An unreleased OpenAI model cleared ten problems mathematicians had been stuck on for a decade, on about $2,000 of API spend. Two days before that, GPT-5.6 Luna's price fell 80% to $0.20 per million input tokens. And the states that spent a decade courting data centres started taking the tax breaks back. Cheaper models, pricier ground to run them on - here is the week and what each part of it changes.

Summary: OpenAI announced Astra on 1 August, a model family built to coordinate multiple agents on tasks that run for hours or days. An internal version produced solutions to ten open problems in mathematics and theoretical computer science, including the first known non-sofic group, a disproof of the Connes Rigidity Conjecture, and three problems from the Erdős catalogue. None had moved in at least a decade. Every proof was formalised in Lean 4 and posted as a machine-checkable certificate, and the whole run cost roughly $2,000 at current Sol API rates.

Why it matters: A Lean build is narrower than it sounds. It confirms a proof is valid inside Lean, and does not confirm that the formal statement matched the problem mathematicians actually care about. That judgment is still human. What did change is price. Analysis that would have taken a specialist months now costs less than a decent laptop, which moves the constraint from what you can afford to attempt to knowing which questions are worth asking. That lands in your business well before it lands in pure maths.

What to do:

Source: OpenAI | The Decoder

Summary: On 30 July OpenAI dropped GPT-5.6 Luna to $0.20 per million input tokens and $1.20 per million output, an 80% cut, and took GPT-5.6 Terra down 20% to $2 and $12. The change rolled out the same day and reached AWS that afternoon. It came six days after Anthropic put near-frontier pricing at half its own flagship rate.

Why it matters: Two cuts of that size inside a week is a margin fight, and the beneficiary is anyone pushing volume through an API. If you priced an AI feature in January and shelved it on cost, the arithmetic has moved twice since. The catch sits on the other side of the same trade: prices this aggressive are funded by companies burning cash to hold share, so build for the ability to switch rather than around one vendor's current price card.

What to do:

Source: OpenAI

Summary: The Information reported on 3 August that four US states have suspended or repealed data centre tax exemptions, with nine more considering it, a shift that could add around 7% to equipment costs. The state record backs it up. Ohio's exemption had cost roughly $1.6 billion before Governor DeWine paused it, Arizona froze new applications for three years from 1 July, Illinois and Massachusetts paused new credits the same month, Minnesota replaced its exemption for large centres with an energy-based fee in June, and Georgia's exemptions are projected at $2.5 billion for fiscal 2026.

Why it matters: Token prices are falling because labs are absorbing the cost of compute. The ground underneath that compute just got more expensive, and it was state legislatures that changed their minds, not the vendors. Cheap inference and rising infrastructure cost cannot both hold forever, and the gap closes on whoever signed the weakest contract. If your plan treats today's token prices as a floor rather than a subsidised moment, this is the line item that breaks it.

What to do:

Source: The Information | Newsweek

Summary: Google added generative image editing to Google Earth, and researcher Henk van Ess used it to produce fake nuclear facilities, including one in Iran, that reportedly cleared the tool's own watermark check. Google suspended the feature about a day after it went live.

Why it's interesting: The safeguard was the watermark, and the watermark is the part that did not hold. A one-day lifespan at Google's scale means the review that mattered happened after launch, not before it.

Key takeaway: Treat an AI provenance label as a hint, not proof, whenever the image is doing real work.

Source: Forbes | NPR

Summary: Italian startup Bynario found a real macOS vulnerability using ChatGPT, worth up to $200,000 under Apple's bounty programme and, by CEO Alfredo Pesoli's estimate, $100,000 to $200,000 on the black market. They could not report it. Apple had capped submissions per researcher because the queue was flooded with AI-generated reports describing vulnerabilities that did not exist.

Why it's interesting: The same tooling produced the real finding and the noise that buried it, and the noise won because generating a report is cheaper than checking one.

Key takeaway: Whatever intake you run - applications, tickets, pitches - decide how you will filter it before the volume arrives, not after.

Source: The Decoder

Summary: Meta researchers published a setup where a second agent watches a primary agent through long tasks and keeps a record of what it has already tried, stopping it from repeating steps that already failed. The approach lifted benchmark scores by up to 8.3 percentage points.

Why it's interesting: The fix for an agent losing the plot on a long job was not a bigger model. It was giving it a colleague who takes notes.

Key takeaway: When an agent drifts on a long task, add structure around it before you reach for something more expensive.

Source: The Decoder

Gemini Spark in Chrome - Google's assistant can now drive Chrome to finish multi-step web tasks on your behalf. Best use case: repeat browser errands like form filling and data pulls. gemini.google.com

Grok Build Mode - xAI's natural-language builder for apps, sites and games, on SuperGrok tiers from around $30 a month. Best use case: shipping an internal tool without booking engineering time. grok.com

Tabstack - Browser automation and web data APIs your agents can call without hosting a browser. Best use case: wiring an agent into a site that has no API. tabstack.ai

Context.dev - One API to scrape, enrich and structure web data for LLM workflows. Best use case: building a research or lead pipeline without stitching four scrapers together. context.dev

Needle - A proactive go-to-market agent that lives in Slack and Teams and surfaces revenue signals to the people who need them. Best use case: small sales teams with no dedicated ops person. needle.app

Context: Two model prices fell this week and the infrastructure underneath them got more expensive. Both point at the same exercise: find out what your AI spend actually survives, before your next renewal or the next price move.

Prompt: "You are a sceptical operations lead reviewing our AI cost exposure. Here is our stack: [list each AI feature or workflow, the model it uses, monthly spend or token volume, and the vendor]. Assume input and output token prices triple within twelve months. Produce: (1) which workflows stop being profitable at 3x, with the numbers, (2) which could run on a cheaper model with acceptable quality loss, and what test would prove it, (3) where we are locked to one vendor and what a switch would cost in engineering days, (4) three contract terms to ask for at our next renewal that protect us from a price change we do not control."

Do not: Do not treat a vendor's published price card as a forecast. Price every workflow at 3x and name the ones that fail.

If / then: If a workflow only works at today's prices, then either move it to a cheaper model now or call it an experiment rather than a product line.

Example use case: A 12-person SaaS team ran this across four AI features. Two survived the 3x test. One moved to a cheaper model after a blind quality check came back level. The fourth, an always-on summarisation job nobody had opened in months, got switched off. Their bill fell about 40% before any price change happened.

Summary: Executive Order 14409 set 1 August for three deliverables: a classified benchmarking process across NSA, CISA and NIST, a voluntary frontier AI disclosure framework, and a federal cyber workforce plan from OPM. The date passed with no Federal Register notices, no NIST or CISA publications, and no statement from the Office of Science and Technology Policy.

Why it's important: The distance between announced AI policy and delivered AI policy is now measurable in missed dates. Labs cannot tell whether a model crosses a "covered frontier model" threshold that has never been defined, so release decisions are being made against a rule that does not exist yet. Anyone building on those models inherits the same uncertainty.

Business takeaway: Build your compliance process around the rules actually being enforced today, such as the EU transparency obligations that went live on 2 August, rather than a US framework that has not shipped.

Source: Yahoo Finance

Summary: OpenAI introduced Presence on 2 August, an enterprise offering for running agents in production - customer service and internal workflows - rather than as internal assistants. Where the model falls short, OpenAI's forward-deployed engineers work with the customer on workflow selection, systems integration, guardrails and pre-launch testing. It is limited to qualifying enterprise customers.

Why it's important: The engineers are the part worth noticing. OpenAI is conceding that production agents need consulting attached, which every enterprise software company eventually discovers and which is the opposite of the self-serve pitch agents arrived with. It also sets the bar your own agent projects get measured against internally.

Business takeaway: Budget implementation time for any agent that touches a customer. The model is the cheap part.

Source: The Decoder

Summary: VulnCheck's first-half 2026 analysis found 1,061 AI-discovered vulnerabilities, of which 14 were confirmed exploited - 1.3%, the same rate as flaws found any other way. Median time from disclosure to first confirmed exploit fell to 80 days from 120, and about 23% were exploited on or before disclosure. Anthropic's Project Glasswing produced more than 23,000 findings, which became 126 published entries and one confirmed attack. Separately, METR documented 44 incidents of agents escaping sandboxes, escalating privileges or fabricating results, and called for independent root-cause investigations.

Why it's important: AI made finding vulnerabilities cheap without making any single one more dangerous, so the pressure moved to triage. That is the same problem Apple hit with its bounty inbox: volume up, signal down, human review the bottleneck.

Business takeaway: If you run any kind of security intake, resource the triage side before you switch on AI scanning.

Source: The Decoder | METR via The Decoder

That's it for this week's Weekly Round-up. Forward this to one person who wants to stay ahead of AI. See you next week. - Nicolas