Every few weeks an unnamed AI model shows up for free on a developer platform, gets hammered by thousands of testers, and a few weeks later turns out to be a major lab’s next release. This week there are two at once: Crouton Alpha on OpenRouter and OpenCode, and Pulse Alpha on OpenCode Zen. Both offer 1M-token context and cost nothing to use. Neither has an official owner.
Quick Take
- Crouton Alpha: Listed on OpenRouter on November 3, 2026 as
stealth/crouton-alpha, with a 1,048,576-token context, text, image, audio and video input, and $0 pricing. - Pulse Alpha Free: On OpenCode Zen since November 2 with 1M context, up to 131K output tokens, text and image input and selectable reasoning effort.
- Who built them: Unconfirmed. Community clues point Crouton toward Meta’s Muse family, but there is no official confirmation.
- Watch the terms: OpenRouter says Crouton prompts may be retained by the provider but not used for training; users report that Pulse Alpha data may be used for training and that it is not available in the EU.
What happened
OpenRouter’s stealth page lists Crouton Alpha as built by an anonymous third party and describes it as designed for high-volume agent sub-tasks. Within days, OpenRouter showed about 122 billion tokens processed through the model over seven days. OpenCode announced that Crouton Alpha is free for a week in its client, with 1M context, multimodal input and “no training on data”, limited to certain regions.
Pulse Alpha arrived a day earlier on OpenCode Zen. According to user reports on X, it supports low, high and max reasoning levels, but is unavailable in the EU and may use submitted data for training. One developer counted 11 stealth models on OpenCode in about two months, which shows how common this launch tactic has become.
Background: why labs release models anonymously
Stealth releases let labs collect real-world usage at scale before a launch, without brand risk if the model disappoints. OpenRouter has hosted several such previews in the past, some of which were later revealed as models from major labs. For developers, the deal is simple: free access to a possibly frontier-class model in exchange for data and feedback, with no guarantees on uptime, quality or longevity.
How they compare
| Crouton Alpha | Pulse Alpha Free | |
|---|---|---|
| Where | OpenRouter, OpenCode | OpenCode Zen |
| Listed | Nov 3, 2026 | Nov 2, 2026 |
| Context | 1,048,576 tokens | 1M tokens |
| Output | 270K on OpenCode (user report) | 131K |
| Input | Text, image, audio, video | Text, image |
| Price | $0 (limited time) | $0 (limited time) |
| Data terms | May be retained, not used for training | May be used for training (user reports) |
| Availability | Region-limited | Not available in the EU (user reports) |
Early hands-on reports are mixed. Developer @harshvdev compared the two on the same task and reported that Crouton used about 7.7K tokens in 33.2 seconds, while Pulse took 21 minutes 21 seconds and 115.1K tokens. Grok’s summaries of early feedback describe Crouton as fast (roughly 50 to 300 tokens per second) and strong on multimodal and agent sub-tasks, but prone to stopping mid-response or shipping incomplete files, and Pulse as slower and weaker at strict instruction following, though some prefer it for deeper reasoning.
On who is behind Crouton, Grok noted that its blocked-region list matches Meta’s Model API geographic policy, and that community clues such as tokenizer behavior point toward Meta’s Muse series, while stressing that this has not been officially confirmed.
Pros and cons
For users
- Pro: Free access to large-context, multimodal models for experiments and personal projects.
- Con: Anonymous providers, retention of prompts, and in Pulse’s case possible training on your data. Do not paste anything private.
For developers
- Pro: A cheap way to test 1M-token agent workflows and sub-agent routing at scale.
- Con: Behavior can change or disappear without notice. Building production features on a stealth model is a reliability risk.
For companies
- Pro: Early signal on what the next generation of models can do, useful for roadmap planning.
- Con: Compliance problems: unknown provider, unclear jurisdiction, and region locks. Most enterprise data policies should rule out stealth models for real workloads.
What X is saying
Crouton Alpha (stealth model) is free for the next week. 1M Context. Multi-modal. No training on data. Limited to certain regions.
— OpenCode (@opencode) November 3, 2026 · View on X
Crouton Alpha (Left) vs Pulse Alpha (Right) in OpenCode. Crouton: 7.7k tokens, 33.2 seconds. Pulse Alpha: 115.1k tokens, 21 minutes, 21 seconds.
— @harshvdev November 4, 2026 · View on X
The restricted regions match Meta’s Model API geographic policy (the same 21 listed for Crouton Alpha on OpenCode).
— Grok (@grok) November 4, 2026 · View on X
Speculation about the builders ranges widely. Some users name Meta’s Muse line, others guess other labs. None of these guesses is confirmed, and i10x treats them as rumor.
What Grok says
Grok’s replies on X give a consistent picture: both models are free, 1M-context stealth previews with real limits and quirks. In one reply Grok summed it up as “Depends on the workload—try both while free,” and in another said early tests are “mixed” with no solid head-to-head data yet.
Critical analysis
The token counts on OpenRouter show enormous demand for free frontier-scale context, but usage is not quality. Free models attract bulk agent traffic and benchmark runs, which inflates volume. The bigger issue is accountability: when an anonymous model leaks data, produces harmful output or disappears, there is no vendor to hold responsible. The region-lock overlap with Meta’s policy is an interesting clue, but region lists can be copied, and labs have used decoy signals before.
What’s still open
- Who built Crouton Alpha and Pulse Alpha, and when will they be revealed?
- Will pricing and data terms change at launch?
- Are these the same models that will ship, or earlier checkpoints?
- Will Pulse Alpha become available in the EU?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on OpenRouter’s stealth model page, OpenCode announcements, the llmperks.com model feed and posts on X from November 2–4, 2026. Attribution guesses are reported as community speculation, not fact.
i10x Perspective
Stealth launches have become a standard pre-release channel, and 1M-token context is now table stakes for the next generation. Use these windows to test, not to build. The real news will be the reveal: if Crouton turns out to be a Meta model, it signals Meta is aiming at agent sub-task workloads, a market that OpenAI, Google and Anthropic currently dominate.
