A new "stealth" LLM was launched offering 100T tokens per day

laurentiurad2 pts0 comments

Ox Alpha

Open Chat

Now in stealth · free to try

Meet Ox Alpha.

The reasoning model that appeared out of nowhere. Built for code, long-horizon agents, and a million tokens of context. Nobody knows who made it — everyone wants to try it.

Start chatting<br>Learn more

No sign-up. No subscription. Chats aren't stored on our servers.

1M<br>tokens of context

131K<br>max output tokens

input modalities<br>text · image · video

$0<br>while it's in stealth

Reasoning

Thinks before<br>it speaks.

Ox Alpha is a reasoning-first model. It plans, checks itself, and only then answers — so hard problems get worked through instead of guessed at.

Multi-step logic Math, systems design, proofs, tricky edge cases — worked through, not pattern-matched.

Visible thinking Watch the reasoning stream in before the final answer lands.

Production-grade output Tuned for real workloads, not party tricks.

1,048,576

A million tokens of memory.

Paste an entire codebase, a 500-page spec, or a week of transcripts. Ox Alpha keeps all of it in mind while it works — no chunking, no retrieval tricks, no losing the thread.

Capabilities

Built for real work.

Everything a frontier model should do — and a few things most of them can't.

Long-horizon coding

Designed for sustained software engineering: debugging, refactoring, and reasoning across large codebases without losing context.

Sees what you see<br>Text, images, and video in. Screenshots, diagrams, and UI — reasoned about, not just described.

Tool calling

Native tools and tool_choice support for agents that actually do things.

Structured output

Ask for JSON and get JSON — schema-shaped responses via response_format.

Sustained agentic work

Keeps its footing across long, multi-step tasks instead of drifting off halfway.

In action

Watch it reason.

A reasoning model shows its work. Here's what a typical exchange looks like.

oxalpha.com/chat

Why does this Python function sometimes return stale data?<br>cache = {}; def get(k): return cache.get(k) or fetch(k)

Ox Alpha

Thinking Two issues: or treats falsy cached values (0, "", []) as misses and re-fetches; and nothing ever writes to cache, so the "stale" data must come from fetch itself — likely a module-level connection reused across requests. Check both.

The bug is in the or : any falsy cached value is treated as a miss. Also, cache is never populated. Use an explicit membership check and store results:<br>def get(k):<br>if k not in cache: # explicit miss check<br>cache[k] = fetch(k) # actually populate<br>return cache[k]<br>If results still look stale, the culprit is fetch reusing a connection with read-consistency lag — want me to look at it?

Benchmarks

Stealth model. Top of the table.

On an independent run of 10 real-world coding tasks, Ox Alpha solved 8 — ahead of every frontier model it was lined up against.

80%<br>Ox Alpha<br>8 of 10 tasks solved

65%<br>Fable 5<br>fable-5 [max]

62%<br>GLM-5<br>glm-5.3 [max]

52%<br>GPT-5<br>gpt-5.6-sol [max]

62%<br>Grok 4<br>grok-4.6 [xhigh]

Task<br>fable-5 [max]glm-5.3 [max]gpt-5.6-sol [max]grok-4.6 [xhigh] ox-alpha

anko-typed-variable-bindings<br>4/44/42/41/4 ✓

arktype-json-schema-refs<br>2/41/43/41/4 ✓

fastapi-deprecation-headers<br>4/43/43/44/4 ✓

helm-unified-manifest-stream<br>4/44/44/44/4 ✓

igel-persist-feature-schema<br>3/43/40/44/4 ✓

katex-multicolumn-array-spans<br>2/44/43/44/4 ✓

meriyah-explicit-resource-decl<br>1/40/40/40/4 ✓

query-persist-restored-state<br>2/43/41/42/4 ✓

scc-bounded-memory-spilling<br>4/43/44/44/4 ✕

vulture-persistent-analysis-cache<br>0/40/41/41/4 ✕

Mean on these 10<br>65%62%52%62% 80%

Independent community benchmark — 10 real-world coding tasks. Reference models: passes out of 4 attempts per task. Ox Alpha: pass/fail. Highlighted row: the task every reference model scored 1/4 or worse on — Ox Alpha solved it. Third-party data, small sample — directional, not definitive.

Full breakdown & methodology

Ox Alpha vs Fable 5

Ox Alpha vs GLM-5

Ox Alpha vs GPT-5

Ox Alpha vs Grok 4

Try it on

Pick a task. Or bring your own.

Coding

Debug a gnarly production traceback

Root-cause analysis from a model built for long-horizon software engineering.

Open in chat

Reasoning

Work through a hard math or logic problem

It plans before it answers — watch it think in steps, not guesses.

Open in chat

Big context

Analyze an entire codebase in one prompt

A 1M-token window means the whole repo fits. No chunking, no RAG.

Open in chat

Vision

Explain a screenshot, chart, or diagram

Combine text with visual context and let it reason about what it sees.

Open in chat

Agentic

Design a multi-step agent workflow

Tool calling, structured output, and the stamina for long tasks.

Open in chat

Coding

Review a pull request like a senior engineer

Correctness, edge cases, simplifications — with the reasoning behind them.

Open in chat

Writing

Turn rough notes into a production-ready doc

Structured, precise writing tuned for production workloads.

Open in chat

Big context

Digest a 500-page PDF into key takeaways

Executive summary, quotes, and...

alpha chat open reasoning model cache

Related Articles