Weekly AI Roundup: What Happened in AI This Past Week (July 4–10, 2026)

Welcome back. The city is in that mid-monsoon rhythm now — grey mornings, sudden downpours, the smell of wet earth that even Dhaka’s concrete cannot fully suppress — and the week just gone had a fittingly two-sided character. On the global stage, the frontier reopened: after a month in which the most powerful models kept getting switched off and gated, OpenAI finally put its new GPT-5.6 family in the public’s hands, xAI shipped a serious new coding model, and — for the first time — the United Nations convened a global dialogue where every country, not just the G7’s inner circle, had a seat to talk about how AI should be governed. And at home, the government finally said out loud, in Parliament, what it is actually doing about the deepfake crisis this newsletter has tracked for weeks. Answers, at last, on both fronts — even if the answers raise as many questions as they settle.

Two halves as always. Five global, five from home. The same two questions on each — who benefits, who pays — and the same effort to read every story from Dhaka as well as from Geneva or San Francisco. Let us begin.

5 Big International Developments

1. OpenAI’s GPT‑5.6 — Sol, Terra, and Luna — moved toward broad release

The frontier began reopening to everyone. OpenAI’s GPT‑5.6 family — previewed on June 26 and rolling out more widely through early July — comes in three tiers: Sol, the flagship reasoning-and-agent model; Terra, a mid-range model that OpenAI prices at half the cost of the previous GPT‑5.5; and Luna, a fast, low-cost option. (A plain-language note: the names now describe durable capability tiers rather than one-off models — think Sol for the hardest thinking, Terra for everyday work, Luna for speed and low cost.) OpenAI’s published per-million-token prices are Sol at $5 in / $30 out, Terra at $2.50 / $15, and Luna at $1 / $6. The preview had been deliberately limited: OpenAI says that, at the U.S. government’s request, it started with “a small group of trusted partners” because Sol’s stronger cybersecurity and vulnerability-finding abilities warranted extra safeguards — the same approval-gate dynamic this newsletter covered a fortnight ago. Notably, OpenAI itself pushed back in its own announcement, writing that it does not “believe this kind of government access process should become the long-term default.” (A note on timing: multiple reputable outlets reported GPT‑5.6 reaching general availability on July 9; OpenAI’s own preview page still read “limited preview” at the time of writing, so I describe this as a rollout completing this week rather than leaning on a single announcement date.)

The significance is twofold. First, the tiering — flagship, mid-range, fast-and-cheap — is the maturing shape of the model market: not one model but a portfolio matched to cost and task, which is how a technology moves from novelty to infrastructure. Second, the story extends the government-gating arc. A frontier model that began life available only to a handful of vetted partners is now moving toward broad availability. On the evidence so far, the state’s approval gate looks like a delay rather than a permanent wall — at least this time.

Who benefits, and who pays? Users and developers benefit from a stronger, cheaper, better-segmented model family; the cheaper Terra and Luna tiers especially benefit anyone cost-sensitive. OpenAI benefits from the goodwill and revenue of a broad release ahead of its eventual IPO. The ones who pay are competitors now facing a strong, aggressively-priced portfolio — and, in the background, the open question of whether the next model’s government gate will lift as smoothly as this one’s did.

The Dhaka view is practical and, for once, mostly good. The cheaper tiers matter enormously for a cost-constrained ecosystem. Terra at roughly half the previous cost, and Luna as a fast low-cost option, are exactly the price points at which Bangladeshi developers, students, and small businesses can actually build — the difference between AI being a demo and AI being affordable in a workflow. The right move for local builders this week is empirical: test Terra and Luna on your real tasks, in Bangla and English, and work out which tier fits which job. The frontier being cheaper and more segmented is the single most useful global development, this week, for the people who read this newsletter.

2. xAI launched Grok 4.5, built for coding and agents

The model race stayed crowded. On July 8, xAI released Grok 4.5, its first model built specifically for coding and agentic work, reportedly trained in part on real developer-session data. Early placement on independent capability indices put it competitively high — above open-weight models and some major rivals — at a price reported to be substantially lower than the top-tier Claude and GPT offerings, with API pricing cited around $2 per million input tokens and $6 per million output. (One note of caution: some coverage used unusual corporate framing around xAI’s structure that is worth treating carefully; the model launch and its pricing, however, are well corroborated.)

The significance is the same competitive dynamic pulling in the same direction: strong coding-and-agent capability, aggressively priced. Grok 4.5, GPT-5.6’s Terra tier, Claude Sonnet 5 from last week — the pattern across three labs in two weeks is capable coding models at falling prices. For anyone who writes software or builds agents, this is a genuinely good moment: the tools are getting better and cheaper simultaneously, which does not always happen.

Who benefits, and who pays? Developers and companies building agentic software benefit from more capable, cheaper options and from the competition holding prices down. xAI benefits from a credible entry in the most commercially important category. The ones who pay are the higher-priced incumbents whose margins the aggressive pricing pressures.

The Dhaka view echoes the previous story: price competition among capable coding models is a direct benefit to Bangladesh’s large developer and freelancer economy. When three labs are racing to offer strong coding models cheaply, the cost of AI-assisted software work — the country’s fastest-growing tech export activity — falls. The strategic caution is the recurring one about portability: with capable models now arriving from multiple labs at similar prices, there is no reason for a Bangladeshi team to lock itself to a single provider. Build for the market of models, not for one vendor, and let them compete for your business.

3. The UN convened its first Global Dialogue on AI Governance in Geneva

The governance story took a genuinely important turn — and this one matters more for Bangladesh than any G7 session. On July 6 and 7, the inaugural UN Global Dialogue on AI Governance was held in Geneva, the first session of a standing United Nations platform where all governments and stakeholders — not just the wealthy few — can shape AI cooperation. It closed with a call to turn principles into “concrete actions” before the next session in New York in May 2027, with heavy emphasis on narrowing the global AI divide and embedding human rights. The UN Secretary-General used the occasion to push priorities including AI safety testing and liability, a child-safety pledge, and powering AI data centres with renewable energy by 2030.

This is the counterweight to the G7’s exclusivity that this newsletter has been waiting for. When AI governance happens only at the G7, Bangladesh has no seat. When it happens at the UN, Bangladesh does — the same seat as everyone else. The Global Dialogue is, for now, long on principles and short on binding commitments, and it would be naïve to overstate what a first session achieves. But the existence of a universal forum, where a country like Bangladesh can actually participate in shaping the norms rather than merely inheriting them, is structurally significant in a way that a communiqué’s specific language is not.

Who benefits, and who pays? The countries outside the G7’s inner circle — the global majority — benefit most from having a forum where their voice counts. Human-rights advocates benefit from the framing. The ones who “pay,” relatively, are the powerful states and labs that would prefer to set the rules in the smaller, friendlier room of the G7. The tension between the exclusive forum (G7, where power sits) and the inclusive one (UN, where legitimacy sits) is now the central structural fact of AI governance.

The Dhaka view is a direct call to action, and it is the most important thing in this week’s roundup for Bangladesh’s policymakers. This is the forum where Bangladesh has standing, and it should use it deliberately. That means showing up with a prepared position, not just attendance; building coalitions with other developing countries who share the interest in affordable access and against a two-tier “trusted partner” world; and pushing hard on the AI-divide and human-rights framing that serves Bangladesh’s interests. The country cannot get into the G7. It is already in the UN Dialogue. The gap between having a seat and using it well is the difference between shaping the norms and being shaped by them — and closing that gap costs diplomatic effort, not money.

4. Crunchbase data showed a record $510 billion in H1 startup funding — AI-dominated

The scale of the money became official. Crunchbase data released around the start of July showed global venture funding hit a record $510 billion in the first half of 2026 — more than all of 2025 combined — with AI capturing the overwhelming majority. OpenAI and Anthropic alone reportedly accounted for roughly $217 billion, about 43 per cent of all first-half startup funding, and AI captured more than 70 per cent of second-quarter capital. The half also saw enormous exits, including a landmark SpaceX public listing.

The significance is the concentration. This is not a broad-based venture boom; it is an AI boom, and within AI, a boom concentrated in a tiny handful of names. Two companies taking 43 per cent of all startup funding is a degree of concentration that should give even AI optimists pause — it means the capital markets are making an enormous, concentrated bet on a very small number of players, and the fragility that implies is real. Booms this concentrated have a way of correcting.

Who benefits, and who pays? The frontier labs and their investors benefit from unprecedented capital. The AI infrastructure ecosystem benefits from the spillover. The ones who pay, potentially, are everyone exposed to the correction if the concentrated bet disappoints — and the thousands of non-AI startups competing for the shrinking slice of capital that AI did not absorb.

The Dhaka view is a sobering piece of context for the country’s own startup ambitions. Bangladesh’s entire startup-funding activity in 2025 was measured in the low tens of millions of dollars; the FY27 fund is roughly $41 million. Against a global half-year of $510 billion, those numbers are a rounding error. That is not a counsel of despair — Bangladesh is not competing to build the next frontier lab — but it is a counsel of realism. The country’s opportunity is not in the capital-intensive frontier; it is in the application layer, in solving local problems, and in building companies that are capital-efficient by necessity. The global funding numbers are a reminder of which game Bangladesh is actually playing, and which one it is not.

5. Apple added Siri voice customization to the iOS 27 beta

The consumer story ticked forward. In early July, the latest iOS 27 beta added the ability for testers to customise Siri’s speaking pace and expressivity — sliders for speed and emotion, and a choice among voices and accents — part of the broader generative-AI overhaul of Siri that Apple unveiled at its June developer conference. The iOS 27 public beta is slated for this month, with general release expected in September; the AI features require recent iPhone hardware.

The significance is incremental but real: it is the slow, careful productisation of the generative-Siri overhaul, the point at which the AI stops being an announcement and starts being something two billion device-owners actually touch. Voice customization is a small feature, but it signals that the substantive reinvention of the world’s most widely deployed voice assistant is moving from keynote to shipping software.

Who benefits, and who pays? Apple users benefit from a more capable, more personalisable assistant. Apple benefits from keeping its enormous installed base engaged with on-device AI. The ones who “pay” are owners of older hardware excluded from the features, and — as always with Apple’s AI — anyone concerned about which data flows to which partner model behind the friendly new voice.

The Dhaka view is the familiar one about Bangladesh’s overwhelmingly Android market: iOS features reach a small, urban, affluent slice of the country directly. But the more interesting angle is language. A more capable, more customisable Siri raises, again, the question this newsletter keeps asking of every consumer AI: how well does it work in Bangla? The frontier assistants are being tuned for the world’s major markets, and Bangla — a language with a quarter of a billion speakers — remains chronically under-served. The most useful thing Bangladesh’s technologists can do about the consumer AI wave is keep pushing, loudly, for Bangla-language capability, because the market will build for the languages whose speakers are visible and vocal.

5 Stories That Hit Closer to Home (Bangladesh)

1. The government told Parliament it has launched two projects against AI-driven misinformation and deepfakes

The answer this newsletter has been waiting weeks for arrived, on the floor of Parliament. On July 8, Information and Broadcasting Minister Zahir Uddin Swapon told Parliament that the government has launched two projects to address the misuse of AI and deepfake technology. Replying to lawmakers, he said the government “is fully aware” of the spread of AI-generated misinformation and deepfakes and acknowledged that many people have already fallen victim; that the country’s existing institutional framework was designed for an “analogue” environment while the new threats are entirely digital; and that, on the prime minister’s instruction, the government had “initially adopted a project-based approach.” He added that the Information Ministry signed memoranda of understanding on sharing China’s AI experience and technology during the prime minister’s recent visit to China. (For transparency: some secondary coverage attached specific tallies — numbers of fact-checks, workshops, and journalists trained — to this statement. Because I could not confirm those figures against the parliamentary record, I have left them out rather than repeat them.)

This is real, and it deserves acknowledgement after weeks of my noting the enforcement gap. The government is doing more than nothing: it has publicly recognised the harm and committed two dedicated projects to it. That is progress worth crediting honestly. The caution — and regular readers will anticipate it — is twofold. First, “two projects” and a project-based approach are not yet the same as the specific capacity I keep calling for: a trained unit that can investigate synthetic-media abuse against ordinary women and secure prosecutions. Second, anti-misinformation machinery in a government’s hands always carries the risk of shading into control of inconvenient speech, which is why any such effort has to be watched in practice, not just noted in the parliamentary record.

Who benefits, and who pays? The public benefits from genuine fact-checking and media-literacy capacity. Journalists benefit from training. The potential cost is the recurring one: that “countering misinformation” becomes a lever against critical journalism rather than a shield for its victims. The value of these projects depends entirely on whether they protect people or police them.

The Dhaka view is measured credit with a sharpened demand. The government has answered the “what are you doing?” question, and the answer is more than I feared and less than is needed. The specific gap remains: the women targeted by deepfakes need an investigative-and-prosecution pathway, not only a fact-checking-and-awareness one. And the EU’s fresh ban on non-consensual synthetic intimate imagery, which I covered last week, is sitting right there as a template Bangladesh could adapt. The right posture is to credit the progress, name the remaining gap precisely, and keep asking — because this week proved that sustained, specific pressure does eventually produce answers on the parliamentary record.

2. Deepfake abuse against women stayed at the centre of the harm story

The crisis behind the government’s response remained vivid. The underlying reality that prompted this week’s parliamentary answer is a documented, escalating pattern of deepfake sexual abuse targeting Bangladeshi women. Survey work has found that the large majority of women social-media users have faced online violence and that most do not report it out of stigma. Among the documented cases is that of a student activist whose face was used, via a fake profile, on more than a hundred confirmed AI-generated deepfake videos that together drew millions of views — a single, concrete illustration of how completely this technology can be turned into a weapon against one person.

I have written the analysis of this crisis in previous weeks and will not repeat it at length; the point of keeping it here, immediately after the government’s response, is to hold the two together. The response (two projects, fact-checking, training) has to be measured against the harm (a woman with a hundred deepfake videos to her name, and countless less-visible victims). The gap between those two things is the exact measure of how much more is required.

Who benefits, and who pays? As every week: inaction, or insufficient action, benefits perpetrators and costs the women targeted and the public life they are pushed out of. The Dhaka view remains that the missing piece is specific investigative-and-prosecution capacity for image-based abuse — cheaper than the AI fund, still not built at the scale the harm demands, and now, at least, the subject of a government that has been forced to say what it is doing.

3. Bangladesh advanced a “design-first” semiconductor strategy

A genuinely forward-looking industrial story. Bangladesh is pursuing a “design-first” path into the semiconductor industry — focusing on chip design, testing, and packaging while deliberately skipping the enormously capital-intensive business of wafer fabrication — with the ambition of growing chip-related exports from single-digit millions of dollars today toward a $1 billion target by 2030. The FY27 budget supported the strategy with tax and duty relief on chip design, testing, and packaging inputs, extended for several years.

The significance is that it is the right kind of ambition — realistic about Bangladesh’s constraints and matched to its strengths. Wafer fabrication requires tens of billions of dollars and geopolitical positioning that Bangladesh does not have. Chip design and packaging require skilled engineers and training — which Bangladesh can build. Choosing the achievable, talent-intensive slice of the semiconductor value chain, rather than fantasising about the whole thing, is exactly the strategic discipline this newsletter keeps arguing for.

Who benefits, and who pays? Bangladeshi engineers and the electronics-export sector benefit if the strategy builds real design capability. The economy benefits from a higher-value export category than garments. The risk — the way it could “pay” — is the perennial gap between a stated strategy and the skilled workforce it requires: a design-first chip industry needs a pipeline of trained VLSI and hardware engineers that the education system does not yet produce at scale.

The Dhaka view is that this is a strategy worth taking seriously and resourcing properly. It connects directly to the talent theme running through this month’s roundups — the olympiad medalists, the future-of-work anxieties, the question of what keeps skilled young Bangladeshis building at home. A credible semiconductor-design industry would be exactly the kind of high-value career path that gives a talented graduate a reason to stay. The chip-design ambition and the talent-retention challenge are the same problem viewed from two ends, and the FY27 tax relief is a start on only one end of it. The other end — the training pipeline — is where the real work is.

4. Post-budget analysis kept scrutinising the Tk 500 crore AI fund

The fiscal story matured from announcement into analysis. Independent post-budget analysis this month continued to scrutinise the FY27 budget’s Tk 500 crore AI and startup fund and its accompanying tax package — the zero turnover tax, the extended income-tax exemptions, the broad VAT relief for early-stage tech firms. The assessment was broadly that this is one of the most startup-friendly budgets Bangladesh has produced, set against the country’s stated ambition to lift the digital economy from a low-single-digit share of GDP toward 10 per cent within five years.

The significance of the shift from announcement to analysis is that scrutiny is where budget promises live or die. A fund that generates favourable headlines in June and disciplined, independent analysis in July is a fund being held to account — which is exactly what should happen. The recurring caveat stands: the tax reliefs are concrete and will benefit real founders and freelancers immediately, while the fund itself is only as good as the governance that deploys it.

Who benefits, and who pays? Founders and freelancers benefit immediately and concretely from the tax relief. The broader ecosystem benefits if the fund is deployed well. Taxpayers pay if it is not, and the analysis serves the useful function of keeping that risk visible.

The Dhaka view, unchanged and worth repeating precisely because it stays true: the tax package is the part that will actually reach ordinary digital workers this year, and it is genuinely good policy. The fund is the part that depends on governance nobody can yet evaluate, because the deployment has not happened. Credit the tax relief now; judge the fund by its investments later; and keep the independent analysis coming, because it is the only thing that turns a budget line into accountability.

5. The National AI Policy remained unfinished at the month’s turn

And so we end where the last three roundups have ended. The National AI Policy 2026–2030 remained in finalization this week, its risk-based framework — the surveillance limits, the ban on social scoring, the mandatory impact assessments for high-risk systems, the coordinating agency — still short of the cabinet approval and gazette notification that would make it enforceable. A month of this newsletter’s coverage has passed; the deployments have multiplied; the policy meant to frame them has not moved to the finish line.

The significance is simply the accumulation. For several weeks running I have reported the policy as unfinished while reporting new AI deployments and directives around it — the central bank’s rule, the cyber amendment, the traffic cameras, the anti-deepfake projects. Each of those is the kind of thing the national policy is supposed to make coherent. Governing AI piece by piece while the unifying framework sits in draft is better than not governing it at all, but it is a second-best that becomes more costly the longer it persists.

Who benefits, and who pays? Near-term flexibility accrues to deployers; the cost falls on the citizens whose protections stay aspirational and on the coherence of the whole governance effort. The Dhaka view is the one I will keep making until it is no longer necessary: the draft is good, the delay is unforced, and finalizing it is the single highest-leverage governance act available to the government this quarter. Four weeks of saying so has not yet moved it. I will say it again next week if I have to.

Wrapping up

Ten stories in seven days. Globally: OpenAI’s GPT-5.6 family — Sol, Terra, and Luna — moved toward broad release after a government-gated preview, pushing capable models to cheaper tiers (Terra at half the price of GPT-5.5); xAI launched Grok 4.5 for coding and agents at aggressive prices, part of a three-lab, two-week race that is making capable coding models cheaper by the week; the UN held its first Global Dialogue on AI Governance in Geneva, the inclusive forum where — unlike the G7 — Bangladesh actually has a seat; Crunchbase data confirmed a record $510 billion in first-half startup funding, dangerously concentrated in a handful of AI names; and Apple added Siri voice customization to the iOS 27 beta as the generative-Siri overhaul moves toward release.

At home: the government told Parliament it has launched two projects against AI-driven misinformation and deepfakes, an answer that is more than I feared and less than the crisis needs; the deepfake abuse of women stayed at the centre of the harm story, measured against a single documented case of one activist with over a hundred deepfake videos; Bangladesh advanced a realistic “design-first” semiconductor strategy aimed at $1 billion in exports by 2030; independent analysis kept scrutinising the Tk 500 crore AI fund and its genuinely good tax package; and the National AI Policy 2026–2030 remained unfinished, as it has for several weeks running.

If there is one thread this week, it is that answers arrived on both stages — and every answer opened a question. The frontier reopened, but on the state’s terms. The UN gave the world a forum, but only principles so far. Bangladesh’s government said what it is doing about deepfakes, and the gap between the response and the harm became measurable rather than merely suspected. That is progress, of the honest kind: not resolution, but clarity about exactly what remains to be done. For Bangladesh specifically, this week handed over two concrete, cost-free opportunities — a seat at the UN Dialogue to use deliberately, and a European ban on synthetic-abuse tools to adapt as law. Both are available. Both require will rather than money. What gets done with them is, as always, the question the next weeks will answer.

If something here caught your interest, share it with one person who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Stay well.