Aurvane · Model launch · 01

Halcyre 1 thinks in light.

Our first frontier model reasons in long, steady steps, shows its working and knows when to stop. In the API from launch day.

The launch filmA lattice of glass over a desert pool

The launch

A model you can watch think.

Most models hand you an answer. Halcyre 1 hands you the reasoning too: the plan it made, the checks it ran and the sources it leaned on, so your team can trust the result or catch it early.

1Mtokens of context, read in one pass
3.1×faster to first token than our previous model
−42%cost per solved task at the same quality

Illustrative figures for a fictional model.

Inside Halcyre 1

  1. 01 / 04 · Reads

    Takes in everything at once.

    A million tokens in one pass: the whole codebase, the full contract, a quarter of email. It finds the line that matters and says where it found it.

  2. 02 / 04 · Plans

    Thinks before it speaks.

    Hard questions get a plan first. It breaks the problem into steps and spends more thought where the problem is harder, less where it isn't.

  3. 03 / 04 · Checks

    Shows its working.

    Every step it takes is there to read. It tests its own code, checks its sums and cites its sources, so a reviewer can follow the path, not just the answer.

  4. 04 / 04 · Knows

    Says when it isn't sure.

    Calibrated to admit doubt. When the evidence runs out, it tells you what it would need to answer properly instead of guessing.

The numbers

A step change, not a nudge.

Halcyre 1 against our previous model on four in-house test suites, and what happens when you give it more time to think.

IllustrativeAurvane, Halcyre 1 and these test suites are fictional. These are not real benchmark results.

Score by test suiteHalcyre 1Previous model
Lattice Reasoningmulti-step logic, 2,400 problems
86.471.2
LongRead 1Manswers buried in a million tokens
91.864.5
Patchwork Codereal repository bug fixes
72.955.1
Grounded Answersclaims checked against sources
94.183.6

050100

More thinking, better answersScore on Lattice Reasoning by thinking budget (tokens)
Illustrative: score against thinking budgetHalcyre 1 climbs from 58 at a 1k-token budget to 89.5 at 64k; the previous model levels off near 64. 4060801001k2k4k8k16k32k64k Halcyre 1 · 89.5Previous model · 64.5

Try it

Ask it something hard.

Pick a prompt and watch Halcyre 1 think, then answer. This playground is a scripted replay that runs in your browser: no model is called and nothing you type leaves the page.

halcyre-1Ready

The answer streams in here, word by word.

Tokens 0First token –Speed –Scripted replay

Model card

Halcyre 1, on one page.

What it takes in, what it gives back, and where it shouldn't be trusted on its own.

# one request, reasoning shown
resp = client.responses.create(
  model="halcyre-1",
  input="Review this pull request",
  reasoning={"effort": "high", "show": True},
)
Model ID
halcyre-1 · halcyre-1-swift
Context window
1,000,000 tokens
Longest answer
128,000 tokens
Takes in
Text, code, images, PDFs, tables
Gives back
Text, code, structured JSON, tool calls
Reasoning
Adjustable effort, visible steps
Knowledge
Trained on data up to mid-2026
Where it runs
Aurvane API, or a private deployment

Use with care

  • Not for medical, legal or financial decisions without a qualified person checking the result
  • Can still be confidently wrong; read the steps on anything that matters
  • Your API data is not used to train our models

Build with Halcyre 1

Start with one question.

Tell us what you're building. We open access in waves and reply to every request.

Request API access