Astraeus 3 is the model path for broader reasoning across engineering work. It is where review, triage, product surfaces, and repo memory need to feel like one system.
The public announcement keeps the lower-level implementation details private. Astraeus is shown as the premium model path, with the memory layer measured as a platform effect around the benchmark frame.
Astraeus benchmark frame
Astraeus 3 is shown as the premium Sigilix route, with the displayed Astraeus task rows carrying the 3.6-point Sigilix context lift. Peer rows use the public benchmark references while lower-level implementation details stay out of the page.
These charts separate the questions that matter: command-line workflow performance, repository repair performance, maximum output room, and native context capacity. Missing public rows are omitted instead of being inferred.
TerminalBench is the closest public frame for terminal-bound coding work: reading the repository, choosing commands, reacting to execution feedback, and making progress through a task where each step depends on the one before it.
Astraeus is built for the longer version of that workflow. It is the route you use when the answer needs to connect repo structure, prior reviews, issue history, and current evidence before proposing the next move.
The Astraeus row applies the 3.6-point Sigilix context lift to the reference TerminalBench row. Separate ultra or agentic mode rows are omitted so the comparison stays focused on base model rows that fit this route.
Values shown: GPT-5.6 Sol, Astraeus 3 with context lift, GPT-5.5, Premium peer A, Premium peer B, Qwen3.7-Max, and NVIDIA Nemotron 3 Ultra.SWE-bench Verified shifts the comparison from terminal flow to repository repair. The task is to locate the relevant code, understand the failure, make the patch, and preserve the behavior the tests are meant to protect.
Astraeus is designed for that kind of broad repair context. Its benchmark row is useful, but the product route matters because your repository memory and prior decisions help keep the repair grounded in how your codebase actually works.
The Astraeus row applies the same 3.6-point context lift to the reference SWE-bench Verified value. The chart only includes peers with a public SWE-bench Verified value; GPT-5.6 Sol and Terra are omitted here because the reference set did not include public SWE-bench Verified rows for them.
Values shown use public SWE-bench Verified rows for Premium peer A, Astraeus 3 with context lift, GPT-5.5, Qwen3.7-Max, Premium peer B, and NVIDIA Nemotron 3 Ultra.This chart shows Astraeus as an effort curve instead of a single detached benchmark point. The rightmost Astraeus point carries the 3.6-point context lift on top of the reference SWE-bench row; earlier points show the working shape for shorter and medium-length repair work.
Astraeus also applies the tested 63% repeated-context reduction at each effort point. The raw output budget still exists, but the chart plots the effective load after memory carries forward the context that would otherwise be repeated.
Peer lines are capacity checkpoints anchored to their public SWE-bench Verified rows. They give visual context for output room and score without pretending every vendor publishes the same per-effort benchmark series.
Hover a dot to see the effort label, effective output tokens, raw output budget when applicable, context window, and score used for that checkpoint.
Scores are anchored to public SWE-bench Verified rows for Astraeus 3, Premium peer A, GPT-5.5, Qwen3.7-Max, and Premium peer B. Astraeus effective tokens use the tested 63% repeated-context reduction.Output capacity is not the same as intelligence. It is the room the route can use when the answer needs to carry planning, code edits, evidence, and follow-up reasoning without being cut short.
For Astraeus, that room matters because premium work often spans a larger surface: multiple files, earlier PR decisions, issue constraints, and a repair plan that has to remain coherent across the whole change.
The chart uses only rows with public max-output limits. Models without a public output-token cap in the reference set are omitted rather than estimated.
Values shown use max-output rows for Premium peer B, Astraeus 3, Premium peer A, GPT-5.5, and Qwen3.7-Max.Context capacity is the surface a route can hold in the current run. Memory is different: it is what Sigilix can carry forward so the same background does not have to be reconstructed every time.
Astraeus sits at the intersection of both. The native window gives the model room to read the current work; memory gives the product continuity across PRs, Slack decisions, Linear issues, triage runs, and CLI sessions.
The chart uses the public context-window rows and leaves out models where the reference set did not provide a public context value.
Values shown use context-window rows for GPT-5.5, Astraeus 3, Premium peer A, NVIDIA Nemotron 3 Ultra, Qwen3.7-Max, and Premium peer B.What Astraeus is for.
Astraeus is for repository-scale questions, longer review chains, and work that needs to connect the current diff to prior decisions and adjacent systems.
It is less about answering one prompt and more about maintaining continuity across the engineering path.
Benchmark framing.
The comparison keeps public benchmark rows separate from product architecture. The page shows Astraeus as the route you use, then compares it against base and frontier peers where TerminalBench, SWE-bench, output, and context rows are available.
Astraeus is not presented as a detached raw-model claim. Its value is the route around the model: repository memory, issue history, review evidence, and prior decisions all give the system a better starting point for long-horizon engineering work.
Each chart answers a different question. TerminalBench is about interactive command work, SWE-bench is about repository repair, output capacity is about how much room a route can use when it has to produce a long answer, and context capacity is about how much current surface can fit before memory has to carry more of the continuity.
Cross-surface memory.
Astraeus benefits most when PRs, Slack decisions, Linear issue trails, triage runs, and CLI sessions are allowed to reinforce each other instead of being pasted into a fresh prompt every time.
What comes next
Astraeus is the model path for work that has to remember how the repository got here.