The mystery is over. Sherlock Think Alpha and Sherlock Dash Alpha, the two unbranded models that appeared on OpenRouter in November 2025, were early snapshots of Grok 4.1 Fast.
For about a week, nobody outside xAI and OpenRouter knew that for certain. Two models showed up with no press release and no attribution. The AI community filled the gap with theories. Some pointed to Grok 4.20. Others pointed to Gemini 3, given the timing and the context window size.
The resolution itself is a small story. The pattern behind it is not. Cloaked model launches are becoming a routine part of how frontier labs test new models before committing to a public release, and this case is a clean example of how that process plays out in public view.
What Appeared on OpenRouter
Two models landed on OpenRouter without an official name attached:
| Model | Type | Context Window | Notes |
|---|---|---|---|
| Sherlock Think Alpha | Reasoning | ~1.8 to 2M tokens | Multimodal, strong tool calling |
| Sherlock Dash Alpha | Non-reasoning | ~1.8 to 2M tokens | Faster, same context size |
Neither model carried a provider name at launch. This is a known pattern on OpenRouter, sometimes called a cloaked model. Labs use it to run a new model past real users before deciding on a public release and a name.
Why the Community Guessed Grok and Gemini
Two theories dominated the discussion.
The Grok theory rested on two observations. The output style leaned heavily on emoji, which matched prior Grok behavior. There was also a direct precedent: Grok 4 Fast had gone through the same cloaked process earlier, appearing on OpenRouter as Sonoma Sky Alpha and Sonoma Dusk Alpha before its September 2025 launch.
The Gemini theory rested on the context window and the calendar. A window in the 1.8 to 2M token range is large, and the timing lined up with the expected release window for Gemini 3.
A separate detail circulated as color rather than proof. Someone tested one of the Sherlock models by asking it to write code for a PS5 controller interface, and it became a widely shared anecdote. It is worth mentioning because it shaped the conversation, but it was never a benchmark claim and should not be read as one.
| Theory | Supporting evidence | Outcome |
|---|---|---|
| Grok 4.20 | Emoji-heavy output style; direct precedent from Sonoma Sky/Dusk Alpha becoming Grok 4 Fast | Right on the lab, wrong on the model name |
| Gemini 3 | Large 1.8 to 2M token context window; timing near Gemini 3's expected release | Incorrect |
| PS5 controller test | Widely shared anecdote, not a benchmark | Color only, no bearing on identity |
What It Actually Was
OpenRouter's own model pages now confirm both Sherlock models. Sherlock Think Alpha was an early snapshot of Grok 4.1 Fast with reasoning enabled. Sherlock Dash Alpha was the same snapshot with reasoning disabled.
Timeline
| Date | Event |
|---|---|
| November 15, 2025 | Sherlock Think Alpha and Sherlock Dash Alpha appear on OpenRouter with no attribution |
| November 19, 2025 | xAI officially launches Grok 4.1 Fast |
| November 21, 2025 | OpenRouter updates both model pages to confirm the identity |
Why Labs Run Stealth Launches
Cloaked releases give a lab a way to collect real usage data before a branded launch carries the weight of an official announcement. Feedback comes in from actual traffic and actual prompts, without the pressure of a public benchmark comparison attached to the company name.
The direct precedent here is Sonoma Sky Alpha and Sonoma Dusk Alpha, which tested on OpenRouter before becoming Grok 4 Fast. The gap between the cloaked test and the branded launch was about two weeks in that case, which is close to what played out with Sherlock Alpha and Grok 4.1 Fast.
This is not unique to xAI. Other labs have used similar unbranded testing patterns on OpenRouter, and it is reasonable to expect this pattern to continue as more labs adopt faster release cycles.
| Cloaked name | Later confirmed as | Cloaked appearance | Official launch | Gap |
|---|---|---|---|---|
| Sonoma Sky Alpha / Sonoma Dusk Alpha | Grok 4 Fast | September 5, 2025 | September 19, 2025 | ~2 weeks |
| Sherlock Think Alpha / Sherlock Dash Alpha | Grok 4.1 Fast | November 15, 2025 | November 19, 2025 | ~4 days |
What This Means for Developers and Marketers
A few practical takeaways from this cycle:
| Takeaway | Why it matters |
|---|---|
| Cloaked models offer free or low-cost access to near-frontier capability during the testing window | Pricing and rate limits are not finalized yet, so access can be cheaper than the eventual official release |
| Do not build production workflows on a model with unconfirmed identity, pricing, or availability | A cloaked model can be pulled or repriced without notice |
| Treat stealth launches as a recurring signal, not a one-off curiosity | Each cloaked model is an early preview of a capability shift that reaches production models within days or weeks |
Where MigmaAI Fits Into This Story
Migma AI's core use case is AI-generated email campaigns and content at scale. That use case depends on one thing improving continuously in the background: the model layer underneath it.
Stealth launches like Sherlock Alpha are an early signal of how fast that layer moves. A new model shows measurable gains in reasoning or tool calling before it even has a public name. Within days, that improvement filters into every product built on top of it, whether or not the product ever mentions which model it runs on.
MigmaAI is not tied to any single model provider. The value of a tool like Migma comes from staying ahead of shifts at the model layer rather than depending on one vendor's release schedule. A few points on why that matters:
- Model releases are no longer spaced far apart. The Sonoma to Sherlock gap alone went from about two weeks to about four days between cloaked testing and official launch.
- A platform built around one vendor inherits that vendor's release calendar, including any slowdown, price change, or deprecation.
- A platform built to work across the model layer picks up gains from wherever they show up next, without a rebuild.
What this looks like in practice, feature by feature:
- Campaign generation. As reasoning and instruction-following improve across the model layer, generated email sequences follow brand voice and campaign structure more consistently, with less manual editing per send.
- Personalization at scale. Better tool calling and larger context windows, the same gains being tested in cloaked models like Sherlock Alpha, support pulling in more customer context per email without losing coherence across a batch.
- Turnaround time. Faster non-reasoning variants, the same category as Sherlock Dash Alpha, support quicker draft generation for high-volume campaigns where speed matters more than deep reasoning.
- Consistency across a content calendar. Stronger long-context handling supports keeping tone and messaging aligned across a full sequence of emails rather than drifting after the first few.
Why this matters for teams already using AI content tools:
The underlying quality of AI-generated content keeps moving, whether a team tracks it or not. A platform that abstracts that complexity away, so users get the benefit of model improvements without needing to track every stealth launch or benchmark cycle themselves, saves real time. Migma AI's reported 89% time savings reflects that same principle: less time spent managing tools and comparing models, more time spent on the output itself.
Closing
The Grok 4.20 versus Gemini 3 guessing game is a useful snapshot of how fast rumor cycles move in frontier AI right now. A model can appear on OpenRouter with no name, get tested by thousands of users, get misattributed to two different labs, and get correctly identified, all within about a week.
That pace is not a one-time event. It is the current baseline for how fast the model layer moves. Every cloaked launch like Sherlock Alpha is a preview of gains that will show up in production tools within days, whether or not any single team is tracking the release cycle closely enough to notice.
Tracking every stealth launch, benchmark, and model update is not a reasonable use of a marketing team's time. Getting the benefit of those improvements without doing that tracking is.
Ready to stop managing the model layer and start managing your output?
MigmaAI generates email campaigns and content at scale while staying ahead of shifts across the model layer, so your output quality keeps improving without any extra work on your end. Try MigmaAI today and see the reported 89% time savings for yourself.
Start with MigmaAI→ https://migma.ai/
For further reading, see the official Grok 4.1 Fast launch coverage or the Gemini 3 launch coverage.
Frequently Asked Questions
Is Sherlock Alpha still available on OpenRouter?
Sherlock Think Alpha and Sherlock Dash Alpha were early testing snapshots and are no longer listed as active cloaked models now that Grok 4.1 Fast has an official release. Grok 4.1 Fast is available directly through xAI and through OpenRouter under its own name.
What was the pricing during the cloaked testing window?
Cloaked models on OpenRouter are typically offered free or at reduced cost during the testing period, since the goal is to gather usage data rather than generate revenue. Pricing changes once the model launches under its official name.
What was the context window size?
Both Sherlock models carried a context window in the 1.8 to 2M token range, consistent with the context window later confirmed for Grok 4.1 Fast.
How does Grok 4.1 Fast perform after launch?
Grok 4.1 Fast launched as a speed and cost-optimized model built on the Grok 4 architecture, with gains over Grok 4 Fast in several benchmark categories. It was positioned as an efficiency-focused release rather than a generational leap.
Why did the Gemini 3 theory gain traction?
The large context window and the timing near Gemini 3's expected release window were enough to make Gemini 3 a plausible guess, even though the two models turned out to be unrelated.
