Self-hosted Claude with memory that sticks, files that just work, and a calm interface that gets out of your way. Never start a thread over.
Aurelius runs on Claude Opus 4.8, the latest frontier model Anthropic shipped. It sets a new high on the GDPVal-AA Elo leaderboard for real-world knowledge work (1890), posts 69.2% on SWE-Bench Pro for autonomous coding, and clears 83.4% on agentic computer use. It also drops measurably on misaligned-behavior eval versus Opus 4.7. You get the best model running, with none of the chrome.
Start a conversation →| Benchmark | Opus 4.8 | Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Agentic coding SWE-Bench Pro | 69.2% | 64.3% | 58.6% | 54.2% |
| Agentic terminal coding Terminal-Bench 2.1 | 74.6% | 66.1% | 78.2% | 70.3% |
| Multidisciplinary reasoning Humanity's Last Exam, no tools | 49.8% | 46.9% | 41.4% | 44.4% |
| Multidisciplinary reasoning Humanity's Last Exam, with tools | 57.9% | 54.7% | 52.2% | 51.4% |
| Agentic computer use OSWorld-Verified | 83.4% | 82.8% | 78.7% | 76.2% |
| Agentic financial analysis Finance Agent v2 | 53.9% | 51.5% | 51.8% | 43.0% |
Opus 4.8 drops measurably on Anthropic's misaligned-behavior eval versus Opus 4.7 (1.83 vs 2.48, a 26% reduction) while gaining capability everywhere else. Higher capability without paying for it in safety is the harder of the two trade-offs, and it's the one Anthropic shipped. The same model is what answers your messages here.
Opus 4.8 keeps the same per-token price as Opus 4.7 but doubles down on long-running work: a 1M token context window by default, 128k max output, and a new "high" effort default that, per Anthropic, "spends a similar number of tokens as Opus 4.7's default, but with better performance." Long-horizon agentic coding gets better long-context handling and fewer compactions. Everything you ran on 4.7 keeps working, on the same bill, with a stronger model behind it.
Try it on a hard problem →Aurelius is built for serious thinking work, homework that needs explaining, code that needs writing, papers that need editing. Switch workspaces from the sidebar and the personality switches with you.
Walk through a derivation step by step. Get the concept behind a textbook chapter. Quiz yourself before the exam. Aurelius remembers what you already know and skips the parts you don't need.
"Explain mitosis like I'm preparing for IB Biology paper 2."
Drop a stack trace, get a fix. Paste a file, get a code-review. Code workspace gives you a side editor on every snippet, tweak it, copy it, download it.
"Find the off-by-one in this loop and explain why it bit me."
Research mode cites sources by domain name and is explicit about what it doesn't know. Web search is a toggle, not a default, you choose when the model goes online.
"Compare the lithium-ion vs sodium-ion grid storage literature from 2024."
Writer mode edits in your voice instead of imposing its own. Paste a draft, get the tightened version plus the diff explained. Headlines, openings, transitions, without sounding like every other AI output.
"Rewrite this email in plain language without losing the ask."
Tutor mode diagnoses what the learner already understands before teaching, then goes one step at a time. Lean toward Socratic if that's what you've picked in your profile.
"Walk me through Cauchy's integral formula, assume I know complex numbers."
General mode is your default, a thinking partner for whatever's on your mind. The thread persists, the context stays, you can come back tomorrow and continue.
"Help me think through how to tell my parents I'm switching majors."
Claude.ai gives you Opus 4.8 the way Anthropic ships it. Aurelius gives you Opus 4.8 the way you want to use it, with your data in your database, smart routing on every message, and zero noise around the chat.
Every reply runs on Opus 4.8 by default, the same frontier model topping Anthropic's leaderboards. We don't quietly downgrade you to keep margins, you pay for tokens, you get the model you paid for.
"Yo waguan fam" doesn't need Opus. Aurelius silently routes trivial messages to Haiku, ~50× cheaper, ~4× faster, and unlocks the full Opus envelope only when the question warrants it. You see the routing decision; the savings are yours.
Every message persists to your own Turso database. Three tables, users, chats, messages, that you can inspect, export to JSON, or wipe at any time. We can't read your data without your token, and we don't sell it ever.
Five onboarding questions and a settings page that lets you tell Aurelius who you are, how you want to be spoken to, what to never do. Every reply gets calibrated to your level and audience, not the average user's.
Zero behavioral tracking. Zero advertising. Zero marketing pixels. The only third-party calls we make are to Anthropic (the model) and your DB (your data). The privacy page is plain English, read it.
One sidebar, one composer, one model picker. No chips clamoring for your attention. Markdown, code blocks with copy buttons, file uploads, pinned chats, and a session that auto-saves to your sidebar, that's the whole surface.
Buy a 50,000-token pack for the model you want. No subscription, no surprise renewal, top up when you're close to empty.
Fast, cheap. Quick Q&A and short edits.
Balanced reasoning. The default pick.
Top-tier reasoning. Newest, hardest problems.
Packs are one-time purchases. Credits roll forward until used. Need volume? Talk to us.
Run Claude on your own infrastructure, with the polish of a product team behind it.