Introducing Janus 1.0 Flash
Janus 1.0 Flash is the first model in the Janus family and, from today, the default in Kaer on every plan. We curated it for one job above all: agents that write and run code, step after step, at a price that makes leaving them running sensible.
Frontier models have become very good at agentic coding, and very expensive to leave running. An agent working through a codebase re-reads its context on every step, so a long session can cost more than the change is worth. We measured the models available to us against that workload, chose the one that gives the most capability per dollar, and put it under one name, inside one subscription, with the rest of Kaer around it. Janus 1.0 Pro, a larger model for long, multi-step work, joins it on all paid plans.
In our testing, Janus 1.0 Flash stood out on:
Agentic coding. An agent sends its context again on every step, so Janus 1.0 Flash prices cached input at $0.03 per million tokens. A 50-step coding session over a 100,000-token context comes to about $0.24 at list prices.
One subscription. In Kaer the model comes with the plan. The operator, its own computer and browser, workflows, AI assistants and publishing all draw on the same credits, with paid plans from $7.99 a month.
Frontier reasoning. With thinking on, Janus 1.0 Flash scores 91.7% on AIME 2026, 6.6 points behind Gemini 3.1 Pro, which costs 14 times as much per token.
Price. Janus 1.0 Flash is priced at $0.25 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. Every model that outscored it in our comparison costs at least 1.6 times as much per token.
Reliability. None of the 1,612 requests in our benchmark runs ended in an error or a timeout.
Context. Its one-million-token context window holds long contracts, large codebases and complete email threads in a single request.
Priced for agentic coding
Agentic coding is where token prices compound. On each step the agent sends its working context again: the files it has open, its plan, the output of the commands it ran. Most of those tokens are repeats, so the price that decides the bill is cached input. Janus 1.0 Flash charges $0.03 per million cached tokens, against $0.25 for fresh input.
Take a session of 50 steps over a 100,000-token context, where each step adds 3,000 new tokens of tool output and writes 2,000. At list prices a step costs about 0.5¢ and the whole session about $0.24. Without the cache discount, the same session would cost $1.34. These sizes are an illustration; yours will differ, and the arithmetic is the same.
It writes whole programs, too. Asked once for each of six small web apps, Janus 1.0 Flash wrote 3,321 lines of code for $0.11 in total. The apps run on this page, exactly as it wrote them, with the flaws we found noted under each.
One subscription
Coding with agents usually means paying for a model, then again for everything around it: somewhere to run the code, a browser for the web, a place to deploy and something to run jobs on a schedule. In Kaer, Janus 1.0 Flash comes with the plan. The operator, its own computer and browser, workflows and automations, AI assistants and publishing what you build all draw on the same credits.
Janus 1.0 Flash is the default on every plan, including the free Starter plan. Paid plans start at $7.99 a month and add Janus 1.0 Pro. On the Kaer API, both models are billed per token at the prices at the end of this post.
Performance
We compared Janus 1.0 Flash with six frontier models on two widely reported evaluations: GPQA Diamond (graduate-level science) and MMLU-Pro (knowledge and reasoning). Every result for another model comes from an independent evaluator rather than the company that built it. Benchmark scores capture only part of a model's usefulness for everyday work, which is why we also show what Janus 1.0 Flash builds and writes, unedited, further down.
| Janus 1.0 Flash | Claude Fable 5.1 | GPT-5.6 Sol | Gemini 3.1 Pro | Gemini 3.8 Flash | DeepSeek V4 Pro | Kimi K3 | |
|---|---|---|---|---|---|---|---|
| Graduate-level scienceGPQA Diamond | 72.1%3 | 93.7% | 94.1% | 94.1% | 95.3% | 92.8% | 93.5% |
| Knowledge and reasoningMMLU-Pro | 83.6%4 | 92.4% | 89.1% | 91.0% | 90.2% | 87.0% | 88.0% |
| Price per 1M tokensInput / output | $0.25 / $0.50 | $10 / $50 | $4 / $20 | $2 / $12 | $0.75 / $3.75 | $1.32 / $3.96 | $3 / $15 |
Janus 1.0 Flash results are from our own runs with thinking on. Other results are as published by Artificial Analysis (GPQA Diamond) and Vals AI (MMLU-Pro), accessed on 6 October 2026.
Bold marks the best result in each row; for price, the lowest. Numbered marks refer to the footnotes at the end of this post.
On GPQA Diamond and MMLU-Pro, the leading models are ahead by 24.0 and 8.8 points respectively. Every model that scores higher than Janus 1.0 Flash costs at least 1.6 times as much per token.
The chart below plots each model's score against its price per million tokens. The closer a point is to the top left, the more capability it delivers per dollar. On MMLU-Pro, the closest model above Janus 1.0 Flash, MiniMax-M3, scores 0.6 points higher at 1.6 times the price.
GPQA Diamond measures 198 multiple-choice questions in biology, physics and chemistry, written by PhD-level experts to be hard to answer even with web search. Janus 1.0 Flash was run by Kaer with thinking on; other results are as published by Artificial Analysis. Prices are each vendor's list price per million tokens, weighted three parts input to one part output.
Building apps
Benchmarks measure narrow skills, so we also asked Janus 1.0 Flash for something more ordinary: six small web apps of the kind people ask Kaer to build, each from a single prompt. Every app below is the model's first and only attempt, running on this page exactly as it wrote it. Switch to Code to read the file.
I run a small plumbing business. Build me an invoice generator I can use in the browser: my business details and the customer's at the top, an invoice number and date, line items (description, quantity, unit price) that I can add and remove, a VAT rate I can set to 0%, 5% or 20%, and the subtotal, VAT and total updating as I type. Add a Print button that prints only the invoice, laid out like a proper A4 invoice. One HTML file with inline CSS and JavaScript, no external libraries.
Generated on 6 October 2026 with thinking on, one request per app. Each app runs in a sandbox on this page, with no network access. At Janus 1.0 Flash's API rates, the six apps cost $0.11 to produce in total.
Janus 1.0 Flash in use
The examples below show Janus 1.0 Flash on four tasks of the kind people bring to Kaer. Each response is the model's first attempt, shown in full and unedited, with the cost of producing it at Janus 1.0 Flash's rates.
A client has not paid invoice #1042 for £2,400, now 14 days overdue. They have paid late before but always paid in the end. Write a short, polite but firm reminder email from me (Sam, owner of Northlight Studio). Under 150 words, with a clear subject line.
Of course. Here is a short, polite, and firm reminder email.
Subject: Friendly reminder: Invoice #1042 for £2,400
Hi [Client Name],
Hope you're having a productive week.
This is a quick reminder that invoice #1042, for £2,400, was due 14 days ago.
To avoid any delays, we’d appreciate it if you could arrange payment at your earliest convenience.
You can pay securely here: [Payment Link]
If you are having any issues with this, please let me know.
Best regards,
Sam Owner, Northlight Studio
Thinking
Thinking lets Janus reason through a problem before it responds. It takes longer and uses more tokens, but it improves results on reasoning-heavy tasks such as maths. In Kaer, Auto turns thinking on for coding, analysis and research requests and leaves it off for routine ones, so simple answers stay fast. On the Kaer API, you can turn it on for any request. All benchmark results in this post were measured with thinking on.
Reliability
We measured reliability in two ways: whether Janus 1.0 Flash completes a request, and whether it reaches the same answer when asked again.
None of the 1,612 requests in our benchmark runs ended in an error or a timeout. Eighty-three responses reached our output limit of 65,536 tokens and seven were declined; we scored all of these as incorrect. On AIME 2026, Janus 1.0 Flash solved 25 of the 30 problems on all four attempts, and on GPQA Diamond, it answered 97 of the 198 questions correctly on all four attempts.
When Janus can't answer a request, Kaer says so directly rather than returning an empty response. Usage in Kaer is prepaid, and Kaer pauses to ask before spending past your credits.
Janus 1.0 Pro
Janus 1.0 Pro reasons through every step before it answers, for long, multi-step work such as building software, research across many sources and agents that run for extended periods. It's available on all paid plans and costs a little more per token than Janus 1.0 Flash.
Getting started
Janus 1.0 Flash is available today in Kaer on every plan, where it's the default model on Auto. Janus 1.0 Pro is available on all paid plans. Developers can use both models on the Kaer API with the model IDs janus-1-flash and janus-1-pro.
| Price per 1M tokens | Janus 1.0 Flash | Janus 1.0 Pro |
|---|---|---|
| Input | $0.25 | $0.26 |
| Cached input | $0.03 | $0.01 |
| Output | $0.50 | $0.78 |
In the Kaer app, usage draws on your plan's credits. On the Kaer API, these are the prices per million tokens.
Footnotes
- Janus 1.0 Flash results are from our own runs between 5 and 7 October 2026, with thinking on, temperature 1.0, top-p 0.95, a limit of 65,536 output tokens and one question per request, using each benchmark's official prompt.
- GPQA Diamond uses the prompt and answer extraction of OpenAI's simple-evals. Each question was answered four times with its options in a different order, and the score is the average. Of the 792 attempts, six were declined and 179 gave no answer in the required “Answer: X” form; all count as incorrect. Of those without an answer, 78 reached the output limit while still reasoning, and most of the rest repeated the instruction's placeholder instead of a letter.
- Janus 1.0 Flash's MMLU-Pro result is on a 700-question subset, with an equal number of questions from each of its 14 subjects, drawn with a fixed seed and run with the official five-shot prompt. Vals AI reports results on the full set.
- Responses that errored, timed out, were declined or reached the output limit were scored as incorrect. No question was re-run.
- 95% confidence intervals for Janus 1.0 Flash: 67.1–76.9 (GPQA Diamond) and 80.6–86.1 (MMLU-Pro).
- Results for other models are as published by Artificial Analysis, Epoch AI and Vals AI, accessed on 6 October 2026, with prices from each vendor's pricing page on the same day. Evaluation setups differ, so small gaps between models may not be meaningful. Artificial Analysis notes that some Claude Fable 5.1 requests are answered by a fallback model.
Frequently asked questions
What is Janus 1.0 Flash?
Janus 1.0 Flash is the first model in the Janus family and the default model in Kaer on every plan, including the free Starter plan.
How much does Janus 1.0 Flash cost?
$0.25 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. In the Kaer app, usage draws on your plan's credits.
How is Janus 1.0 Pro different?
Janus 1.0 Pro reasons through every step before it answers, for long, multi-step work such as building software and research across many sources. It costs a little more per token and is available on paid plans.
Can I use Janus through an API?
Yes. Both models are available on the Kaer API with the model IDs janus-1-flash and janus-1-pro, at the same per-token prices as in the app.
Does Janus think before it answers?
In Kaer, Auto turns thinking on for coding, analysis and research requests. On the Kaer API, you can turn it on for any request. Thinking uses more tokens and improves results on reasoning-heavy tasks.
How large is the context window?
One million tokens, for both Janus 1.0 Flash and Janus 1.0 Pro.
How were the benchmarks run?
We ran every Janus result ourselves with thinking on, following each benchmark's official setup, and counted every failed or unanswered response as incorrect. Results for other models come from independent evaluators, linked and dated in the post.