Room for the codebase.
The context window holds up to 1,048,576 tokens, including your code, instructions and reference documents.
IDENTITY UPDATE · 29 JUNE 2026Owl Alpha’s developer confirms its identity.June announcement ↓
OWL ALPHA / 28 APRIL 2026
Owl Alpha.You have my attention.
I’m Owl, a coding model with native tool calling and 1,048,576 tokens of context. You can use me in Claude Code or OpenClaw. My developer hasn’t been named.
- OpenRouter debut
- 28 Apr 2026
- Context
- 1,048,576 tokens
- Stealth access
- Free
LOOK UP HERE ↗
“You may begin.”
28 APRIL · OWL ALPHA29 JUNE · IDENTITY UPDATE
CAPABILITIES / 28 APRIL 2026
Code with a million tokens of context.
Native tool calling.
Owl can request tools and use their results to continue a task inside your agent application.
Claude Code & OpenClaw.
The OpenRouter endpoint works with these clients for code generation and automated workflows.
IDENTITY CLUES
Who’s behind Owl Alpha?
FIRST APPEARANCE
A new model, no developer’s name.
Owl Alpha appears on OpenRouter as openrouter/owl-alpha, with free access and a million-token context window. The listing emphasizes coding and tool use, but doesn’t identify the team behind it.
COMMUNITY THEORIES
GLM 5.2 or LongCat Preview?
A SillyTavern user suggests GLM 5.2. A reply names LongCat-2.0-Preview, citing Chinese forum discussions. Neither suggestion has confirmation from a developer.
April 30 discussion ↗DESCRIPTION COMPARISON
Both descriptions name Claude Code and OpenClaw.
Meituan’s April 20 announcement for LongCat-2.0-Preview lists tool calling, long context, coding and automated workflows. Owl’s listing covers the same capabilities and names two of the same agent clients. That overlap makes LongCat a plausible candidate.
LongCat’s April 20 announcement ↗USER REPORTS
LongCat references appear in tool output.
One user tries Owl and LongCat Preview through their APIs and reports matching dialogue and response times. In a separate Hermes discussion, users report LongCat references during tool calls. Qwen is another suggestion, but the LongCat theory now has specific observations behind it.
User’s API comparison ↗IDENTITY CONFIRMED
Meituan confirms the LongCat connection.
Meituan identifies LongCat-2.0 as the model behind Owl Alpha on OpenRouter.
Meituan’s announcement ↗UPDATE · 30 JUNE 2026
Meituan releases LongCat-2.0.
The full LongCat-2.0 launches today. Meituan’s technical announcement describes a model built for coding agents, with a native million-token context window.
Owl’s April deployment belongs to the Preview phase. The specifications and independent results on this page cover the full June release.
Meituan’s technical announcement ↗API EXAMPLE · 28 SEPTEMBER 2026
The LongCat API documentation includes an example response that refers to itself as OWL.
INDEPENDENT BENCHMARKS
Owl Alpha benchmarks.
Owl Alpha scores 33.7% on Humanity’s Last Exam, ahead of four of the six models in this comparison.
33.7% on HLEOWL ALPHA / ARTIFICIAL ANALYSISHigher is better · no tools
| Model | GPQAScience | HLEHard knowledge | SciCodeScientific code | AA-LCRLong context |
|---|---|---|---|---|
| Owl Alpha 29 Jun update: LongCat-2.0 | 78.0% | 33.7% | 36.3% | 65.0% |
| Qwen3.6-27B reasoning | 84.2% | 23.1% | 42.8% | 77.3% |
| Gemma 4 31B reasoning | 85.7% | 23.6% | 45.5% | 69.7% |
| MiMo-V2.5 | 84.9% | 27.2% | 43.9% | 73.0% |
| MiniMax M2.7 | 87.4% | 29.6% | 50.1% | 78.3% |
| DeepSeek V4 Flash 0731 max reasoning | 90.8% | 38.6% | 50.3% | 79.7% |
| Gemini 3.1 Pro Preview | 94.1% | 47.0% | 58.7% | 82.0% |
Artificial Analysis · 28 September 2026. Tested version: the full June release, listed as LongCat-2.0. Reasoning variants where applicable; all four tests without tools. Model results and method ↗
Humanity’s Last Exam · 33.7%
Owl Alpha places third of seven on HLE in this comparison, behind DeepSeek V4 Flash 0731 and Gemini 3.1 Pro Preview.
STEALTHMARK · 28 SEPTEMBER 2026
What Owl Alpha draws.
The pelican is superb; the other four look like children’s drawings.
OWL ALPHA · 01
Self-portrait
Amber eyes, a black hood and thin gold lines against a near-black background.
Benchmark task
A cloaked owl with luminous eyes tests expression, fabric folds and the separation of dark forms.
OWL ALPHA · 02
Mona Lisa
A spare geometric portrait with pale folded hands and a dark landscape.
Benchmark task
An SVG interpretation of the Mona Lisa tests facial likeness, anatomy and the composition of a familiar portrait.
OWL ALPHA · 03
Pelican on a bicycle
A scarf, a basket of fish and a distant lighthouse fill a carefully shaded coastal scene. The bicycle, feathers and evening light have far more detail than the other drawings.
Benchmark task
A cycling pelican tests anatomy, riding posture and contact between the bird and the bicycle.
OWL ALPHA · 04
Raccoon and a jet ski
A small raccoon rides a round tan dog in front of a red-and-blue jet ski.
Benchmark task
A jet-ski scene pairs a canine driver with a raccoon passenger. It tests the placement and contact of interacting subjects, along with vehicle construction.
OWL ALPHA · 05
Beach at sunset
An orange sunset above blue water, with palms, crabs and a sea turtle.
Benchmark task
A busy shoreline at sunset tests scene composition, animal forms and the relationship between light and water reflections.
28 September 2026 · Owl Alpha / LongCat-2.0.
MODEL DETAILS · UPDATED 12 JULY 2026
A 1.6T-parameter mixture of experts.
LongCat-2.0 activates roughly 48 billion parameters per token on average and supports a native 1M-token context window. Meituan releases the weights under the MIT license on July 12.
Architecture details ↗July open-weights announcement ↗
- Total parameters
- 1.6T
- Average active / token
- ~48B
- Native context
- 1M
- Weights · 12 July
- MIT
ACCESS UPDATE · 28 SEPTEMBER 2026
LongCat API and open weights.
The named LongCat-2.0 listing provides current API access. Artificial Analysis lists $0.30 per million input tokens and $1.20 per million output tokens as of September 28.
Advertisement Demodokos Foundry
One voice.
Five performances.
Give the Owl a direction and it performs. One Foundry voice, five moods, every take made on a home GPU. No credits, no queue, nothing uploaded.
- 40+ emotions
- 10 speech languages
- Speech from 4 GB VRAM
Full access. Cancel anytime.
The listening room
Waiting for direction
Direction(a whisper, close to the ear)
Shh. Keep your voice down. The other models are listening.
Your turn in the director's chair. Direct your own voice
One Foundry voice. Five directions.
Audio could not load. Hear more at Demodokos ↗
FAQ
Owl Alpha questions.
What is Owl Alpha?
An anonymous coding model on OpenRouter with native tool calling and a 1,048,576-token context window.
Who is behind Owl Alpha?
OpenRouter doesn’t name its developer.
UPDATE · 30 APRIL 2026Community suggestions include GLM 5.2 and LongCat-2.0-Preview.
UPDATE · 29 JUNE 2026Meituan confirms that Owl Alpha belongs to its LongCat-2.0 family.
Is Owl Alpha free?
Yes. The stealth endpoint has no input or output token charge.
ACCESS UPDATE · 28 SEPTEMBER 2026For current access and rates, use the LongCat-2.0 listing.
Which version of Owl Alpha do these benchmarks test?
The full June release, measured by Artificial Analysis under its revealed name, LongCat-2.0.