How this is made, and who is accountable for it.
Cafiyn Pulse gives advice about security and money, so the methods behind it should be legible and the person behind it should be nameable. Both are on this page, along with every correction we have had to publish.
One person, one company, no sponsors.
Cafiyn Pulse is published by Cafiyn Innovations, a company based in Bangalore, India. The tools, the data and every word of the written content are by Karthik Kumar, the founder. There is no editorial team.
Nobody pays to be here. No provider named in the Cost Comparator, no product in a comparison page, and nothing in the weekly briefing is sponsored, and there are no affiliate links anywhere on the site. Where a page recommends one option over another, that is a judgement call and the reasoning is shown so you can disagree with it.
You can reach me at hello@cafiyn.com, X, LinkedIn, or Instagram. Corrections are welcome and get published rather than quietly applied.
How the Stack Grader scores.
The grader asks 10 questions across 10 categories: auth, secrets, backups, ratelimit, payments, monitoring, https, validation, access, deps. Every question carries a weight from 1 to 12, set by how much damage getting it wrong actually causes rather than by how hard it is to fix.
Each answer is a tier that earns a fraction of that weight, from 0 for not done at all through to 1.0 for a production-grade answer. Your score is earned weight divided by applicable weight. Questions you mark not applicable are removed from both the numerator and the denominator, so skipping payments because you do not take payments neither helps nor hurts your grade.
| Grade | Score | What it means |
|---|---|---|
| A | 90 and above | No ship blockers. Worth re-checking after your next big change. |
| B | 75 to 89 | Solid. One or two gaps worth closing before you promote it. |
| C | 60 to 74 | Shippable to a small audience, not to a paying one. |
| D | 40 to 59 | Real exposure. Fix the top three before launch. |
| F | Below 40 | Ship blockers present. Something here will bite you. |
How model prices are verified.
The roster covers 79 models from 22 providers. Prices come from each provider's own pricing page, never from a pricing aggregator or a summary blog, because those are wrong often enough that citing them would make the whole exercise worthless.
There are two passes:
- Weekly, automated. A GitHub Action checks every price we publish against the provider page we cite for it, and opens an issue when one no longer appears there. It deliberately does not edit anything: provider pages differ too much to parse safely, and a scraper that guesses wrong would publish a wrong price with the authority of an automated check.
- Monthly, by hand. A person re-reads every provider pricing page and reconciles the roster against it. This is the pass that catches a model moving to a legacy list, an introductory rate with an expiry date, or a per-request fee that per-token pricing hides.
Roster version 2026-09-05. Model pages verified 2026-09-05. Last full reconciliation 2026-09-07. Providers covered: Anthropic, AssemblyAI, Azure, BFL, Cohere, DeepSeek, Deepgram, ElevenLabs, Google, Groq, Ideogram, Kling, Luma, Mistral, OpenAI, Perplexity, Pika, Recraft, Runway, Stability, Together, xAI.
What happens when we get it wrong.
Corrections get published, dated, with the wrong figure and the right one side by side. We do not quietly edit a number and carry on, because anyone who made a budget decision on the old figure deserves to see that it moved and by how much.
There are currently 3 published corrections, all of them cases where a price we published was wrong:
- o3-mini, o4-mini (our listing) (2026-09-07): We listed both as current with no caveat. They have had a published shutdown date since 22 April 2026. Our own weekly drift check flagged both on 5 September and we recorded it as a likely false positive.
- GPT-5 (2026-09-05): Our roster carried GPT-5 at $5 input and $15 output. OpenAI lists $1.25 and $10. Corrected.
- Claude Sonnet 5 (2026-09-05): Our roster carried Sonnet 5 at $3 input and $15 output. Anthropic lists $2 and $10. Corrected.
The full record, including provider-side changes and increases that are already scheduled, is in the price change log.
Rules the weekly briefing follows.
- Every item traces to a primary source: provider documentation, an official announcement, or established trade press. The link is always attached.
- Pricing aggregators and SEO summary blogs are not sources. They are wrong often enough to be worse than nothing.
- Every item carries a line on what a team shipping a product should do differently as a result. If there is nothing to do, the item does not run.
- Hot takes are opinion and are labelled as opinion, kept separate from the sourced reporting.
- We never state that a named company failed, shut down or was breached unless the linked source says exactly that.
- Published figures are not rewritten after the fact. Past issues keep permanent URLs so a number you read stays readable.
Common questions.
Who publishes Cafiyn Pulse?
Cafiyn Innovations, a company based in Bangalore, India. The tools and all written content are by Karthik Kumar, the founder. There is no editorial team and no sponsored content: nobody pays to appear in the Cost Comparator, the comparison pages, or the weekly briefing.
How does the Stack Grader calculate a grade?
Every question carries a weight from 1 to 12 reflecting how much damage getting it wrong actually causes. Each answer is a tier that earns a fraction of that weight, from 0 for "not done at all" to 1.0 for a production-grade answer. Your score is earned weight divided by applicable weight, expressed as a percentage. Questions you mark "not applicable" are removed from both sides, so they neither help nor hurt. Grades are A at 90 and above, B at 75, C at 60, D at 40, and F below that.
How are AI model prices verified?
Two passes. An automated drift check runs every Monday and compares every price we publish against the provider's own pricing page, opening an issue when one no longer matches. It reports rather than edits, because a scraper that guesses wrong would publish a wrong number with the authority of an automated check. On top of that, a person re-verifies the whole roster monthly against each provider page. The roster covers 79 models from 22 providers and is at version 2026-09-05.
What happens when you get something wrong?
The correction is published, dated, with the wrong figure and the right one side by side, in the price change log. We do not quietly edit a number and move on. As of 2026-09-07 there are 3 published corrections, all of them cases where our own published price was wrong.
Do I need an account?
No. All three tools run without one. An account exists only so your progress follows you between devices, and nothing is gated behind it. There is no newsletter you have to join to read anything, and no paid tier.
Is any of this sponsored?
No. No provider, vendor or tool named anywhere on this site has paid for placement, and there are no affiliate links. Where a comparison page recommends one option over another, that is a judgement, and the reasoning is on the page so you can disagree with it.
How do you decide what goes in the weekly briefing?
Every item has to trace to a primary source: provider documentation, an official announcement, or established trade press. Pricing aggregators and SEO summary blogs are excluded because they are wrong often enough to be worse than nothing. Anything in the Hot takes section is opinion and labelled as such. We never state that a named company failed, shut down or was breached unless the linked source says exactly that.
Open the tool.
Three minutes, no account, and you will know which of the ten categories above apply to you.
Run the Stack Grader