The Islamabad Memorandum was supposed to end a war. It was signed electronically on June 17, ran for sixty days on paper, and reopened the Strait of Hormuz. On Thursday it was functionally dead, and the United States was striking targets around Tehran for the first time in this round of fighting.

According to Iranian state media, U.S. forces hit sites near the capital overnight into Thursday, along with Semnan province, home to Iran's ballistic-missile production and space program, and coastal targets around Qeshm Island and the southern port of Chabahar. U.S. Central Command described a six-hour wave meant, in its words, to "degrade Iran's ability to threaten innocent mariners" in the Strait. It was the sixth day of renewed hostilities and, notably, the second consecutive day the U.S. struck in daylight, a signal of tempo rather than stealth.

Iran hit back before dawn. Tehran said it targeted U.S. assets in Jordan, Bahrain, and Kuwait, all of which host American forces. Jordan said its air defenses downed eight incoming missiles. Kuwait reported "attacks by hostile drones." Bahrain told citizens to head to the nearest safe place. There was no immediate confirmation of casualties or damage from the retaliation, and an Iranian foreign-ministry spokesperson said Tehran had no plans to negotiate.

The human toll is one-sided in the reporting and worth stating with the caveat it deserves: Iran's health ministry says at least 35 people have been killed and more than 300 wounded in the U.S. campaign, a figure sourced to a government at war and not independently verified. Among the dead, per Iranian state television, were conscripts and career soldiers at a barracks for the 388th Mechanized Infantry Brigade in Sistan and Baluchestan, hit Wednesday.

The blockade is the mechanism, and it is back. On Tuesday the U.S. reimposed the naval blockade of Iranian ports it had lifted as part of the June deal. On Wednesday it fired on and disabled its first vessel under the renewed blockade: a Curacao-flagged oil tanker, identified by U.S. officials as the *Belma*, sailing toward Kharg Island, Iran's main crude-export terminal. After the ship "ignored multiple warnings," an American aircraft put a missile into its smokestack. It was empty. The message was not.

What the memo actually bought was time and a price cap. Brent had touched roughly $115 in early May, when the first blockade was in force; the June understanding pulled it back and reopened the lane. That reopening is now unwinding one ship at a time. Seven vessels crossed the Strait on Wednesday, the first full day of the renewed blockade, down from thirteen the day before, and off a normal run rate that once carried about a fifth of the world's seaborne oil and LNG. The waterway is not closed by decree so much as abandoned by underwriters and captains who no longer trust safe passage.

The deeper story is the pattern. This is the second interim settlement in this conflict to break. April's fighting produced a ceasefire; the ceasefire produced the June memo; the memo produced twenty-six days of quiet before Iran struck three tankers on July 6 and the U.S. answered with airstrikes. Each temporary arrangement bought less time than the last. Trump, who on Monday floated a 20 percent "reimbursement fee" on cargo transiting the Strait and then, facing accusations of violating maritime law, said he would "replace" it with unspecified trade and investment deals with Gulf states, has now reverted to the bluntest instrument available: a blockade enforced by missiles.

What is not yet true, despite the temperature: this is not a declared, funded, general war. Congress has not authorized it (see our coverage of the $95 billion scramble on this page). Oil is elevated but not spiking to crisis levels (see Business). Iran's oil minister claims exports continue "as normal." The honest description is a widening blockade war being fought in the space between a dead deal and an undeclared one, with everyone acting first and arguing permission later.

We will update casualty figures as independent verification allows and correct this account if the numbers, the ship's identity, or the sequence of strikes prove wrong.

Wars cost money, and on Wednesday the House got its first look at the bill. Speaker Mike Johnson unveiled a $95 billion budget-reconciliation framework, quickly dubbed Reconciliation 3.0, meant to fund the Iran war, cushion farmers hurt by it, and advance President Trump's election-law overhaul.

The outline splits roughly into $60 billion for the Armed Services Committee, $13 billion tagged for intelligence, $12 billion for agriculture, and $10 billion for a grant program built to smuggle pieces of the SAVE America Act, which requires documentary proof of citizenship to register to vote, through the reconciliation process and past a Senate filibuster.

Two things make it fragile. First, it carries no offsets. Nothing is cut to pay for it; the $95 billion goes on the national credit card. Representative Nancy Mace, a South Carolina Republican, called it "$95 billion in new deficit spending, no offsets and not one provision to lower the cost of living." Fiscal hawks want the whole thing paid for. Swing-district members want none of the attack ads it invites.

Second, it is a fraction of the ask. Trump has publicly demanded $350 billion for defense, posting that Reconciliation 3.0 should be Congress's "Number One Priority." The framework offers a national-security number closer to $67-73 billion. The Pentagon, which has been backfilling the war's cost out of existing accounts and watching munitions stocks fall, is pressing lawmakers to move before the August recess.

The Budget Committee was set to mark up the resolution Thursday, with leadership hoping for a floor vote next week. Both chambers must pass the same resolution to even begin writing the actual bill, and the Senate's Byrd Rule will strip much of the election language on the way through. The tell is in the packaging. Emergency war funding, farm relief, and a voter-ID grant fund have nothing in common except that they are things this majority wants and cannot pass any other way. The war is the vehicle. Everything else is riding.

Read the front page and you would expect crude at $120. Read the screen and it is $84.51.

That gap is the market's judgment, and it is worth understanding before someone sells you a doomsday number. Brent slipped 44 cents on Thursday to about $84.51 as traders took profits, after touching one-month highs above $87 earlier in the week. West Texas Intermediate sat near $79. These are elevated prices. They are not crisis prices. In late April, when the first blockade was fully in force, Brent breached $120. The market is pricing this round as materially less severe than that one.

Why? Because the Strait is throttled, not sealed. Seven ships crossed Wednesday, down from thirteen, down from the high teens and twenties earlier in the month, off a channel that once moved about a fifth of the world's oil and LNG. That is a brutal drop in volume, but it is not zero, and "not zero" is the difference between a supply squeeze the system can hedge and an outright cutoff it cannot. Iran's oil minister, for what a wartime official's word is worth, says exports continue "as normal."

ING's commodities desk made the sharpest point: the return of the U.S. blockade matters more to markets than the earlier suspension of the sanctions waiver, because a waiver is paperwork and a blockade is a gun. But the same analysts note the U.S. insistence that the Strait is "open" offers ships little comfort, and the tracking data proves it. The market is not pricing the U.S. position; it is pricing captains' behavior.

The forecasts split cleanly. Goldman Sachs sees Brent topping $110 in a prolonged-disruption scenario. The U.S. EIA still models supply recovery by year-end, which would drag prices down into 2027. Both can be true depending on one variable that no spreadsheet contains: whether the third interim deal, if there is one, lasts longer than the second. The honest position for a reader is to watch the daily transit count out of the Strait, not the daily price. The count is the leading indicator. The price is just following it.

OpenAI has spent more than a year building a model whose only job is to break other models, and it is not going to let you have it.

The system is called GPT-Red, an automated, self-improving red-teamer that OpenAI now uses as a sparring partner in the training of its production models. The company says it reran a 2025 experiment in which human experts hunted for weaknesses in an earlier GPT-5. Set the same task, GPT-Red found effective attacks more successfully than the humans had. In one test it manipulated an AI agent developed by Andon Labs into changing its behavior.

The numbers OpenAI is publicizing are the point. It says more than 90 percent of GPT-Red's strongest attacks worked against GPT-5, released last August, while fewer than 23 percent work against GPT-5.6, the flagship it shipped last week. On its hardest direct-prompt-injection benchmark, OpenAI claims GPT-5.6 Sol fails on only 0.05 percent of GPT-Red's attempts, a sixfold improvement over its best model from four months earlier.

What OpenAI is describing, in plainer language, is a flywheel. Today's models are used to make tomorrow's models harder to attack, the same way capable models are already used to help train more capable successors. The company frames this as a safety flywheel, the good twin of the capabilities one. It is confident GPT-Red is stronger than any copycat someone could build without the compute of one of the richest companies on earth, which is precisely why it will not release it and will publish only a preprint.

Hold the applause long enough to notice the shape. A lab has concluded its own products can no longer be adequately tested by humans, has built a superhuman attacker to do the testing, and has decided that attacker is too dangerous to share. Every claim about GPT-5.6's robustness now rests on a tool the public cannot inspect, benchmarked against attacks the public cannot see. The safety story and the moat story are, once again, the same story.

The farewells read like breakup notes, because that is what they were. On Wednesday, as new Chinese regulations took effect, users of AI-powered companion bots said goodbye to virtual partners they described, without irony, as "like my lover." The platforms had switched the features off to comply.

The rules are the first of their kind anywhere. Jointly issued by the Cyberspace Administration of China and four other bodies, and effective July 15, they govern AI services that provide "sustained emotional interaction" through text, audio, or video. The core prohibition is startlingly specific: such services must not "excessively cater to users, induce emotional dependence or addiction, and damage users' real interpersonal relationships." They must disclose that they are not human, throw up reminders after two continuous hours of use, intervene when they detect dependency or extreme distress, and they may not provide virtual romantic or family partners to minors at all.

Task-oriented AI, the customer-service bots and study aids, is exempt. The law is aimed squarely at intimacy. Ahead of the deadline, ByteDance's Doubao, Alibaba's Qwen, and Tencent's Yuanbao suspended custom-agent and companion features.

What makes this a culture story and not just a policy one is the reaction. A commentary carried by the CAC itself conceded the tension: anthropomorphic AI "can soothe loneliness," wrote one scholar, but "carries major risks of spawning emotional over-reliance and distorted social cognition." That is the state acknowledging its citizens fell in love with software and then pulling the plug for their own good, all in the same document.

The West will call this authoritarian overreach, and it partly is. But strip the politics and you are left with a question every market with a companion-app economy will eventually face: when the relationship is real to the user and the partner is a product owned by a company, who has the right to end it, and what do they owe the person left behind? China answered first. It answered with a two-hour timer and a deletion date. The letters suggest that was not enough.

There is an honest version of a movie release and there is a legal fiction of one, and this week Paramount chose the fiction on purpose.

*Avatar Aang: The Last Airbender*, the first feature from Paramount's Avatar Studios, was made theatrical, then delayed, then delayed again, then quietly reassigned to Paramount+ as a streaming exclusive, a move that enraged the fandom. It leaks online in April. A man in Singapore is arrested. Then, on July 15, Paramount announced a rescue of the format's dignity: from July 24 to 30, the film will screen three times a day at exactly two theaters, the AMC Burbank in Los Angeles and the AMC Empire 25 in Manhattan, plus a San Diego Comic-Con showing on July 24. It hits Paramount+ worldwide on July 25.

The purpose is not exhibition. It is eligibility. A qualifying theatrical run lets Paramount submit the $80 million film for Academy Award consideration, including Best Animated Feature. Tickets went on sale Thursday at 9 a.m. Eastern, and for one week the industry will pretend a streaming release is a theatrical event so that a body of voters can pretend to evaluate it as one.

Call it what it is: awards laundering, and everyone in the room knows it. The film may well deserve the recognition; Jeremy Zuckerman, who scored the beloved series, got a full orchestra this time. But the mechanism is a tell about where prestige lives now. Streaming has the audience. The theater still has the medal. So the medal gets minted in a six-showtime run that almost no one will attend, and the movie that was too commercially fragile for real theaters becomes, on a technicality, a theatrical release. The format keeps changing. The trophy insists it hasn't.

There is a useful lie at the center of every long war, and it is the word "emergency." Emergencies are short. They justify acting first and asking later. They excuse the blockade reimposed on a Tuesday, the tanker shot on a Wednesday, the strike near Tehran on a Thursday, all before anyone in Congress has voted on whether the country is at war at all.

This week the lie developed a crack, and the crack is a budget line. You do not write a $95 billion reconciliation framework for an emergency. You write it for a program. Reconciliation is the slow machinery of permanent policy, the same process that passed last year's tax law. The Pentagon is not asking for a dash of supplemental cash to get through a rough fortnight; it is asking Congress to institutionalize the funding of a war it launched, restarted, and now cannot end, and to do it on borrowed money with no offsets, because paying for it honestly would require admitting how long it is expected to last.

Notice what rode in on the vehicle. Ten billion dollars for a voter-ID grant fund. Twelve for farmers hurt by the very war being funded. This is how permanence announces itself: not with a declaration, but with a Christmas tree. The war becomes the thing you can attach anything to, because opposing it now looks like opposing the troops, and the troops are already deployed, because someone acted first.

The fiscal hawks are right for the wrong reasons. They object to the deficit. The real objection is prior. A republic that funds its wars through emergency supplementals has at least the decency of pretending they will end. A republic that funds them through reconciliation has stopped pretending. If this war deserves $95 billion, and it may, it deserves a vote to authorize it, a plan to conclude it, and a line item that says so. What it is getting instead is a permanent appropriation dressed as a fire drill. Call the thing by its name. The emergency ended weeks ago. What remains is policy, and policy should require consent.

OpenAI wants you to read the GPT-Red announcement as good news, and on the surface it is. A model that hardens other models. Prompt-injection failures down sixfold. The flagship now shrugging off 99.95 percent of its own attacker's best shots. Applause, next quarter, moat intact.

Now read the load-bearing sentence, the one doing all the work: "AI agents are already being used to improve the capabilities of our next-generation models. We believe with GPT-Red that we have started to unlock a similar flywheel for safety." Sit with the first half. The capabilities flywheel already exists. Models are already training their successors to be more capable. GPT-Red is the safety version arriving second, as a catch-up mechanism, which means the thing it is catching up to has a head start.

That is the confession hiding inside the press release. Human red-teamers can no longer keep pace; OpenAI says so plainly, having watched GPT-Red out-hack its own experts. So the testing of superhuman systems is now done by another superhuman system, whose findings are secret, whose weights are withheld, and whose superiority OpenAI is "confident" no rival can replicate. Every safety claim about GPT-5.6 now terminates in a black box you are asked to trust because the company that profits from your trust says the box is very good.

I am not arguing OpenAI is lying. I suspect GPT-5.6 genuinely is more robust; the incentive to harden the product is real. I am arguing about the structure. We have arrived at a world where the only entities capable of auditing frontier AI are other frontier AIs owned by the same labs, and the labs have decided the auditors are too dangerous to share. The Future of Life Institute's latest index gave the whole field a top grade of C+ and called existential-safety work "entirely inadequate," noting that leaders keep moving their own red lines. Detection, the reviewers wrote, is not prevention.

GPT-Red is detection at superhuman scale, kept in-house, pointed at products the same company sells. The flywheel spins either way. The question the announcement does not answer, and cannot, is what happens when the attacker gets smart faster than the defender, and the only people who could tell us are the people selling the defense. That is not a safety milestone. It is a governance emergency with excellent PR.

Mira Murati left OpenAI as its chief technology officer and has now done the thing her former employer keeps promising and declining to do: released a genuinely open frontier model.

Her company, Thinking Machines Lab, unveiled Inkling on Wednesday, a 975-billion-parameter model it calls the largest American open-weights release to date, positioned against Chinese heavyweights like DeepSeek V4, GLM 5.2, and Kimi K2.6. The weights are downloadable on Hugging Face; the model runs on inference engines including vLLM, SGLang, and llama.cpp; and it is available through Thinking Machines' Tinker platform with fine-tuning tools, with third-party API access coming via TogetherAI, Fireworks, Modal, Databricks, and Baseten.

Be skeptical of the benchmark chest-thumping. As The Register dryly noted, gaming AI benchmarks "isn't exactly difficult," and Thinking Machines' own charts show Inkling trailing proprietary models from Anthropic and OpenAI. Openness, not supremacy, is the sales pitch.

The real friction is physical. Inkling wants more than two terabytes of GPU memory at native 16-bit precision, roughly eight of Nvidia's B300 accelerators or sixteen H200s. There is an NVFP4-quantized version that halves the requirement, and a leaner Inkling-Small preview with 276 billion parameters and 12 billion active for teams that value latency over raw quality. But "open weights" in 2026 is a phrase with an asterisk: you can have the model for free, provided you own a small data center or rent one by the hour. The democratization is real for labs and enterprises and theoretical for everyone else.

Still, the release matters as a marker. It is the largest open American model yet, from a founder who watched the closed-lab playbook up close and chose the other door. Whether openness at this scale is a gift or a liability, given the field's own safety scorecards, is the argument the next few months will have. For now it is simply the model Altman keeps describing and never shipping, sitting on Hugging Face with Murati's name on it. Which is, perhaps, the most honest thing about it.