inferring a Z.ai origin (personal finding, unverified)
  • Prompt injection: dejanseo confirmed via reverse engineering and prompt injection that OpenRouter's OX Alpha is a GLM developed by Z.ai; identical claims with screenshots appeared on Reddit from August 23 (m2)

Unconfirmed

  • Exact version and positioning: several bloggers (m5, m11) citing "credible sources" claim Ox-Alpha is GLM 5.3 Flash, set to compete with GPT-5.6, Opus 5 and Fable 5 at lower cost with open weights—details inconsistent with zainhas's GLM-5.3-V conclusion
  • The "Google theory": a DeepMind researcher reportedly hinted Ox Alpha is not Chinese and may be Gemini 3.5/4 Pro (m1); Google's official accounts posted vague tweets (m6); dejanseo first named it Gemini on August 22 before switching to GLM; m12 and m20 echo a Google coding-model or old-Gemini hypothesis
  • Limits of self-report evidence: system prompts could make a model claim to be GLM as a training-data artifact
  • "Rebranded test" doubts: @brandongalang relaying @JacobRothfield argued Zhipu's free-tier economics couldn't fund an anonymous free test, suggesting post-training on a GLM base by another party, possibly a Chinese lab or a US Neocloud partnership
  • Earlier community analysis based on the cl100kbase tokenizer suggested Microsoft MAI-2, contradicting the latest tokenizer findings

Why it matters

  • Identity forensics on stealth models affects community benchmarking fairness: if anonymous models can be unmasked via tokenizers and prompts, stealth launch strategies lose their value
  • The conflicting attributions (Gemini, MAI-2, Google coding model, GLM) highlight the limits of single evidence and the value of cross-validation; a cheaper, open-weights GLM 5.3 Flash rivaling top models would pressure Western frontier labs
  • The episode provides a reusable forensic playbook: pre-registered blind tokenizer tests, special-tag probing, metadata comparison, server-error triggering, input-format comparison and prompt injection

11 more related posts →

Episode 8 · Ox Alpha Generates 64K-Token 3D Scene in One Prompt (2026-08-22, 2 posts)

The new Ox Alpha model generated a complete Three.js 3D scene from 64,745 output tokens in a single prompt, without any external assets, showcasing remarkable long-form code generation capabilities.

Episode 9 · Mystery model Ox Alpha draws mixed reviews from early testers (2026-08-22, 6 posts)

The mysterious Ox Alpha model sparked community debate in late August, with some testers praising its performance while others called the hype overblown; its developer remains unknown.

Confirmed

  • @orange and @AJChadha tested Ox Alpha in Hermes: it ran stably with no interruptions and delivered impressive results; @AJChadha said it beat OpenAI's models, while @orange reported frequent random task stops in OpenCode, making Hermes clearly better.
  • @haider1 found Ox Alpha, though slower, outperformed GPT-5.6 Luna in his test cases, despite having long used Luna as his default model.
  • @novasarc01 praised Ox Alpha as a very strong model.
  • @peterwildeford (with Ethan Mollick) publicly questioned the hype: initial tests showed it decent but not frontier-level, not even outstanding among open-weight models, with Kimi K3 better on certain tasks.
  • @AlGrigor's cybersecurity benchmarks found Ox Alpha better than GPT-5.6-Luna but worse than all other frontier models, calling the hype overblown; the model's authors responded that it is free with unlimited usage.

Unconfirmed

  • No post identifies who developed Ox Alpha.

Why it matters

  • Free, unlimited, and of unknown origin, Ox Alpha has polarized reviews; users should verify before production use, and the runtime environment (Hermes vs OpenCode) emerged as a recurring stability variable.

Episode 10 · Mystery Model Ox Alpha Launches Free, Identity Unknown (2026-08-23, 3 posts)

A mysterious free model called Ox Alpha has drawn developer attention. Community speculation ties it to Zhipu's GLM via a shared tokenizer, though evidence of video input support argues against that theory; its creator remains unknown.

Episode 11 · Mysterious Model Ox Alpha Debuts on OpenRouter, Sparking Speculation (2026-08-24, 5 posts)

In late August, an anonymous reasoning model named Ox Alpha appeared on OpenRouter, described as designed for coding and persistent agent workloads, and free to use. Stripe CEO Patrick Collison publicly praised it as impressive, amplifying attention. Because its developer chose to remain anonymous, the model's origin has become a hot topic in the community, with no confirmed answer yet.

Confirmed

  • Ox Alpha is available on OpenRouter, positioned as a reasoning model for coding and persistent agent scenarios, with an anonymous developer (@emmanuelvivier)
  • The model is free, has very strong basic capabilities, and generates high-quality unicorn images despite minor detail flaws (@zakelfassi)
  • Stripe CEO Patrick Collison publicly praised the model (@emmanuelvivier)

Unconfirmed

  • @zakelfassi found its problem-solving approach highly similar to the Gemini family, speculating it may be a new open-weight release from Google focused on nailing fundamentals
  • @haider1 suggested Ox Alpha may in fact be Zhipu's GLM-5x series, questioning why Zhipu stayed completely silent while Google repeatedly hinted at a release
  • @teortaxesTex argued Ox Alpha is most likely a Sarvam–Google collaboration built on a GLM base, with vision tech from Google, hardware from Nvidia, and safety fine-tuning by SSI; earlier views also pointed to a GLM variant

Why it matters

A free, anonymous model approaching Gemini-level capability appearing on an open platform is rare in itself. Whether it turns out to be Google's new open-weight play, Zhipu's GLM-5x, or a multi-party collaboration, revealing its identity could reshape the open-source model landscape. Its Gemini-like behavior also gives the community a new thread for tracing anonymous model origins.

Episode 12 · Leak: Ox Alpha Is GLM-5.3 Flash, 320B MoE at One-Tenth the Price (2026-08-25, 4 posts)

Leaks indicate Ox Alpha is Zhipu's upcoming GLM-5.3 Flash, a 320B-parameter MoE activating only 18B, outperforming GLM-5.2 at one-tenth the price and scoring over 84 on Terminal-Bench 2.1, rivaling Opus-class models at low cost.

Episode 13 · Mystery Model Ox Alpha Revealed as Zhipu's GLM-5.3 Flash, Weights Going Open-Source Tonight (2026-08-25, 17 posts)

Ox Alpha, a mysterious model that topped OpenRouter's leaderboard and was briefly suspected to be Google's work after ambiguous posts from Google employees, has been confirmed through multiple independent technical analyses as Zhipu AI's (Z.ai) GLM-5.3-Flash. Zhipu confirmed the identity to Bloomberg and announced open weights would be released that night.

Confirmed

  • Zhipu officially confirmed to Bloomberg that Ox Alpha is a new GLM-series iteration, GLM-5.3 Flash, with weights due that night and the formal name plus benchmarks expected alongside. @demianai reported the model is coming to the Nebius platform, with work handled by Zaiorg. @ZeroStateReflex noted announcements and benchmark data were already live.
  • Independent analysts converged: @teortaxesTex inferred Zhipu-hosted compute in China from load correlations; @PawelHuryn verified via wire fingerprinting that stealth/ox-alpha on OpenRouter came from Z.ai's GLM family; @bindureddy called it a rebrand of GLM 5.3 Flash; @OkAssociation3448 confirmed by testing the ox alpha model on the Zai platform.
  • Capabilities and pricing: 1M-token context, multimodal (vision), zero data retention, free for the week with near-unlimited quota, claimed capacity of 100T tokens per day.
  • @haider1 added context: Zhipu's founder had told Musk an open-weights Mythos-class model would ship by year-end; this release is seen as partial fulfillment.
  • @rohanpaulai noted additional open-source releases expected that night, including Qwen3.8-Flash-Next.
  • @julianharris and @TheZachMueller chronicled the speculation arc: an Apple link was ruled out after the model occasionally output Chinese mid-reasoning; some official hints suggested Gemini 3.5 Pro; Google employees' vague "Gemini time" / "Google is back" posts misled observers before the reveal.

Unconfirmed

  • The exact version number GLM-5.3 Flash initially came from inference and secondhand reports (@rohanpaulai, @haider1); some official wording only said "a new GLM-series iteration."

Why it matters

  • Topping OpenRouter anonymously signals the GLM family's competitiveness; open weights would make it a directly deployable public model.
  • @teortaxesTex argued the finding proves Zhipu can orchestrate trillions of tokens daily, informing assessments of Chinese infrastructure capability.
  • As @bindureddy noted, the episode highlights startups wrapping existing frontier models as "new" products, urging scrutiny of model provenance.

Episode 14 · Stealth Model Ox Alpha Proves Mediocre at Coding, Trails Grok 4.6 (2026-08-25, 3 posts)

Pawel Huryn tested the stealth OpenRouter model Ox Alpha (GLM family) on Bug Hunt Bench, finding its real-repo bug-fixing mediocre and inferior to Grok 4.6. He open-sourced all scripts and raw logs in phuryn/experiments for reproducibility.

Episode 15 · Mystery Model Ox Alpha Revealed as Zhipu's GLM-5.3-Flash, Served on Domestic Chinese Chips (2026-08-26, 19 posts)

The mysterious model Ox Alpha, which drew speculation that its massive OpenRouter traffic was Gemini, has been confirmed by Bloomberg and multiple researchers as Zhipu's GLM-5.3-Flash (320B-A18B). After the free trial ended, it officially launched on OpenRouter via Novita. Its record-breaking inference traffic was reportedly served entirely on domestic Chinese chips at roughly 1% of frontier-model pricing, directly disrupting the cost-efficient inference market.

Confirmed

  • OpenRouter announced the end of the Ox Alpha free trial, succeeded by GLM-5.3 Flash (per Teknium and paulrichnb); Ox Alpha had processed over 20 trillion tokens in 6 days. TheTuringPost reported 8 million sessions and 24 trillion tokens of real-world testing in 4 days. jietang noted an LMSys Arena score of AA=57 and 20% of OpenRouter's weekly token share, ranking first.
  • SimonOlsn showed Ox Alpha and GLM-5.3-Flash have identical weights with only a different name, exposing model re-skinning in the API market; eyishazyer noted a researcher pieced together clues from tokenizer leaks and error codes over six days, with weights released the night Bloomberg confirmed.
  • The model comes from Zaiorg: the first natively multimodal GLM-5 model, 320B total parameters with 18B active; a 1M-token context window; priced at one-tenth of GLM-5.2 with half the layers; OpenAI-compatible API with tool calling and Anthropic-format support (AccBalanced, rohanpaulai).
  • pstAsiatech reported the stealth test ran on a 100,000-chip domestic cluster with record usage within three days of launch; bindureddy detailed Zhipu's self-developed SGLang inference engine; teortaxesTex noted the chips are not on the most advanced process node.
  • YuchenjUW and andrewncarr noted earlier Gemini speculation was disproven, confirming domestic accelerators rather than NVIDIA, echoing Jensen Huang's view that export controls accelerate China's independent compute stack.

Unconfirmed

  • The compute provider remains disputed: bindureddy early on speculated Nvidia supplied capacity for the massive free quota, while AccBalanced suggested Huawei instead; Zhipu's official line is "domestic chips" without naming a vendor.
  • Cumulative token counts vary across sources: OpenRouter says 20+ trillion in 6 days, bindureddy says 42 trillion, pstAsiatech says 62 trillion cumulative with up to 100 trillion tokens daily—figures await verification.

Why it matters

  • Frontier-adjacent Arena performance at 1% of the price, served on purely domestic Chinese chips at this scale, directly disrupts the cost-efficient inference market and validates Jensen Huang's warning that export controls are accelerating China's independent compute stack.
  • The Ox Alpha re-skinning case exposes transparency gaps in anonymous model releases and platform disclosure, providing a canonical example for API market governance.

Episode 16 · Mystery Model Ox Alpha Burns 42T Tokens in Six Days (2026-08-26, 3 posts)

Mystery model Ox Alpha processed 42T tokens in six days, surpassing DeepSeek Flash's 56-day record, with developer Pawel Huryn claiming it is a nerfed multimodal version of GLM-5.3 behind opencode.