When something is given away, there is usually a reason

Highlight

When something is given away, there is usually a reason

Over the past fortnight, three of the large providers have given something away. OpenAI has cut the list price of its flagship model GPT-5.6 Sol by, in its own words, more than 20 per cent, to 4 US dollars per million input tokens and 20 US dollars per million output tokens. Anthropic has extended, once again, the 50 per cent increase to the weekly Claude Code limit that has been running since early summer. And Google’s introductory price for Gemini 3.7 Flash, familiar from issue 12, stands at 0.75 and 3.75 US dollars before rising to 1.50 and 7.50 at the turn of the year. It looks like a price war, and the obvious explanation would be that demand is softening.

That explanation does not hold up, although everything that speaks against it comes from the providers themselves, two of them ahead of a stock market listing. Anthropic justifies its own restraint with the opposite of a lull: it would like to make the increase permanent, but „strong demand for our models means that capacity may be tight over the coming weeks“. The company’s annualised revenue run rate stood at 65 billion US dollars at the end of July, up from 47 billion in May. For OpenAI, more than 40 billion was reported in mid-August. And on 16 August DeepSeek actually raised its prices. Competitive pressure with falling inference costs is a more likely reading than weak demand. From the outside the two cannot be separated cleanly in any case: revenue is volume times price, and prices have just fallen.

The three moves are worth telling apart, because only one of them is a concession. OpenAI is cutting an existing list price and guarantees it until at least 21 November. Google’s 0.75 is an introductory price that was always meant to rise; the company is stepping back from nothing, and the price ends on 31 December. Anthropic, for its part, is not raising anything but extending, once more, an increase that has been running since early summer; after 31 August the limits return to the plan standard, as its own usage terms state. What the three have in common is the expiry date and nothing else.

For an institution the more interesting figure is a different one. Anyone building a business case on today’s unit costs is building it on a promotion. And while the token becomes cheaper, nothing becomes cheaper about what actually makes deployment expensive in a supervised environment: the outsourcing assessment, the governance, the liability. How quickly that difference becomes visible is shown by a report from Munich further down.

LinkedIn Featured

Sovereignty is not an address: what Mistral’s opening to a Chinese model means

Mistral is opening its platform to a Chinese model, promising inference in Europe while letting Microsoft and Google run the data centres. The article shows that the European sovereignty debate conflates two questions: who built the model, and where the infrastructure sits. Both answers can diverge, and in the case of GLM-5.2 on a European endpoint they do. The piece works out what an institution actually acquires here, and counts the 25 sub-processors Mistral lists in its own trust centre.

Sovereignty is not an address: what Mistral’s opening to a Chinese model means →

Agentic AI

tools, skills & what’s trending

Pricing: three conditions on a clock

OpenAI cuts GPT-5.6 Sol to 4 US dollars input and 20 US dollars output per million tokens at standard context length, guaranteed until at least 21 November. Anthropic extends the 50 per cent increase to the weekly Claude Code limit, for now until 31 August. Google sells Gemini 3.7 Flash at 0.75 and 3.75 US dollars, rising to 1.50 and 7.50 from 1 January 2027. Issue 12 carried a correction on this point: the 50 per cent discount circulating at the time belonged to a reseller. The genuine cut is smaller, comes from the vendor, and is time-limited.

Open weights: the pressure comes from below

On 12 August Alibaba published the weights of Qwen3.8-Max, a model with 2.4 trillion parameters, though under a licence of its own; two days later a 27-billion-parameter model followed under Apache-2.0. Nine days after that DeepSeek showed an experimental multimodal model that, by its own account, brings its agent capabilities close to Opus 4.8. For an institution with data residency requirements the second item is the more interesting one: open weights can be run in-house, and the question of the endpoint then does not arise at all.

Use case: a broker opens its custody account to third-party agents

Since 25 August the Munich broker Scalable Capital has allowed ChatGPT, Claude and Grok to access account functions through a CLI and a Model Context Protocol server: quotes, watchlists, portfolio analyses, savings plans and securities orders that the investor confirms individually before execution. Login and two-factor confirmation remain mandatory, and deposits and withdrawals stay confined to web and app. The terms of use are unambiguous in one place: the third-party applications operate „independently and outside Scalable’s control“, and what they produce does not constitute investment advice and is executed at the investor’s own risk. The firm opens the door and declares in the same breath that it does not own the room behind it. By its own account it is the first bank in Europe to take this step. Who answers for what a third-party agent produces is settled by the terms of use themselves: the investor.

Banking & Regulation

what really counts right now

The ECB asks about the bubble in public

On 17 August five ECB economists published a blog post under the title „The AI boom: rational enthusiasm or the next dot-com bubble?“. They leave the answer open. Their more important observation is that the answer barely matters for the risk. The valuation of the S&P 500, measured by the cyclically adjusted price-to-earnings ratio (CAPE), is close to its historical peak. Euro area households hold around 440 billion euros of exposure to US technology equities, largely through low-cost ETFs and, in the words of the text, „without necessarily being aware of the associated concentration risk“. The European equity market itself is dominated by the old economy and shows little of the boom, which according to the ECB is precisely no protection against contagion. Unlike in 2000, there is little room this time to cushion a correction through monetary or fiscal policy.

Two consultations redraw the perimeter

On 25 August the EBA opened a consultation on three draft regulatory technical standards. They govern how investment firms with total assets above 30 billion euros are reclassified as credit institutions: calculation of total assets under Art. 8a (6) b CRD, reporting obligations under Art. 55 (5) IFR already from 5 billion, and waiver criteria under Art. 8a (7) CRD. The deadline runs until 25 November, with a public hearing scheduled for 30 September. A week earlier ESMA had opened a consultation on draft technical standards for annual reporting of clearing activity at recognised third-country CCPs, Art. 7d EMIR, deadline 12 October. Neither proposal shifts a threshold dramatically, but both decide who will in future be measured under which regime.

Signal & Noise

what your time is worth

• The Fed ends its Deutsche Bank action after nine years – Federal Reserve Board. The cease and desist order dated 20 April 2017 was terminated on 13 August; it covered Deutsche Bank AG in Frankfurt am Main, DB USA Corporation and the New York branch. The notice does not say what the 2017 order concerned. It closes a chapter of supervision that had been open for more than nine years.
• Singapore’s DBS puts credit agents in the hands of 1,500 bankers – DBS. Following a pilot with 150 users, several agents now handle more than 70 tasks per credit memo and deliver a review-ready first draft. Up to 40 per cent of a relationship manager’s time currently goes into that research and documentation work. The 30 per cent saving quoted is a stated goal, not a measured result. The scale is notable nonetheless: this is no longer a pilot.
• OpenAI offers zero data retention for frontier models – OpenAI. For eligible API customers, prompts and responses are not stored after processing and do not feed into training; alongside this comes a mechanism for abuse detection intended to work without staff seeing the content. It does not replace an outsourcing assessment, and so far it is not an externally audited property either. It does remove the objection that most often stalls agentic AI pilots in supervised institutions.
• The SEC proposes a Regulation Crypto Assets – U.S. Securities and Exchange Commission. SEC Chairman Paul S. Atkins wants to give issuers clear paths to raising capital under the federal securities laws; the proposal builds on the interpretive guidance of March 2026. The comment period runs 60 days from publication in the Federal Register. For European institutions the comparison with the ongoing MiCA review is the real reason to read it.
• Deutsche Bank helps shape Google’s Financial Research Agent – Deutsche Bank. After the trade surveillance work this newsletter covered on 17 August, this is the bank’s second Google collaboration in the field: as a design partner it helped shape the Financial Research Agent in Gemini Enterprise for Financial Services and intends to deploy it across its corporate bank. That is no longer a single project.

„A US AI fallout would not remain a US problem.“