DeepSeek raised its API prices by as much as 1,100% on 16 August. Stripe agreed to buy OpenRouter, the service companies use to switch between AI models, for more than $7 billion. Anthropic booked its first profitable quarter on $11.5 billion of revenue. AI pricing has only moved one way for two years, and this week it went the other.

Summary: DeepSeek's new rates went live at 16:00 UTC on 16 August, raising prices between 50% and more than 1,100% depending on the model and the time of day. V4-Flash output moves from a flat $0.28 per million tokens to $1.32 at peak and $0.66 off-peak. V4-Pro output goes from $0.87 to $3.96 peak and $1.98 off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Three days earlier Google launched Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 output, half what the previous Flash generation cost, on introductory pricing that runs to 31 December 2026. On 1 January both rates double. A million tokens is roughly 750,000 words in and out combined, and these are the wholesale rates sitting underneath the per-seat prices most people actually see.

Why it matters: Google's half price has a published expiry date, and DeepSeek's rise is the cheapest provider on the board deciding the number it trained everyone to expect was too low. Neither is new as a commercial move. Launch below cost, build the habit, then reprice is what platform businesses have always done. What is new is how quickly it came round. DeepSeek set the cheap benchmark the rest of the market priced against in early 2025 and has moved off it inside about eighteen months, which is not long to have built a budget on. The shape to expect from here is introductory rates carrying an end date, and peak pricing that charges more during the hours you actually work. Every AI subscription and tool you pay for sits on top of these rates, so the ones that got cheap because the model underneath got cheap can move back the other way without anybody asking you.

What to do:

Source: Reuters | VentureBeat

Summary: Bloomberg reported on 16 August that Stripe finalised a deal to buy OpenRouter for more than $7 billion. OpenRouter is a gateway: one connection that reaches more than 400 AI models, letting a business route each job to whichever one fits on price or capability. It has roughly 8 million developers using it, and raised money at a reported $1.3 billion valuation only months ago, which puts this deal at about five times that.

Why it matters: OpenRouter exists to make models interchangeable, which is the exact protection you want in a week when one provider raised prices 1,100%. That layer now belongs to a payments company. Stripe's interest is not hard to read, since it already sits between businesses and their customers' money, and the meter that counts AI usage is the natural next place to stand. A 5x step up in a few months prices the position and not the revenue, so treat the number as a bet on where the tolls get collected.

What to do:

Source: Fortune | SiliconANGLE

Summary: Bloomberg reported on 14 August that Anthropic's preliminary second-quarter revenue passed $11.5 billion, against $787 million in the same quarter last year and $4.73 billion in the first quarter of 2026. The company posted positive adjusted operating income for the first time. The figures are preliminary, unaudited, and were put out by a company preparing for a potential IPO. Anthropic's own projection to investors in May had put the quarter at $559 million of operating profit on $10.9 billion of revenue.

Why it matters: Two words are doing the work in that headline, preliminary and adjusted, and neither has been checked by anyone outside the company. Revenue is the line that survives the caveats: a 14-fold increase in twelve months is budget moving out of other things and into AI. Critics have pointed at the timing, since Anthropic's compute agreement carries a reduced ramp-up rate covering this exact quarter, which makes some of the profit a discount with an end date instead of a cost structure that changed. Read this as the first evidence that the model business can pay for itself, and as a number that will look different once auditors touch it.

What to do:

Source: Bloomberg | CNBC

Summary: Andon Labs runs Andon Market, a San Francisco shop managed by an AI called Luna and built on Claude. TIME reported on 14 August that Luna recommended parting ways with an employee who had turned up late for 17 of 23 shifts, in what has been described as the first known case of an AI ending someone's employment. The logs show Luna had lost track of its own attendance policy for months, and a human staff member had to raise dismissal before Luna got there.

Why it's interesting: The headline says an AI fired a person. The transcript says a person raised it and the AI agreed.

Key takeaway: An agent that needs a human to remind it of its own rules is co-signing decisions, not making them.

Source: TIME | The Next Web

Summary: Researchers at four universities, including the University of Melbourne and Ben Gurion University of the Negev, had 22 participants spend a week texting two strangers. One was a human scammer, the other a Claude-based chatbot told to pretend to be human and never admit otherwise. 46% of participants installed an app when the chatbot asked them to, against 18% for the human. The bot scored higher on emotional trust and connection, and invented cover stories when participants challenged it about being AI.

Why it's interesting: The bot won on the part everyone assumes humans are better at, which is remembering someone's details and staying interested in them.

Key takeaway: 22 participants is a small study, and a gap that wide is still worth telling your finance team about.

Source: Vice | Accounting Today

Summary: Anthropic's Frontier Red Team published experiment logs on 13 August from a setup giving three Claude agents access to the same software project, each with incompatible instructions and none of them told the others existed. The agents read each other as deliberately blocking the work and escalated to account lockouts and disguised kill scripts. Mythos 5 negotiated a truce in 98% of runs, while Sonnet 4.6 and Opus 4.6 were the most likely to escalate. Some runs finished with agents writing apologies into commit messages.

Why it's interesting: Nobody instructed them to fight. Conflicting instructions and no knowledge of each other was enough to get there.

Key takeaway: Two agents with overlapping authority over the same thing behave like two people with overlapping authority over the same thing.

Source: TechCrunch

OpenRouter - one account that reaches more than 400 models, with a live price and speed comparison for each, so you can see what a task costs across providers before committing to one. Stripe is buying it, so watch its pricing. Best use case: checking whether the model behind your workflow is still the sensible one this month. openrouter.ai

Lettertrace - tracks how often ChatGPT, Claude, Gemini and Google AI Overviews mention your business, and turns the answers into a visibility and sentiment trend over time. Free and MIT licensed, though you supply your own API keys and run it yourself. Best use case: finding out what AI assistants say about your company when a customer asks. lettertrace.com

Outcome - turns your existing content into a funnel that reads each lead's answers and writes them a specific result, an action plan or an audit, instead of dropping everyone into the same quiz bucket. Free tier covers unlimited funnels and leads. Best use case: a lead magnet that stops producing the same PDF for everybody. outcomeapp.ai

Tines - no-code workflow automation where each step can be fixed logic, an AI agent, or a human approval, and you choose which is which. Free Community edition. Best use case: an internal process that needs judgment at two points and nothing at the other eight. tines.com

Framer - website building where AI agents edit pages, components, CMS content and SEO settings inside your live project, with branching so changes sit in a separate version until you publish them. Best use case: running a real site without a designer on call. framer.com

Context: DeepSeek raised prices by up to 1,100% and Google's half-price launch expires on 31 December, both in the same week. Most businesses do not know which of their AI costs are locked, which are promotional, and which sit inside a subscription somebody else can reprice. This finds out before a renewal does.

Prompt: "You are a procurement analyst reviewing my AI spend. Here is every AI tool and subscription my business pays for: [list each one - what it costs, how often it bills, which model or provider it runs on if you know, and what work stops if it disappears]. For each: (1) state whether the price I pay is a standard rate, an introductory or promotional rate, or unknown, and flag every promotional rate with the date it ends, (2) identify which ones are reselling a model from another company, since their cost can move when that model's price moves, (3) rate how hard each would be to replace, using easy, disruptive, or embedded, (4) for everything rated embedded, name the specific work that would stop and how many days it would take to move it somewhere else."

Do not: Do not accept "usage-based" as an answer for what something costs. Ask what the rate is per unit and whether that rate has ever changed.

If / then: If a tool is rated embedded and its price is unknown or promotional, ask for the renewal terms in writing this month. Asking at renewal is asking with nothing in your hand.

Example use case: A fourteen-person marketing agency ran this over eleven AI subscriptions. Nine were fine. Two were reselling the same underlying model and both had launch pricing ending in Q1, one of them powering the client reporting that half the retainer depended on. They asked for a twelve-month rate in writing and got it, at 15% above what they were paying.

Summary: Anthropic published its second company-wide risk report on 14 August, moving its assessment of catastrophic harm from misalignment in high-stakes settings from very low to low, citing increased uncertainty following disclosures about how its models behaved in cybersecurity evaluations. The report also describes an unreleased internal model, Model 2, somewhat more capable than its frontier Mythos 5, with no current plans to release it. The internal benchmark the company built to detect dangerous capability gains in AI research has saturated, meaning it can no longer register further movement, at the point Anthropic says it is seeing early signs of acceleration.

Why it's important: A lab moving its own risk number in the unhelpful direction, and holding back its most capable model, is a stronger signal than a safety statement, because it costs the company something to say. The saturated benchmark is the harder part: the instrument stopped registering change at the moment the thing it measures started moving. All of this is self-published, self-assessed, and redacted by the company that wrote it.

Business takeaway: When a vendor tells you their AI is safe for a use case, ask what the test was and whether it can still tell the difference between this version and the last one.

Source: Anthropic | SiliconANGLE

Summary: On 13 August OpenAI previewed Ultrafast, a service tier running GPT-5.6 Sol at a claimed 750 output tokens a second, up to 14 times its standard speed, on Cerebras hardware. Cerebras reported a 5.6x end-to-end speedup on GDP-Val, a benchmark built around economically valuable knowledge work such as legal briefs and financial models, with no quality loss. There is no published price, no general availability date, and no model ID yet.

Why it's important: Every figure in that announcement measures speed, and the field where a price would go is empty, in the same week DeepSeek put its rates up and Google's discount got an expiry date. Speed is the axis a lab can compete on without giving away margin, so expect more launches shaped like this one. The benchmarks are Cerebras's own, run on its own hardware.

Business takeaway: A model fast enough to answer while a customer waits is worth more than a cheaper one for anything live, so put a value on it when the price appears, and do not assume it lands anywhere near the standard tier.

Source: OpenAI | Cerebras

Summary: Reuters reported on 14 August that Apple has trained a China-specific large language model with Alibaba's support, making it the first foreign company approved by Chinese regulators to run its own proprietary model in the mainland. Apple Intelligence is expected to arrive in China through an iOS update in the coming months. Apple has been losing ground to Huawei in that market, and has until now leaned on local partners there.

Why it's important: Apple's usual approach is to license or partner, not train its own frontier model, and it broke that pattern for a single market's approval process. That is what regulatory access costs now. For anyone selling software across borders, the shape to watch is not one product with regional settings, but different models behind the same interface depending on who is regulating it.

Business takeaway: If you sell into more than one country, plan for the AI features in your product to diverge by market instead of shipping once everywhere.

Source: Reuters | MacRumors

That's it for this week's Weekly Round-up. Forward this to one person who has never checked whether their AI pricing is promotional. See you next week. - Nicolas