Skip to main content
Model release · v0.47

Janus 1.0 Pro

The larger Janus model, for long, multi-step work: software, research and agents that run for hours.

Introducing Janus 1.0 Pro

Janus 1.0 Pro is the larger model in the Janus family, for work that runs long: software built across a whole repository, research that reads many sources, and agents that stay on a task for hours. It's available on every paid Kaer plan and on the Kaer API.

It joins Janus 1.0 Flash, the default model on every plan. Flash is priced for agents that take many quick steps; Pro reasons through every step before it answers, at $0.14 / $0.70 per million input / output tokens against Flash's $0.14 / $0.35. In our testing, Janus 1.0 Pro stood out on:

Agentic coding. 76.8% on SWE-bench Verified and 58.3% on SWE-bench Pro, the harder set of real repository tasks, where Janus 1.0 Flash scores 38.4%.

Agents on the web and the desktop. 69.6% on BrowseComp, which asks for facts buried deep in the web, and 67.0% on OSWorld-Verified, which runs tasks in real desktop applications.

Reasoning. 58.7% on ARC-AGI-2 and 81.1% on GPQA Diamond, graduate-level science questions written to resist search.

Price. $0.14 per million input tokens, $0.03 per million cached input tokens and $0.70 per million output tokens. Janus 1.0 Flash is $0.14, $0.03 and $0.35.

Context. A one-million-token context window, the same as Flash, for whole codebases and long research in one request.

Benchmarks

We report Janus 1.0 Pro on 32 benchmarks across software engineering, agentic work, professional tasks, reasoning and knowledge, long context and security. The charts and the tables compare it with the same models throughout: Janus 1.0 Flash, Gemini 3.8 Flash, GPT-5.6 Luna and GLM-5.3-Flash. The charts cover six headline benchmarks. Of the 11 benchmarks both Janus models report, Janus 1.0 Pro scores higher on five, among them ARC-AGI-2, BrowseComp and SWE-bench Pro. Janus 1.0 Flash scores higher on DeepSWE v1.1, ProgramBench, Terminal-Bench 4.0, Automation-Bench v1.0.6, HLE (No Tools) and MMLU-Pro.

  • Janus 1.0 Pro
  • Janus 1.0 Flash
  • Other models

Agentic coding

SWE-bench Verified

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by Vals AI
  4. GPT-5.6 Luna$0.20 / $1.20As published by Vals AI
  5. GLM-5.3-Flash$0.15 / $0.50As published by Vals AI

Agentic terminal coding

Terminal-Bench 4.0

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by Vals AI
  4. GPT-5.6 Luna$0.20 / $1.20As published by Vals AI
  5. GLM-5.3-Flash$0.15 / $0.50As published by Vals AI

Graduate-level science

GPQA Diamond

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by Artificial Analysis
  4. GPT-5.6 Luna$0.20 / $1.20As published by Artificial Analysis
  5. GLM-5.3-Flash$0.15 / $0.50As published by Artificial Analysis

Expert reasoning, no tools

Humanity's Last Exam

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by Artificial Analysis
  4. GPT-5.6 Luna$0.20 / $1.20As published by Artificial Analysis
  5. GLM-5.3-Flash$0.15 / $0.50As published by Artificial Analysis

Knowledge and reasoning

MMLU-Pro

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by Vals AI
  4. GPT-5.6 Luna$0.20 / $1.20As published by Vals AI
  5. GLM-5.3-Flash$0.15 / $0.50As published by Vals AI

Abstract reasoning

ARC-AGI-2

  1. Janus 1.0 Pro$0.14 / $0.70Kaer's result
  2. Janus 1.0 Flash$0.14 / $0.35Kaer's result
  3. Gemini 3.8 Flash$0.75 / $3.75As published by ARC Prize
  4. GPT-5.6 Luna$0.20 / $1.20As published by ARC Prize
  5. GLM-5.3-Flash$0.15 / $0.50As published by ARC Prize

Janus 1.0 Pro and Janus 1.0 Flash scores are Kaer's. Every panel compares the same three models, the low-cost models from Google, OpenAI and Z.ai with an independent result on all six benchmarks; each benchmark's scores for them come from one evaluator, listed below, whose harness and settings can differ from Kaer's. Prices are list prices per million input / output tokens, as of 30 October 2026.

Sources for the other models: Vals AI, Artificial Analysis, ARC Prize.

The full results, by area, against the same models. Janus 1.0 Flash's results are from its release; each benchmark's scores for the other models come from one independent evaluator.

Software engineering: scores in percent unless marked Elo
Software engineeringJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Agentic codingSWE-bench Verified76.8%74.8%80.0%93.0%92.0%
Agentic codingSWE-bench Pro58.3%38.4%———
Agentic codingDeepSWE v1.159.9%71.2%———
Whole-program generationProgramBench16.5%21.5%1.0%0.0%0.0%
Olympiad programmingIOI V224.3%—56.9%61.8%52.5%
Agentic / terminal: scores in percent unless marked Elo
Agentic / terminalJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Agentic terminal codingTerminal-Bench 2.173.9%—87.6%80.9%84.3%
Agentic terminal codingTerminal-Bench 4.022.9%33.5%19.7%11.6%37.4%
Agentic web searchBrowseComp69.6%41.6%———
Business workflowsAutomation-Bench v1.0.642.1%52.0%———
Business workflowsAutomationBench-AA42.6%—59.9%50.2%60.4%
Computer useOSWorld-Verified67.0%————
Tool use across appsToolathlon-Verified62.9%————
Long-horizon agent tasksAgents' Last Exam21.6%————
Professional work: scores in percent unless marked Elo
Professional workJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Professional tasksGDPval 2.11422 Elo—1445 Elo1468 Elo1649 Elo
Professional tasksAA-Briefcase v1.11284 Elo—1203 Elo1339 Elo1451 Elo
Data analysisAA-AnalystAgent25.5%—46.3%—25.0%
Professional tasksGDP.pdf9.2%—21.0%24.0%15.4%
Professional tasksGDPval-AA v2.1 (normalised)48.2%————
Professional tasksJobBench49.0%————
Reasoning / knowledge: scores in percent unless marked Elo
Reasoning / knowledgeJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Graduate-level scienceGPQA Diamond81.1%72.1%95.3%91.1%91.2%
Multidisciplinary reasoningHLE (No Tools)35.4%37.8%47.8%39.5%39.9%
Scientific codingSciCode48.9%—56.6%53.6%51.6%
Research-level physicsCritPt16.6%—18.3%20.6%15.4%
Knowledge and reasoningMMLU-Pro73.3%83.6%90.2%86.0%86.1%
Abstract reasoningARC-AGI-180.0%—98.5%88.0%91.0%
Abstract reasoningARC-AGI-258.7%9.7%89.2%59.5%65.8%
Mathematical proofsProofBench 1.152.0%————
Long context: scores in percent unless marked Elo
Long contextJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Long-context reasoningAA-LCR v1.173.3%—81.3%83.7%80.0%
Security: scores in percent unless marked Elo
SecurityJanus 1.0 Pro$0.14 / $0.70Janus 1.0 Flash$0.14 / $0.35Gemini 3.8 Flash$0.75 / $3.75GPT-5.6 Luna$0.20 / $1.20GLM-5.3-Flash$0.15 / $0.50
Vulnerability reproductionCyberGym76.0%————
CybersecurityExploitGym7.8%————
CybersecurityExploitBench30.9%————
CybersecuritySEC Bench Pro51.3%————

Janus 1.0 Pro and Janus 1.0 Flash scores are Kaer's. Each benchmark's scores for the other models come from one evaluator, listed below, whose harness and settings can differ from Kaer's. A dash means no published result on that benchmark version. GDPval 2.1 and AA-Briefcase are Elo ratings, not percentages. Prices are list prices per million input / output tokens, as of 30 October 2026.

Sources for the other models: Kaer, Vals AI, Artificial Analysis, ARC Prize, Mercor.

Pro or Flash

Janus 1.0 Flash stays the default: on Auto, Kaer uses it for everyday work and turns its thinking on for coding, analysis and research. Choose Janus 1.0 Pro in the model menu when the task is long and the steps depend on each other, such as a change that touches many files, a report that weighs dozens of sources, or an agent left to run on its own.

Janus 1.0 Pro reasons through every step before it answers, so it takes longer and uses more tokens than Flash on the same request. In Kaer, both models draw on your plan's credits. Read more about Flash in the Janus 1.0 Flash release.

Getting started

Janus 1.0 Pro is available today on every paid Kaer plan, from $9.99 a month. Developers can use it on the Kaer API with the model ID janus-1-pro, and Janus 1.0 Flash with janus-1-flash.

Price per 1M tokens
Price per 1M tokensJanus 1.0 ProJanus 1.0 Flash
Input$0.14$0.14
Cached input$0.03$0.03
Output$0.70$0.35

In the Kaer app, usage draws on your plan's credits. On the Kaer API, these are the prices per million tokens.

Frequently asked questions

What is Janus 1.0 Pro?

Janus 1.0 Pro is the larger model in the Janus family, for long, multi-step work such as building software, research across many sources and agents that run for hours. It's available on every paid Kaer plan.

How much does Janus 1.0 Pro cost?

$0.14 per million input tokens, $0.03 per million cached input tokens and $0.70 per million output tokens. In the Kaer app, usage draws on your plan's credits.

How is it different from Janus 1.0 Flash?

Janus 1.0 Pro reasons through every step before it answers and scores higher than Janus 1.0 Flash on benchmarks such as ARC-AGI-2, BrowseComp and SWE-bench Pro. Janus 1.0 Flash is faster, costs less per output token and remains the default on Auto.

Can I use Janus 1.0 Pro through an API?

Yes. It's on the Kaer API with the model ID janus-1-pro, at the same per-token prices as in the app.

How large is the context window?

One million tokens, the same as Janus 1.0 Flash.

Janus 1.0 Pro is available in Kaer today.

On every paid plan, and on the Kaer API.