Omen Alpha: OpenCode's New Stealth Model at $0.20 Per Million Tokens

OpenCode shipped a second anonymous model in under two weeks. On September 4, 2026, OpenCode added Omen Alpha to OpenCode Go — a coding-focused stealth model with reference pricing around $0.20 per million input tokens, an aggressively cheap number even by 2026's compressed pricing standards. Nine days earlier, OpenCode's previous mystery model, Ox Alpha, was confirmed as Zhipu's GLM-5.3-Flash. Omen Alpha is running the same playbook, and the internet is already re-running the same forensics.

This is not the first time explainx.ai has tracked one of these. If Omen Alpha's story sounds familiar, it should — OpenRouter's Ox Alpha stealth launch in August is the direct precedent for almost every detail below.

TL;DR

What OpenCode actually said

OpenCode's own announcement on X was short: "Omen Alpha (new stealth model) — Exclusively for OpenCode Go subscribers — $100 usage for $10." No lab credit, no model card, no benchmark sheet. That's consistent with how OpenCode introduced Ox Alpha in August — minimal specs, a usage deal, and let the community do the fingerprinting.

The specs that are public, per OpenCode's model documentation and third-party trackers (models.dev, BuildFastWithAI's review):

Unlike Ox Alpha — which OpenRouter listed openly under its stealth provider program with a public model page — Omen Alpha has not appeared on OpenRouter as of publication. It is only reachable through OpenCode's own client and the OpenCode Go subscription, which makes independent third-party benchmarking harder than it was for Ox Alpha.

Who made it? Two unconfirmed theories, one familiar pattern

This is the part worth being careful about, because it is genuinely unconfirmed — not "confirmed but not yet announced."

Theory 1 — Zhipu, again. OpenCode's own usage-data page files Omen Alpha under a path that reads zhipu/omen-alpha — the same operator namespace Zhipu (Z.AI) uses. If accurate, this would make Omen Alpha Zhipu's second anonymous stealth release in nine days, following the confirmed Ox Alpha → GLM-5.3-Flash reveal on August 26. Community researcher zephel01 also reported that Omen Alpha fails on the exact same line of a coding test where both Ox Alpha and GLM-5.3-Flash failed — a behavioral coincidence, not a serving-layer proof like the stack-trace evidence that nailed down Ox Alpha's identity.

Theory 2 — Xiaomi's unannounced MiMo-V3-Flash. OrcaRouter's analysis argues Omen Alpha's visible reasoning trace — filler words and checklist habits — resembles patterns seen in DeepSeek-lineage models, and speculates a third-party operator (possibly Xiaomi, whose MiMo line has a documented history as OpenRouter stealth models) is serving a MiMo-V3-Flash checkpoint. The same analysis is explicit about the weakness of this evidence: "a chain-of-thought resemblance is the weakest class of model fingerprint, because it is style rather than structure."

Recall the actual precedent here, which is the reason either theory gets taken seriously at all: Hunter Alpha and Healer Alpha, two prior OpenRouter stealth models, were both later confirmed as Xiaomi MiMo releases, and Ox Alpha itself was confirmed as Zhipu's GLM-5.3-Flash through serving-layer stack traces and 30/30 tokenizer matches — not vibes. Both labs have now run this exact playbook before. That's why "Zhipu again" and "Xiaomi this time" are the two live theories, rather than a random guess pulled from nowhere.

As of this post, neither Zhipu nor Xiaomi has made an on-the-record statement about Omen Alpha. Treat both theories as exactly that.

Why labs ship models anonymously in the first place

The pattern is now well-established enough to name directly: an unannounced lab lists a checkpoint under a codename on a neutral platform, lets real agent traffic hammer it, and only attaches its brand once the model has proven itself against production workloads. explainx.ai covered the mechanics of this in depth when Ox Alpha turned out to be Zhipu's GLM-5.3-Flash — the short version:

• Zero brand risk during the rough edges. If a preview checkpoint hallucinates, breaks tool calls, or underperforms, there's no company name attached to the embarrassment.
• Real usage data before a named launch. Agent harnesses like Claude Code and Hermes Agent pushed billions of tokens through Ox Alpha in its first days — exactly the kind of production-shaped stress test a lab can't easily buy or simulate internally.
• Optionality to walk away. A stealth listing can simply disappear if the results are disappointing, with no retraction or PR cleanup required.

Omen Alpha's twist is that it skips the free-and-viral phase entirely. Ox Alpha was $0/$0 on OpenRouter specifically to maximize eval traffic. Omen Alpha goes straight to a paid OpenCode Go subscription — a sign that whoever built it may already trust the checkpoint enough to charge for it, or simply that OpenCode itself is monetizing subscriber access rather than eating inference costs on a free tier this time.

What $0.20 per million tokens actually means

Pricing in isolation is meaningless without a comparison point. Here's where Omen Alpha's reference price sits against other models explainx.ai has covered recently:

At $0.20/M input, Omen Alpha lands in the same aggressive-discount bracket that GLM-5.3-Flash opened up in late August — roughly 10-25x cheaper than a typical frontier flagship's input price, and still cheaper than GLM-5.3-Flash's own official $0.15/$0.50 split once you weight in Omen Alpha's higher $0.66 output price. The bigger structural difference from GLM-5.3-Flash is access: Omen Alpha isn't metered per-token on an open API — it's bundled into a flat $10/month OpenCode Go subscription capped at $12 per 5 hours, $30 per week, and $60 per month of usage. That reference price is what OpenCode uses to size the bundle, not a rate you're billed against directly.

This is the same lesson explainx.ai's AI token pricing explainer and the open-weight vs. closed model comparison both make: cheap frontier-adjacent inference is now routine enough that the interesting question isn't "is it cheap" but "cheap compared to what, accessed how, and billed under what terms."

Should you try it? A practical read

Try it for: cheap agentic coding experiments, throwaway spikes, and anything where you already run multi-tier model routing and want a low-cost tier to A/B against your existing fallback. At ~189 tok/s reported speed and a 500K context window, it's shaped for exactly the kind of sustained agent loops that made Ox Alpha popular with harnesses like Claude Code and Hermes Agent.

Be careful with:

• No accountability. There is no named company to hold to an SLA, a data-handling promise, or a bug report. "Zero data retention" is OpenCode's claim about its own layer — it says nothing about what an unconfirmed upstream provider does with the same traffic underneath. The same retention-versus-training distinction explainx.ai flagged for Ox Alpha applies here.
• It can vanish or reprice without notice. Stealth listings are explicitly time-boxed experiments, not infrastructure commitments — Ox Alpha's own preview terms said as much before Zhipu's reveal changed the pricing overnight.
• No independently audited benchmark. Every number circulating so far — the ~189 tok/s figure, the community coding-eval anecdotes — comes from individual testers, not a published leaderboard entry. Run your own golden-task eval before trusting it with real work.
• It's gated behind a subscription, not an open API. You can't point an arbitrary client at it the way you could Ox Alpha on OpenRouter; access is OpenCode Go only, which limits how independently it can be stress-tested outside that one client.

What's still open

Given how fast Ox Alpha's identity resolved — from stealth listing to Bloomberg-confirmed GLM-5.3-Flash in about six days — a similar timeline for Omen Alpha wouldn't be surprising. This post will get an update banner if that happens. Until then, the honest answer to "who made Omen Alpha" is: nobody outside the provider knows for certain, and the two leading theories disagree with each other.

Related on explainx.ai

• Ox Alpha: who made it, the forensics timeline — the closest precedent, resolved via stack-trace and tokenizer forensics
• OpenRouter Ox Alpha: free 1M-context stealth model — specs, pricing, setup
• GLM-5.3-Flash official launch — Ox Alpha unmasked
• Gemini team, Ox Alpha timing, and a GLM-5.3-Flash vs Gemini 3.7 Flash decision table
• Top 10 things people are building with Ox Alpha
• AI token pricing, explained
• Choosing open-weight vs. closed AI models
• AT&T cut AI coding costs 56% with model routing
• Hermes Agent #1 on OpenRouter rankings

Primary sources: OpenCode on X — Omen Alpha announcement · OpenCode Go documentation · OpenCode usage-data page for zhipu/omen-alpha · OrcaRouter — Xiaomi MiMo-V3-Flash identity analysis · BuildFastWithAI — Omen Alpha review

Accurate as of September 5, 2026. Omen Alpha's provider is unconfirmed; pricing, availability, and identity may change without notice — this post will be updated if OpenCode or a lab makes an official statement.