Jev AI by TypeSafe Explained: How It Works, Real Use Cases, Cost and Benchmarks (2026)

If you searched Jev AI, JEV AI model, TypeSafe Jev, or Jev AI pricing, here is the short answer.

Jev is an AI model from TypeSafe AI, a San Francisco startup. It went into early access on 15 September 2026. It is not a chatbot. It never writes a sentence. You give it some text and a few questions with fixed answer options, and it gives back a pick, a score or a yes/no probability, along with how sure it is.

That makes Jev less like a chat assistant and more like a very smart if statement inside your software.

This guide explains what Jev is, how it works, where it is useful for Indian businesses, what it really costs, and what the benchmarks say. Every number below has its source named next to it. Some come from TypeSafe and some from independent testers, and we say which is which.

Disclosure: We (SR Infobiz) did not run our own tests on Jev. Everything here comes from TypeSafe's official pages and from independent results that others have published. Our opinion is in the "SR Infobiz verdict" section, and it is based only on that data.

Quick facts

Sources: TypeSafe Models page and launch post, both re-checked 30 Sep 2026.

First, clear up the name (and avoid fake Jev sites)

"JEV AI" is not an AI from Reliance Jio, and it has nothing to do with "JEV", the short form of Japanese encephalitis virus that shows up in medical AI papers. When people say JEV AI in September 2026, they mean TypeSafe's Jev.

The name comes from William Stanley Jevons, the 19th-century economist behind the Jevons paradox: when something gets cheaper to use, people end up using far more of it. TypeSafe expects the same to happen with cheap machine intelligence. "System One" comes from Daniel Kahneman's Thinking, Fast and Slow, where "System 1" means fast, intuitive thinking. (Source: TypeSafe launch post FAQ, 15 Sep 2026.)

Warning: lookalike and reseller sites

Jev became popular very fast, and copycat sites followed:

1. Security firm Eye Security found reseller storefronts such as jev-ai.pro, jevtypesafeai.com and jev-agent.org. They sell access to TypeSafe's own API at 3 to 11.5 times the official price, and your prompts pass through their servers first. The same researchers counted about 670 new domains with "jev" in the name in the eight days after launch. (Source: Eye Security, "Rise of the Jev-Clones", 25 Sep 2026.)
2. Some explainer sites run their own "Jev" API endpoints that TypeSafe does not operate.
3. The jev.ai domain is not TypeSafe's either.

Only use these official routes:

1. TypeSafe: typesafe.ai, docs.typesafe.ai, console.typesafe.ai. The official API lives at api.typesafe.ai.
2. OpenRouter: openrouter.ai/typesafe
3. Vercel AI Gateway: vercel.com/ai-gateway/models/jev

Never paste a TypeSafe API key into any other site.

What is Jev AI, in plain English?

Most AI models you know, like the ones behind chat apps, generate text one word-piece at a time. That is great for writing and explaining. But a lot of business software doesn't need an essay. It needs a decision:

1. Is this ticket about billing or a technical bug?
2. Is this email spam?
3. How hot is this sales lead, on a scale of 1 to 3?
4. Should this comment be removed?

Today many teams ask a large language model (LLM) to answer those questions in text and then parse the reply. TypeSafe's argument is that this is slow, costly and fragile. So it built a model that only makes decisions:

1. Input: your data (the "state") plus typed questions
2. Output: a typed answer for each question, with probabilities, ready for your code to use

TypeSafe calls this new class "System One models", and Jev is the first one. (Source: TypeSafe docs, Introduction.)

Who is behind it? TypeSafe AI was founded in 2024 by Diogo Almeida (CEO), Erik Gafni and Sasha Sheng. According to SiliconANGLE, Almeida previously worked at OpenAI on reinforcement learning from human feedback (RLHF), InstructGPT, ChatGPT and GPT-4. The company came out of stealth with $40 million in seed funding led by DCVC. (Sources: Business Wire press release, 15 Sep 2026; SiliconANGLE, 16 Sep 2026.) Forbes reported a $200 million valuation, according to SiliconANGLE.

How Jev works

The three question types: Choice, Score and Noul

Jev understands exactly three kinds of question. TypeSafe calls them "primitives":

Source: TypeSafe API reference, checked 30 Sep 2026.

"Noul" is TypeSafe's own name for a yes/no question. A Noul value of 0.95 means "very likely yes", and 0.05 means "very likely no".

Many questions, one call, in parallel

You can mix all three types in one request. Jev reads the input once and answers every question in parallel and independently. TypeSafe says adding more questions "barely changes the response time". This is very different from a chat model, which has to write out each answer in sequence.

There is a flip side. Because each question is answered independently, one answer can't "see" another. TypeSafe's advice is to keep each question small and specific, then combine the answers in your own code.

Example: what a Jev call looks like

This example is taken from TypeSafe's official quickstart documentation. The numbers are TypeSafe's published sample, not our test.

Example request:

Example response (shortened):

Source: TypeSafe Quick start.

Your code can now route the ticket to the technical team, flag it as urgent, and send it to a human when confidence drops below whatever threshold you choose.

One detail matters for cost. That ticket is only about 25 words, yet the call billed 392 input tokens. The questions and option descriptions count as input too. We use this in the cost section below.

"Calibrated confidence": what it means

TypeSafe trains Jev with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD). The goal is honest probabilities: when Jev says 90%, it should be right about 90% of the time. The practical idea:

1. High confidence → let the software act on its own
2. Low confidence → send the case to a person or a bigger model

TypeSafe has not published a paper on RLCD, the model's size or its architecture. (Source: launch post.) As the benchmark section shows, independent testers found that the confidence is most trustworthy at the extremes and should be tuned on your own data.

"Jev can't hallucinate": what that really means

TypeSafe says Jev "can't hallucinate" and never makes type errors. Read that carefully:

1. True: Jev can only answer with the options you defined. It can't make up a category, return broken JSON, or suddenly start chatting.
2. Not true: that every answer is correct. TypeSafe's own FAQ says so plainly: "Jev guarantees the shape of its answers, not that every decision is correct… it can choose the wrong one." (Source: typesafe.ai homepage FAQ, checked 30 Sep 2026.)

In the launch post, TypeSafe also admits its "0% hallucination" figure is "not empirical". It follows from how the output format is designed, not from a measurement.

Our reading: "can't hallucinate" really means "can't go off-script". That is valuable for automation, but it is not the same thing as accuracy.

Real use cases for Jev (with an India and business angle)

Jev fits anywhere your software makes many small, repeatable judgment calls on short text. Below are the use cases we think make sense for Indian businesses. These are our suggestions, based on the task types that independent testers found Jev handles well: short input, clear labels.

1) Support ticket routing

An e-commerce brand, a SaaS company or a fintech app can route every incoming ticket, email or WhatsApp message in one call:

1. Choice: billing / technical / delivery / KYC
2. Score: frustration level
3. Noul: "asks for a refund?", "mentions a legal complaint?"

Clear cases go straight to the right queue. Low-confidence ones go to a human.

India note: Many Indian customers write in Hindi, Hinglish or regional languages. TypeSafe itself advises testing non-English content before relying on it (see "Where not to use Jev" below).

2) Spam and fraud-signal filtering

Spam filtering on short text is where independent testers saw Jev perform best (numbers below). For fraud, use Jev to flag signals, for example "Does this message ask the user to share an OTP?" or "Does it pressure the user to pay urgently?". Keep the final block/allow decision in your own rules. Jev is a filter, not a fraud engine.

3) Lead scoring for sales teams

Score each inbound enquiry (a website form, a B2B marketplace lead, an email) on separate small questions:

1. Score: budget clarity, urgency
2. Noul: "Is this a real business enquiry?"
3. Choice: product line

Then combine them with your own weights in code. This follows TypeSafe's "composite scoring" advice: when your priorities change, you change a number in code instead of rewriting a prompt. (Source: TypeSafe docs, Introduction.)

4) Content moderation

Community apps, marketplaces and review sections can ask Jev whether a post is abusive, contains an external link or promotes something off-platform. Because the answers are probabilities, you can auto-remove only the very confident cases and queue the rest for moderators.

5) Invoice and document triage (first pass only)

Jev can sort documents (invoice vs purchase order vs receipt), check whether a page is relevant, or flag a document for review. It should not do the matching, totals or GST calculations. On TypeSafe's own eval board, invoice processing is where Jev trails the leader by the widest margin (61.8% vs 79.1%), and TypeSafe warns against using it for maths and dates. Use it to sort the pile, not to approve payments.

6) E-commerce review tagging

Tag thousands of product reviews at once by sentiment, topic (delivery, packaging, quality, sizing) and whether the customer says they will buy again. At Jev's price, tagging every review becomes affordable even for a small D2C brand.

7) The "cheap filter before an expensive model" pattern

This is the pattern we find most convincing, and independent testers reached the same conclusion:

1. Send every item to Jev first.
2. Confident yes / confident no → act immediately. This is often the bulk of the traffic.
3. Unsure (the middle zone) → send only those to a stronger LLM or a human.

That way you only pay LLM prices on the hard cases. LangChain has built tools around the same idea: routing requests to cheaper or stronger models, and checking an agent's risky tool calls before they run. (Source: LangChain blog, 17 Sep 2026.) For the stronger model, see our breakdowns of Claude Opus 5.5 and Grok 4.7.

Where NOT to use Jev

TypeSafe publishes a candid list of Jev 1.13's weak spots (the "jaggedness" page). In simple terms:

Source: Jev 1.13 jaggedness, last reviewed by TypeSafe on 17 Sep 2026; Models page.

One more surprising example from that page: asked "Is the customer asking for a refund?" and "Is the customer asking for something other than a refund?" about the same ticket, Jev gave 0.72 and 0.47, which add up to 1.19. So don't expect logical consistency across separately worded questions.

Jev AI pricing and cost

The official price

Source: TypeSafe Models page, re-checked 30 Sep 2026. OpenRouter lists the same $0.042/M input and $0/M output for Jev 1.13 (OpenRouter, checked 30 Sep 2026). Vercel AI Gateway also lists $0.042/M input (Vercel).

Worked example: 1 million support tickets

Simple version (300 tokens per call in total):

1. 1,000,000 tickets × 300 tokens = 300,000,000 tokens (300 million)
2. 300 million ÷ 1 million = 300 "million-token units"
3. 300 × $0.042 = $12.60 for all one million tickets
4. Per ticket: $12.60 ÷ 1,000,000 = $0.0000126

More realistic version (the questions count too):

Remember that TypeSafe's own ~25-word sample ticket with three questions billed 392 input tokens. If your ticket text is about 300 tokens and your questions add roughly 400 more, each call is about 700 tokens. (That 400 is our planning assumption, not a TypeSafe figure. Measure your own.)

1. 1,000,000 × 700 = 700,000,000 tokens
2. 700 × $0.042 = $29.40 for one million tickets

In Indian rupees: at 1 USD = ₹95.98 (European Central Bank reference rate for 29 Sep 2026, via the Frankfurter API, fetched 30 Sep 2026):

1. $12.60 × 95.98 ≈ ₹1,209
2. $29.40 × 95.98 ≈ ₹2,822

Your actual card charge may differ because of bank FX margins, GST and payment fees.

For context: TypeSafe's launch post says input tokens on existing LLMs cost "from $0.20 to $10 / MTok", with output tokens roughly 5x more expensive. That is TypeSafe's own comparison. Input alone for the same 300 million tokens would then cost $60 to $3,000, before any output tokens.

Rate limits: how long would 1 million tickets take?

Official limits (TypeSafe says they are "adjusting dynamically" while demand is high):

1. 100,000 tokens per second
2. 40 requests per second

The math for one ticket per request:

1. Token limit: 300,000,000 ÷ 100,000 = 3,000 seconds (about 50 minutes)
2. Request limit: 1,000,000 ÷ 40 = 25,000 seconds (about 6.9 hours)

So the request limit is the real bottleneck. You could pack several tickets into one request, but TypeSafe warns that accuracy drops as the input grows, so test that before relying on it. Higher limits are available on custom/enterprise plans. (Source: Models page.)

Context and input limits

1. 64k tokens per request in total, and 32k for the state plus the longest single question (TypeSafe docs)
2. OpenRouter's listing shows a 32,000-token context for Jev 1.13
3. Text only. Images, PDFs and audio must be converted to text first

Free credits and sign-ups

1. At launch (15 Sep) access went through a waitlist.
2. Sign-ups were paused around 22 Sep because of demand (reported by Eye Security, 25 Sep 2026).
3. Sign-ups reopened on 27 Sep, but the $5 free credit for new users was temporarily switched off. This was reported by aifront-page.com (28 Sep 2026), quoting a post on X by CEO Diogo Almeida that blamed "a few bad actors". We could not confirm the free-credit status on an official TypeSafe page.
4. On 30 Sep 2026 we confirmed that the TypeSafe console shows a normal sign-in page (Google or email code) with no waitlist message.

Is the price subsidised?

TypeSafe gives two messages:

1. Launch post: "We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)."
2. Homepage FAQ: "We can serve Jev profitably at our current prices."

Our take: plan with the current price, but don't build a business case that only works at $0.042. Nobody outside TypeSafe can verify the cost of serving Jev.

Jev benchmarks: what TypeSafe claims vs what others measured

Benchmarks are where Jev's story needs the most care. TypeSafe deliberately does not publish results on public benchmarks (launch post FAQ). So we split the numbers into two groups.

(a) TypeSafe's own claims (self-reported)

Caveats TypeSafe itself states (launch post): the 193.6x / 444.6x figures are "on the higher end of real world gains". The test workflows "were made by individuals on our model capabilities team, so some bias could exist". Timings were run "from our laptops on the West Coast".

TypeSafe's workflow eval board. "Accuracy" here means agreement with the average answer of two large models (GPT-6 Astra and Claude Fable 5.1 at high thinking), not with human-checked labels. Mean over four workflows (security incidents, agent traces, invoices, customer service):

Source: evals.typesafe.ai (TypeSafe, self-reported), "workflow" setup, checked 30 Sep 2026.

How to read this: even on TypeSafe's own board, Jev is not the most accurate. It sits mid-pack on accuracy and is by far the cheapest and fastest. Per workflow, Jev scored 76.0% on customer service but only 61.8% on invoice processing, where the best model scored 79.1%.

(b) Independent results (not by TypeSafe)

Aman Kumar's test on four public datasets (300 items each, about 16,000 calls in total):

Source: Aman Kumar, "Testing Jev on public and private data: classifier or filter?", amankumar.ai, 18 Sep 2026. Independent blog post, not peer-reviewed; code and per-item answers published on GitHub.

He also reported a median response of about 0.8–0.9 seconds from his laptop, compared with 1.4–5.0 seconds for the two small OpenAI models. Answers where Jev was highly confident were right 90–100% of the time on tasks it could do, while mid-range answers were close to a coin flip. His conclusion: Jev works best as a filter, not a replacement.

A review of independent studies, eight days after launch (dev.to):

The table below is from one six-model comparison (200 items drawn from BANKING77, BoolQ, Yelp and ChaosNLI, by Manjunath Janardhan), which the review re-scored from the published logs:

Source: xbill, "Jev After Eight Days of Independent Tests: Level With Mid-Price LLMs, Behind the Frontier", dev.to, 24 Sep 2026. Community review, not peer-reviewed.

Other findings collected in the same dev.to review. Several of these come from arXiv preprints that were only days old and not peer-reviewed:

1. Spam: 98.33% on 18,514 emails, level with a trained logistic-regression baseline (98.39%)
2. Re-ranking: 0.692, level with a dedicated re-ranker at 0.691
3. Format reliability: zero invalid answers across 23,703 decision-model calls in the largest study
4. Big study (arXiv 2609.24574, 7,977 human-labelled items): Jev trailed the best of 19 LLMs on 14 of 15 tasks, by a median of 11.6 macro-F1 points
5. Languages: Russian (XNLI) accuracy fell from 88.3% to 77.3%; Spanish cost 3 to 6 points
6. Real-world speed and cost gains measured across studies ranged from 0.5x to 12.1x faster and 0.6x to 478x cheaper, depending on what Jev was compared against
7. Re-checking the headline: the review found that the 444.6x figure matches Jev vs Opus 5 in TypeSafe's workflow setup. Averaged across all eight setups, it works out to 97.8x faster and 149.2x cheaper

What about Hindi and Indian languages?

As of 30 September 2026, we found no Hindi or Indic-language benchmark for Jev. The closest data point is a small test by MarketDX using hand-written sentences in 23 languages including Hindi. It reported only averages across all languages, with no separate Hindi score. (Source: MarketDX on Hashnode, 19 Sep 2026.) TypeSafe itself says English is where accuracy is best. If your customers write in Hindi or Hinglish, test on your own data before you trust it.

SR Infobiz verdict (our opinion)

This is our opinion, based only on the official and independent data cited above. We did not test Jev ourselves.

Pros

1. Extremely cheap. $0.042 per million input tokens with free output. A million short classifications costs dollars, not thousands of dollars.
2. Fast. Independent testers measured under a second per call from their own machines, and TypeSafe claims 70–500 ms end to end.
3. Answers always fit your options. No parsing, no broken JSON, no chatty replies. For automation, this is a real advantage.
4. Probabilities you can act on. Independent tests suggest the very confident answers are highly reliable, which makes "automate the easy cases, escalate the rest" practical.
5. Honest documentation. TypeSafe publishes its own weak spots and the biases in its evals. That is rarer than it should be.

Cons

1. Not the most accurate. Both TypeSafe's own board and independent studies place Jev at mid-price-LLM level, and behind the top models.
2. The headline multipliers are self-reported. Independent comparisons show a much wider, often smaller, range.
3. Black box. No paper, no model size, no open weights. RLCD has not been explained publicly.
4. English-first. There is no Hindi benchmark yet, and other non-English results show accuracy drops.
5. Early-stage access. Rate limits "adjusting dynamically", sign-ups paused and then reopened, free credits reportedly switched off, and pricing whose long-term sustainability TypeSafe says it can't prove yet.
6. Weak at maths, dates, long documents and multi-step reasoning. These are exactly the parts of invoice or finance workflows where mistakes are costly.

Who should try Jev

1. Yes: teams with high-volume, short-text English decisions such as ticket routing, spam, tagging and moderation, where a confident first pass can remove most of the load from a costlier model.
2. Maybe: Hindi or Hinglish workflows. Only after testing on a few hundred of your own labelled examples.
3. No: anything that needs written output, calculations, date logic, or data that can't leave your network. For that last case, look at open models you can host yourself.

In one line: Jev is best treated as a very cheap, very fast first-pass filter, not as a replacement for a strong LLM.

How we researched this

1. Official sources first. We read TypeSafe's launch post, homepage, docs (Models, API reference, Quick start, known weaknesses, legal) and eval board. On 30 Sep 2026 we re-checked the price, rate limits, context limits and the console sign-in page against the live official pages.
2. Distribution partners. We checked the Jev listings on OpenRouter and Vercel AI Gateway for price and context.
3. Independent results. We used only published tests that name their method and data, and we name the author and date next to every number. We flag where results are self-published or from preprints that are not peer-reviewed.
4. No hands-on Jev test. We did not create a TypeSafe account or run Jev. Jev needs an account and an API key, and new users reportedly no longer get free credits. No number in this post comes from our own Jev testing.
5. Our own work is limited to the cost arithmetic (shown step by step), the rate-limit maths, and the analysis and opinions, which are labelled as ours.
6. Model names such as "GPT-6 Astra", "Claude Fable 5.1", "sol", "terra" and "luna" are written exactly as they appear in the cited sources.

FAQ

What is Jev AI?

Jev is a "System One" decision model from TypeSafe AI, launched in early access on 15 September 2026. It doesn't generate text. It answers typed questions (Choice, Score, or yes/no "Noul") about the data you send and returns probabilities that your software can act on.

Is Jev AI the same as Jio AI?

No. Jev comes from TypeSafe AI, a San Francisco startup, and has no connection to Reliance Jio. It is named after the economist William Stanley Jevons.

How much does Jev AI cost?

$0.042 per million input tokens, and output tokens are free (TypeSafe Models page, checked 30 Sep 2026). For example, one million requests of about 300 tokens each would cost about $12.60.

Does Jev AI have a free tier or free credits?

New sign-ups reopened on 27 September 2026. According to aifront-page.com, which quoted the CEO's post on X, the $5 free credit for new users was temporarily disabled. We could not confirm this on an official TypeSafe page, so check the console before you sign up.

Can Jev AI hallucinate?

It can't produce an answer outside the options you define, which is what TypeSafe means by "can't hallucinate". But it can still pick the wrong option. TypeSafe's own FAQ says so.

Does Jev AI work in Hindi?

It accepts Hindi and other languages, but TypeSafe says English is its primary language, and no Hindi benchmark had been published as of 30 September 2026. Test it on your own Hindi or Hinglish data first.

Bottom line

Jev AI is a genuinely different kind of AI model. It doesn't talk; it decides. For Indian businesses handling large volumes of tickets, reviews, leads or messages, it could make AI triage affordable at scale, as long as you keep maths, dates and final judgment in your own code and send the unsure cases to a human or a stronger model.

TypeSafe's headline numbers are impressive but self-reported. Independent testers see a cheaper, faster, mid-accuracy model that shines as a first-pass filter. Test it on your own data before you trust it.

Handling customer data while you test AI tools? Keep sensitive data in your browser. SR Infobiz offers free, privacy-first tools that run entirely in your browser with no sign-up, including a JSON Formatter, JWT Decoder, Hash Generator and One-Time Secret Sharing for API keys: explore the free tools.

Want help designing an AI triage workflow, choosing between a cheap filter and a bigger model, or building the software around it? Email inquiry@srinfobiz.com.

Sources

All accessed 30 September 2026 unless noted.

Official (TypeSafe)

1. "Introducing System One Models & Jev", TypeSafe AI (Diogo Almeida), 15 Sep 2026: https://typesafe.ai/blog/introducing-system-one-models-and-jev
2. TypeSafe AI homepage and FAQ, TypeSafe AI, checked 30 Sep 2026: https://typesafe.ai
3. "Introduction", TypeSafe docs: https://docs.typesafe.ai/introduction
4. "Models", TypeSafe docs (price, rate limits, context, languages): https://docs.typesafe.ai/models
5. "API reference", TypeSafe docs: https://docs.typesafe.ai/api
6. "Quick start", TypeSafe docs: https://docs.typesafe.ai/introduction/quickstart
7. "Jev 1.13 jaggedness", TypeSafe docs, last reviewed 17 Sep 2026: https://docs.typesafe.ai/model-jaggedness/jev-1.13
8. "Parallel questions" cookbook, TypeSafe docs: https://docs.typesafe.ai/cookbooks/parallel_questions
9. "Re-ranking" cookbook, TypeSafe docs: https://docs.typesafe.ai/cookbooks/rerank_typesafe
10. "Structured workflows for automation tasks" (workflow eval board), TypeSafe AI: https://evals.typesafe.ai/
11. TypeSafe console sign-in page, checked 30 Sep 2026: https://console.typesafe.ai
12. "TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI", Business Wire, 15 Sep 2026: https://www.businesswire.com/news/home/20260915525333/en/

Distribution partners

1. "Jev 1.13" and "Typesafe" provider pages, OpenRouter, checked 30 Sep 2026: https://openrouter.ai/typesafe/jev-1.13 and https://openrouter.ai/typesafe
2. "Jev" model page, Vercel AI Gateway: https://vercel.com/ai-gateway/models/jev
3. "TypeSafe AI's Jev now available on AI Gateway", Vercel changelog, 16 Sep 2026: https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway
4. "Building a Harness with Jev", LangChain blog, 17 Sep 2026: https://www.langchain.com/blog/building-a-harness-with-jev

News and independent sources

1. "TypeSafe AI exits stealth with $40M to build AI for use by software", SiliconANGLE, 16 Sep 2026: https://siliconangle.com/2026/09/16/typesafe-ai-exits-stealth-with-40m-to-build-ai-for-use-by-software/
2. "Testing Jev on public and private data: classifier or filter?", Aman Kumar (amankumar.ai), 18 Sep 2026: https://amankumar.ai/blogs/jev-measured
3. "Jev After Eight Days of Independent Tests: Level With Mid-Price LLMs, Behind the Frontier", xbill (dev.to), 24 Sep 2026: https://dev.to/gde/jev-after-eight-days-of-independent-tests-level-with-mid-price-llms-behind-the-frontier-1kln
4. "Is Jev Smart Enough? Testing a Decision-Model AI (Part 1)", MarketDX (Hashnode), 19 Sep 2026: https://marketdx.hashnode.dev/is-jev-smart-enough-language-and-world-knowledge-en
5. "Rise of the Jev-Clones", Eye Security Research (Dion Fieret and Lucas Hop), 25 Sep 2026: https://research.eye.security/rise-of-the-jev-clones/
6. "TypeSafe AI Reopens Jev Sign-Ups, Suspends Free $5 Credit After Surge in Demand", aifront-page.com (Vaibhav Jha), 28 Sep 2026: https://aifront-page.com/typesafe-ai-reopens-jev-sign-ups-free-credit-suspended/
7. USD/INR reference rate for 29 Sep 2026 (European Central Bank data), Frankfurter API, fetched 30 Sep 2026: https://api.frankfurter.dev/v1/latest?base=USD&symbols=INR