Gemini 4 Carbon tops the charts, banks may owe scam victims, UPI gets a fee
Google's best model is a rank without a release, the Supreme Court is pushing banks to share digital arrest losses, UPI merchants face a 0.4% fee, and China's AI export draft decides who gets the weights. Plus a consumer helpline worth saving.
- Long Press · Daily Five30 Oct 202601 / 05
commentaryGemini 4 Carbon ranks first. Almost nobody can try it.
Google announced Gemini 4 Carbon on October 28, 2026, and it tops the Artificial Analysis Intelligence Index with a score of 59, but only a small group of cyber defenders can use it this week. Everyone else gets 'coming soon' with no date. A ranking you can't test yourself is mostly a press release with a number on it. Google says internal teams have seen it work on engineering tasks for days at a stretch, which is the part worth watching once real developers get access.
The questionIs Google's Gemini 4 Carbon available to the public yet?
In one lineNo. Google announced Gemini 4 Carbon on October 28, 2026, but access starts this week only with a subset of cyber defenders. Broader availability to developers, businesses and consumers is described as coming soon, with no date. It ranks first on the Artificial Analysis index at 59.
The problem
Every model launch arrives with a leaderboard, and by the time you can try the thing the leaderboard has moved on. Gating this one to defenders is a reasonable call for a model that good at finding and exploiting bugs, but it makes the headline claims hard to check.
What changed
Gemini 4 Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench for patching, per Google's announcement. It leads the Artificial Analysis index at 59. Introductory API pricing is $2 per million input tokens and $10 output, doubling after the intro period. Access starts with a subset of cyber defenders through the Fairwind Program.
The catch
Most of the numbers are Google's own internal tests. There's nothing on India pricing, a free tier or the Gemini app, and 'broader availability' has no date.
What it means for you
Nothing to do today. Treat the ranking as a preview, and wait for independent testers to get their hands on it before you rebuild any workflow around it.
Questions people ask
How much does Gemini 4 Carbon cost?
Google's introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period it rises to $4 and $20. Google's announcement doesn't mention India pricing, a free tier or availability in the Gemini app.
What is Gemini 4 Carbon good at?
Google pitches it at long-running agentic work, saying internal teams have seen it run for days on multi-step engineering tasks. It scored 88.2% on SEC-bench Pro for finding vulnerabilities and 71.0% on CWE-bench for patching them. Those are Google's own reported numbers.
SourcesGoogle (The Keyword) ↗
- Long Press · Daily Five30 Oct 202602 / 05
commentaryIndiaSupreme Court wants banks to pay for digital arrest losses
The Supreme Court, hearing its own case on digital arrest scams, has asked RBI to examine making banks compensate victims at least the amount lost where inadequate due diligence let the fraud through, according to The420. The bench led by CJI Surya Kant also floated pausing unusual transactions and alerting the local police. Banks have spent years telling customers to be careful. This turns the instruction around and points it at the bank.
The questionWill banks in India have to compensate digital arrest scam victims?
In one lineNot yet, but the Supreme Court has asked RBI to consider it. A bench led by CJI Surya Kant proposed that banks compensate victims at least the amount lost where weak due diligence let fraud through. The Attorney General says RBI should handle liability through regulation instead.
The problem
Victims of digital arrest and similar frauds are usually told the transfer was authorised, so the loss stays with them. The CJI cited an estimated net loss of about Rs 800 crore to the economy this year.
What changed
The Court asked RBI to examine its proposed KYC Amendment Directions, 2026, expected to be notified by November 30 and operational in December, with institutions indemnifying victims where due diligence failed. It also floated suspending unusual transactions, auto-notifying the local police and physically verifying the transaction, per The420. RBI must submit a list of all authorised financial institutions at the next hearing.
The catch
Nothing is binding yet. The Attorney General argued that liability belongs in RBI regulation, not judicial directions, and the next hearing date wasn't reported.
What it means for you
Until a rule exists, assume the bank will not cover you. Never move money for a caller claiming to be police or CBI, and if it happens, report it at once on 1930 or cybercrime.gov.in.
Questions people ask
When could new RBI KYC rules on cyber fraud take effect?
The proposed RBI KYC Amendment Directions, 2026 are expected to be notified by November 30 and become operational in December. The Supreme Court wants RBI to examine them with victim compensation in mind, and to give the court a list of all authorised financial institutions at the next hearing, which has no reported date.
What other safeguards did the Supreme Court suggest for banks?
The Court floated temporarily suspending unusual transactions, automatically notifying the local police station, and physically verifying whether a transaction is genuine. The CJI also cited an estimated net loss of about Rs 800 crore to the Indian economy from these frauds this year, according to The420's report of the hearing.
SourcesThe420 ↗
- Long Press · Daily Five30 Oct 202603 / 05
tipIndiaBad online order? The helpline most people skip
The National Consumer Helpline takes complaints on WhatsApp at 8800001915, on toll-free 1915 (8 AM to 8 PM) and through the NCH and UMANG apps, and it forwards your grievance to the company, with resolution taking up to 30 days. It's the step before a consumer court, and I suspect most people skip it. A Delhi commission just made the point the hard way: a buyer who got soap instead of a Realme phone in 2021 has a Flipkart refund order in July 2026.
The questionHow do I complain about a wrong e-commerce order in India?
In one lineRegister it on the National Consumer Helpline: call 1915, WhatsApp 8800001915, or use the NCH or UMANG apps. The helpline forwards your grievance to the company, and resolution can take up to 30 days. If you're not satisfied, you can approach the consumer commission.
The problem
A wrong or empty parcel, then a refund that never comes, then a return rejected for a missing photo. That was this buyer's story with Flipkart, whose defence was that it is only an intermediary.
What changed
The District Consumer Disputes Redressal Commission, South II, Delhi ordered Flipkart on July 31, 2026 to refund Rs 10,574 with 7% interest plus Rs 5,000 for harassment and Rs 5,000 costs, rejecting its safe harbour defence, per MediaNama. The helpline route is on the National Consumer Helpline site: 1915, WhatsApp 8800001915, and the NCH and UMANG apps.
The catch
The buyer waited from October 2021 to July 2026, and the commission didn't examine the E-Commerce Rules. NCH is a pre-litigation channel, so it forwards your complaint rather than guaranteeing a refund.
What it means for you
Save 8800001915 now, and keep photos or a video of the unboxing plus every order and reference number. Use the helpline first, and go to the commission only if it stalls.
Questions people ask
Is the 1915 consumer helpline toll-free in India?
Yes, according to the helpline's own page, 1915 is toll-free and agents answer from 8 AM to 8 PM. The page also lists 1800-11-4000 for grievances. You can sign up once on the web portal to get a user ID and file online, and track a grievance without logging in.
What happens if the National Consumer Helpline doesn't fix my complaint?
NCH is a pre-litigation step, so unresolved cases go to the consumer commission. In one Delhi case, a buyer who got soap instead of a phone approached the commission and won a refund of Rs 10,574 with 7% interest, Rs 5,000 for harassment and Rs 5,000 costs, all from Flipkart.
- Long Press · Daily Five30 Oct 202604 / 05
hot takeIndiaUPI's 0.4% merchant fee is coming, and Paytm is pricing it in
NPCI has notified a 0.4% MDR on person-to-merchant UPI payments above Rs 2,000, MediaNama reports, with the start date still to be notified. Paytm's CFO told analysts it could add up to Rs 700 crore a year for the company. That's the tell: one fee, and the first number out is a payments company's windfall. Free UPI was always paid for by somebody. Now the bill may move to the shop, and shops have a long habit of passing costs along to you.
The questionWill UPI payments to merchants above Rs 2,000 carry a 0.4% fee?
In one lineYes, per MediaNama, NPCI has notified a 0.4% merchant discount rate on person-to-merchant UPI payments above Rs 2,000. The start date is still to be notified, and reports don't say who bears it. Paytm says it could earn up to Rs 700 crore a year.
The problem
UPI has been zero-cost for merchants, and every payments app has run on cashbacks, float and hope. A fee changes who pays and who profits.
What changed
NPCI notified a 0.4% MDR on P2M UPI transactions above Rs 2,000, per MediaNama. On the Q2 FY27 call, Paytm's CFO Madhur Deora said a 2.5 basis point margin uplift on Rs 28 lakh crore of GMV works out to about Rs 700 crore a year, maybe Rs 900 crore by FY28.
The catch
Reports don't say who bears the charge, and the start date is unclear: one report ties it to after Diwali, another says NPCI delayed the rollout. Paytm itself calls the number an upside, not part of its model.
What it means for you
Watch how shops handle bills above Rs 2,000 once this starts, since a cash discount or a UPI surcharge is the likely first sign. Small purchases are untouched.
Questions people ask
How did Paytm arrive at Rs 700 crore from UPI MDR?
CFO Madhur Deora said on the Q2 FY27 analysts' call that Paytm expects at least a 2.5 basis point uplift in payments processing margin. Applied to roughly Rs 28 lakh crore of GMV, that gives about Rs 700 crore of extra EBITDA a year, possibly Rs 900 crore by FY28, excluding any platform fees.
What are the UPI market shares of PhonePe, Google Pay and Paytm?
Per NPCI's September data, as cited by MediaNama, PhonePe holds 45.6% of UPI volumes, Google Pay 32.2% and Paytm 8.1%. Paytm CEO Vijay Shekhar Sharma says the company aims for a one-third share of UPI consumer payments, led by product rather than cashback.
SourcesMediaNama ↗
- Long Press · Daily Five30 Oct 202605 / 05
hot takeChina wants to sell AI access and keep the models
China's commerce ministry released a draft on October 16, 2026 that would add 23 AI technologies to its export restriction list, including models at 10^26 training compute, 2 trillion parameters, or a top-10 global ranking, per Reuters. The design is simple: overseas users keep using Chinese AI through websites and clouds, but the weights and know-how stay home. An analyst quoted by Reuters calls it a counterpart to US restrictions on China. The open question is whether labs can still run their best models on foreign-owned servers abroad.
The questionWhat do China's draft AI export restrictions cover?
In one lineChina's commerce ministry draft, released October 16, 2026, would add 23 AI technologies to its export restriction list, covering models with 10^26 training compute or 2 trillion parameters. Overseas users could still reach Chinese AI via websites and clouds, while some transfers or deployments abroad would need approval.
The problem
Frontier AI is now treated like a strategic export, and the rules for who gets the model, not just who gets the chatbot, are being written fast.
What changed
The draft lists 23 technologies, including advanced models, autonomous agents, chip-design technology and methods for coordinating more than 10,000 AI chips, per Reuters. A model is covered if it meets any one of three tests: 10^26 training compute, 2 trillion parameters, or a top-10 global ranking. Comments close November 15 and adoption is expected by the end of 2026.
The catch
It's a draft. The ranking test doesn't say which ranking counts, and law firm Han Kun questioned how size counts for models that activate only some components per task. China's commerce ministry did not comment.
What it means for you
If you build on Chinese models, hosted APIs look like the route the draft protects. Self-hosting open weights is the part to watch, and the rules could still change before the comment window closes.
Questions people ask
When will China's AI export rules take effect?
Public comment on the draft closes November 15, 2026, and adoption is expected by the end of 2026, with implementation guidelines later. A temporary registration requirement for certain transfers that put Chinese technology under foreign ownership or control took effect on October 16, including deals already in progress.
Will DeepSeek, Moonshot and Z.AI models still be available overseas?
Per Reuters, overseas customers currently reach DeepSeek, Moonshot and Z.AI models through websites or cloud platforms, and the draft would generally keep that access. What's unclear is whether labs can train or run top models on foreign-owned overseas servers, and developers may need separate domestic and overseas versions.
SourcesReuters (via WHBL) ↗
Taps
Quick hits. One line each, for when a long press is too much.
- IndiaMediaNama ↗
Sarvam signs AI MoU with HQ IDS for the Army, Navy and Air Force
No value, duration or named projects were disclosed, and the details come only from Sarvam, per MediaNama.
- IndiaMediaNama ↗
Bombay HC refuses to send Blink Digital's Meta suit to California
Order of Oct 27: California courts cannot decide IT Rules violations. Blink wants about 9,000 ad accounts restored; hearing Nov 2, per MediaNama.
- IndiaThe420 ↗
Sonbhadra trust at centre of Rs 11 crore cyber fraud, eight arrested
Over Rs 4 crore went from Rewa to the trust's account in March, then on to 150-plus accounts, with links to 10 states, per The420.
- GlobalMistral AI ↗
Mistral releases Large 4 as open weights, a 1-trillion-parameter model
Modified MIT licence and $0.70 per million input tokens on the API. Cheap to call, not cheap to host, per Mistral.
- GlobalHelp Net Security ↗
Five AI coding tools have browser flaws, Codex and VS Code unpatched
Per Help Net Security on Forever Security's EMBRACE research, Cursor and Copilot are fixed as of Oct 29. Update both, and be careful opening untrusted projects.