April 29, 2026

Weekly AI Roundup: What Happened in AI This Past Week (April 22-29, 2026)

A friendly walk-through of the biggest global AI stories from April 22-29, 2026 (GPT-5.5, DeepSeek V4, Gemini-powered Siri, Anthropic-AWS, Gemma 4) plus five Bangladesh-specific stories from the same week.

Syed Hasibur Rahman

AIWeekly RoundupOpenAIGoogleDeepSeekAnthropicBangladesh

Technology

1886 Words min read8 Minutes, 34 Seconds

2026-04-28 23:00 +0000

Pull up a chair. The kettle is on. If you have been hearing your nephew or your colleague say “GPT this” and “Gemini that” all week and nodding politely without really following along, this post is for you. The last seven days have been one of the busiest stretches in AI I can remember, and most of the news actually matters for ordinary people, not just engineers in San Francisco. Let us go through it together, slowly, the way you would catch up with a friend after Maghrib.

I have split this week into two parts: five big stories from the global stage, then five more that hit closer to home in Bangladesh, whether you are a freelancer in Mirpur, a student in Sylhet, or a garment worker in Gazipur.

5 Big International Developments

1. OpenAI rolled out GPT-5.5

On Thursday, April 23, OpenAI released GPT-5.5, a new version of the AI that runs ChatGPT. The previous version, GPT-5.4, came out only six weeks earlier.

GPT-5.5 is better at writing and fixing computer code, doing research, and what OpenAI calls “carrying more of the work itself.” Until recently, AI was a smart assistant who only did exactly what you asked. The new version is more like one you can hand a whole task to, walk away, and come back to a finished result. On a coding test called Terminal-Bench 2.0 it scored 82.7 percent. GPT-5.5 is paid-only for now, but new features always trickle down to the free tier within a few months.

2. DeepSeek V4 dropped, and it is dirt cheap

A day later, on April 24, the Chinese company DeepSeek released a preview of V4. They made global news last year by showing you do not need billions of dollars or the latest American chips to build a top-tier AI. This week’s release is the follow-up, and the headline is the price.

V4 comes in two sizes: V4-Pro (1.6 trillion parameters) and V4-Flash, the small fast version. Both read up to one million words at a time, roughly four middle Harry Potter books stacked together. Input tokens cost as little as 14 US cents per million for Flash, against several dollars for the equivalent OpenAI tier. The weights are released openly on Hugging Face, so any developer anywhere can download and run it for free. V4-Pro hit a 3,206 rating on the Codeforces coding ladder, ranking 23rd among human competitors worldwide.

3. Google quietly took over Siri

If you use an iPhone, you have probably tried Siri and given up because it cannot really hold a conversation. That is about to change. At Cloud Next 2026 in Las Vegas, Google’s cloud chief Thomas Kurian confirmed on April 22 what had been rumoured for months: Apple’s new Siri, due to land fully when iOS 27 ships in September, will run on Google’s Gemini. Apple is reportedly paying Google roughly one billion dollars a year for the privilege.

Apple and Google are the two biggest competitors in the smartphone world. And yet, when it came to AI, Apple decided building its own was too slow and went to its rival for help. The new Siri will actually understand what is on your screen, remember what you talked about earlier, and chat back like a real assistant.

4. Anthropic signed a $100 billion deal with Amazon

Anthropic, the company behind Claude (another ChatGPT competitor), announced on April 20 it would spend over 100 billion US dollars over ten years on Amazon’s cloud and custom Trainium chips. In return, Amazon will pour 5 billion in immediately and up to 20 billion more later, on top of 8 billion already invested.

To put 100 billion in perspective, Bangladesh’s entire annual budget is roughly 80 billion dollars. One company is committing to spend more than that on cloud computing alone over a decade. Anthropic also locked in five gigawatts of electricity, more than several mid-sized countries use at peak.

5. Google’s Gemma 4 went big at Cloud Next

Earlier in April, Google released Gemma 4, a free family of AI models built for “agentic” work, and Cloud Next this week put serious wind behind it. Gemma 4 comes in tiny versions small enough to run on a phone and bigger versions that hold their own against paid models. The flagship 31-billion-parameter model scored 85.2 percent on a tough knowledge test called MMLU Pro.

Why does this matter outside Silicon Valley? Because “open” means any developer in Bogura, Dhaka, or Khulna can download Gemma 4 and run it on a modest server. No permission, no monthly dollar bill. Together with DeepSeek V4 the same week, that is two giant pieces of free AI infrastructure dropped onto the global table inside seven days.

5 Stories That Hit Closer to Home (Bangladesh)

1. “Hour of AI” reached every secondary school

On Sunday, April 27, the Directorate of Secondary and Higher Education (DSHE) launched a month-long “Hour of AI in Bangladesh-2026” campaign in every secondary school in the country. Students from grades six to nine will get basic coding and a simple introduction to how AI works. The campaign runs until May 26, supported by the global non-profit Code.org and our own a2i team.

If you have a school-going child, ask them this week whether their teacher has mentioned it. In many schools the answer will be “not yet” because rolling out anything to thousands of schools takes time. But the official letter has gone out, and a generation that grows up writing even one line of code will think very differently about AI than one that only consumes it.

2. Outsourcing exports keep climbing thanks to AI

Bangladesh’s BPO and outsourcing exports crossed 900 million US dollars in the first half of the last fiscal year, nearly matching the entire previous year, and the full year 2025-26 is on track to clear one billion dollars for the first time. The reason is not a flood of new clients. It is that our 650,000-plus freelancers are using AI tools like ChatGPT, Claude, and now DeepSeek V4 to do work that was previously out of reach.

A graphic designer in Cumilla who used to charge five dollars for a simple logo can now deliver a full brand kit in a day. A young woman in Khulna who taught herself basic copywriting can produce ten times the volume by editing AI drafts. Multiply by hundreds of thousands and you have a real second pillar of foreign exchange, alongside RMG and remittance.

3. DeepSeek V4 makes Bangla-first apps actually possible

Back to DeepSeek V4 with a Bangladesh hat on. Globally, V4 is a story about China catching up with America at a fraction of the cost. For us, it unlocks something different. Local startups like Speaklar (Bangla AI call centres) and Intelsense AI (bilingual Bangla-English tools) have been bottlenecked for years by the cost of running foreign models — every ChatGPT call cost dollars.

With V4-Flash at fourteen US cents per million tokens and weights freely downloadable, a small team in Banani or NSU can host their own Bangla-tuned chatbot for the price of a modest server. Expect the next twelve months to bring a wave of made-in-Bangladesh AI tools, from farmer advisory bots to homework tutors that actually understand Sylheti and Chittagonian accents.

4. The deepfake crisis against women is now a national emergency

This is the heaviest story this week, and I will not soften it. On April 25, multiple international outlets reported Bangladesh is in the middle of a deepfake abuse crisis aimed at women. Perpetrators pull a normal photo from someone’s Facebook, run it through a free AI tool to generate explicit fakes, then use those to extort the victim or destroy her reputation by sending the file to her family, college, or employer.

Research cited in the reports finds 89 percent of Bangladeshi women on social media have faced online violence at least once. Cases have included suicides. Practical steps any family can take: lock down social media privacy, do not share full-face photos publicly, and if you or someone you know is being extorted, go to the police cyber unit immediately and do not pay. Paying never ends the threat.

5. AI deepfakes hit national security and the Army

On April 28, our Inter-Services Public Relations (ISPR) publicly condemned a coordinated AI-driven disinformation campaign, allegedly linked to a foreign intelligence network, targeting Chief of Army Staff General Waker-Uz-Zaman. AI-generated audio and video were reportedly circulated on social media to manufacture a fake controversy around the Army leadership.

This matters even if you do not follow defence news, because the same technology destroying private women’s lives in story four is now being deployed at the geopolitical level. There is no longer a clean separation between “fun AI tools” and “national security.” The draft National AI Policy 2026-2030, which finished public consultation in February, tries to address this with mandatory algorithmic impact assessments and a ban on social-scoring surveillance. Whether the policy gets the teeth it needs is the fight to watch.

Wrapping up

Ten stories in seven days. Globally: a smarter ChatGPT, a much cheaper Chinese rival, Google quietly taking over Siri, Anthropic and Amazon shaking hands on a sum bigger than our national budget, and a free agentic model from Google. At home: AI in every classroom, a billion-dollar freelancing economy lifted by AI tools, an opening for Bangla-first startups, a deepfake crisis hurting women, and disinformation against our own Army chief.

AI is not coming in some distant future. It is here, improving every few weeks, getting cheaper, and being weaponised. The best response is not to panic and not to ignore it. Pick one tool. Play with it thirty minutes a week. Talk to your kids about what is real online and what is fake. And if your chacha asks you what AI is, send him this post.

See you next week. Until then, take care.

Sources

International stories

Bangladesh stories

See Also