Executive Summary
Ownership and distribution moved this week. No new flagship model shipped. Stripe has agreed to buy OpenRouter, the multi-model gateway a lot of small teams already use. NVIDIA has filed residual-value guaranties, capped at $105 billion, behind an OpenAI campus in Ohio. That is a credit backstop, not a $105 billion cheque. SpaceX has closed its acquisition of Cursor. Anthropic told investors its July run-rate was about $65 billion, a figure from an investor briefing rather than audited accounts. In London, the Cabinet Office is studying what happens if Britons lose access to frontier models again.
The tools got more useful and easier to leave unsupervised. Claude Cowork can now send Gmail and manage Drive. GitHub Copilot put Grok 4.6 on the picker. DeepSeek’s cheap flat rates ended on 16 August. GPT-5.6 Sol is half-price on two gateways until 18 September, though OpenAI’s own list price is unchanged. For most businesses the job this week is still a narrow trial: one write-access workflow with approval left on, one priced promo, and a check that the Copilot URL and any Deep Research files survived the 18 August account changes.
What to Try This Week
Put Claude on one Gmail label, with approval left on
If you have a paid Claude plan, open Cowork on web or mobile and connect Google Workspace. Create a throwaway Gmail label, not the client inbox. Ask it to draft and send one defined reply, and leave approval required. If you are a Team or Enterprise owner, leave the approval prompt on this week. Measure what it sent versus what you would have sent, anything it tried to trash or forward, and whether you would leave write access on.
Price one GPT-5.6 Sol job on the gateway discount
If you already route through Vercel AI Gateway or OpenRouter, point one real Sol workflow at the discounted provider. Do not use bring-your-own-key: that still pays full list. Cap the key. Write down promo cost versus OpenAI list. The cut ends 18 September. Measure tokens, dollars, and whether quality changed.
Finish the Copilot cutover and lock the new host
If anyone still uses consumer Copilot, save remaining Deep Research reports in Word and download any podcasts. IT should confirm copilot.cloud.microsoft and *.cloud.microsoft are allowed. If engineers live in JetBrains, update the Copilot plugin so enterprise MCP allowlists apply. Measure blocked hosts, missing reports, and whether Autopilot or Bypass Approvals still works after the plugin update.
Geopolitics, Governance and Big Moves
Stripe agrees to buy OpenRouter
Stripe and OpenRouter both posted on 19 August: Stripe has agreed to acquire the multi-model API gateway. Stripe says OpenRouter routes across more than 400 models from more than 80 providers, and is already used by NVIDIA, Zoom and Lovable. OpenRouter says the name, product and roadmap stay the same, and that nothing about existing integrations should change. Close is expected in the coming weeks, subject to customary conditions. Neither company disclosed a price. The New York Times, via CNBC, heard about $7.5 billion including about $1.5 billion for founders; Reuters heard slightly more than $8 billion. Treat those as unofficial. The 5.5 per cent credit fee, bring-your-own-key terms and data residency were not addressed. If you route production traffic through OpenRouter, inventory the keys and the DPA this week. Dual-source anything you cannot afford to have sitting inside a payments company.
Sources: Stripe newsroom, OpenRouter announcement, CNBC reporting
NVIDIA puts a $105 billion residual-value cap under an OpenAI campus in Ohio
NVIDIA filed an 8-K on 17 August on a multi-year partnership with SoftBank-linked SB Energy at the PORTS-Pike campus in Pike County, Ohio, with OpenAI as tenant on 20-year leases. NVIDIA will be the exclusive AI compute provider. Initial committed IT load is about 4.25 GW; first capacity is expected from 2028. NVIDIA is also investing $1.5 billion in SB Energy. The $105 billion figure in headlines is not an equity cheque. The 8-K says NVIDIA’s aggregate payment obligation under residual-value guaranties on the 4.25 GW leases is cumulatively capped at $105 billion, payable if OpenAI is insolvent or misses rent. OpenAI indemnifies NVIDIA. Obligations start only after ready-for-service conditions, expected from 2028. NVIDIA’s own blog claims much larger lifetime GPU revenue at the site; those figures are company claims, not in the 8-K.
Sources: NVIDIA 8-K, press exhibit, OpenAI on PORTS-Pike
Anthropic’s $65 billion run-rate comes from an investor briefing
CNBC, confirming Bloomberg, reported that Anthropic told investors its annualised revenue run-rate reached $65 billion at the end of July, versus $47 billion in May and about $9–10 billion at the end of 2025. A source also cited preliminary second-quarter revenue of about $11.5 billion. Anthropic declined to comment. CNBC puts OpenAI’s recent run-rate at about $40 billion. The Wall Street Journal put OpenAI’s second-quarter revenue at $6.7 billion and an operating loss of $12.3 billion, and Anthropic’s second-quarter sales at $11.6 billion. None of these figures are audited public results, and run-rate and recognised revenue are not the same measure. Enterprise demand for Claude is still climbing, which matters for buying power. Nothing here changes the product.
Sources: CNBC reporting, TechCrunch reporting
SpaceX has closed the Cursor acquisition
SpaceX’s Form 8-K (event date 14 August) says the merger became effective that day. Cursor survives as a wholly owned SpaceX subsidiary. Outstanding Cursor stock converted into SpaceX Class A shares at an implied Cursor equity value of $60 billion. Cursor’s own post confirms the close and frames Grok 4.6 as an early joint product. Cursor’s claim that it will get “the largest fleet of GPUs in the world” is a company line, not in the 8-K. No customer pricing, data-handling or third-party model change was stated at close. The risk here is concentration and roadmap control.
The UK is studying another frontier-model access shock
The Financial Times reported on 18 August that the Cabinet Office has asked for an assessment of the economic hit if Britons are blocked from new Anthropic and OpenAI models. Work is due within weeks. The trigger was the June episode when a US directive forced Anthropic to pull foreign access to Fable 5, later restored. This is an assessment only. The SME point is simpler: a UK firm on Claude or ChatGPT has already lived a real access shock. Dual-source the workflows you cannot pause.
Source: FT reporting
OpenAI has paused its largest planned frontier training run
On 18 August OpenAI said that after a July Hugging Face containment failure it paused reinforcement learning for two weeks, then restarted many lower-risk runs. The largest planned frontier RL run remains on hold. New monitoring of tool actions, reasoning traces and logs has a 30-minute alert target; if a critical-boundary flag is not cleared, teams are expected to pause the run. OpenAI claims monitoring will cost about 20 per cent of the watched process’s compute. The same post says there is preliminary evidence that Astra may meet the Critical cybersecurity threshold under its Preparedness Framework. That is a company claim; the evals are not published. Customer ChatGPT and API products continue, so there is nothing to change on seats this week.
Sources: OpenAI, TechCrunch reporting
Dynatrace will buy Arize for $915 million
Dynatrace signed a definitive agreement on 13 August to buy Arize, the AI observability and evaluation company, for $915 million (about $815 million cash plus replacement equity awards). Close is expected later this quarter or early in Dynatrace’s third quarter, subject to regulatory review. Dynatrace’s own guidance on ARR accretion and margin dilution is not a fact about completed results. A public company has put $915 million behind eval-through-production as core infrastructure.
Source: Dynatrace
Tools and Releases
Anthropic will watermark future Claude text
Last week’s Pulse said Anthropic had not added invisible watermarks. On 14 August it said future Claude models will generate watermarked text, using a variant of Google DeepMind’s SynthID-Text, to comply with the EU AI Act. The mark is global at launch because Anthropic says it cannot yet scope by region. It carries no user, organisation or chat identity; detection answers only “likelihood Claude was involved”, and is weak on short, factual or code-heavy text. A detection API is coming; it has not shipped. Supported files get C2PA credentials in metadata. Models launched before 2 August 2026 have a legal transition period. Heavy rewrite can strip the mark. The same day, Gemini and Flow made the visible media watermark optional in countries that do not require it; invisible SynthID and C2PA stay on.
Sources: Anthropic, TechCrunch on Gemini
Claude Cowork can send Gmail and manage Drive
Cowork is now on claude.ai and iOS/Android for all paid seats. Enterprise remains admin-gated. Workspace connectors can send, reply and forward Gmail, and share, move or trash Drive files. Approval is required by default; Team and Enterprise owners can allow those actions to run without asking. Data retrieved sits with the chat on Anthropic’s servers. Write access to Gmail and Drive is the change worth reviewing. The Chrome side panel already carried Cowork last week.
Sources: The Verge, Claude Help
GLM-5.3 is on the API; thinking cannot be turned off
Z.ai now lists GLM-5.3 at $1.40 per million input tokens and $4.40 per million output tokens, the same as 5.2. No release date for open weights has been given. Thinking cannot be disabled; reasoning_effort is low, high or max, and max is the default. Port a 5.2 job that used thinking.type=disabled and you will get a 400, then a bigger bill. China-origin plus the lab’s own earlier cyber-eval notes still need a procurement check.
Sources: Z.ai pricing, Z.ai announcement
Cursor wants to host your code, and keep cloud agents on
Origin, an early beta on paid plans from 17 August, adds native repos, pull requests and GitHub sync. Enterprise admins can opt out. Pricing, residency and DPA are not on the changelog. Do not migrate production on day one. Separately, Cloud Agents now default to pre-warmed “builds”, and from 19 August they can subscribe to a pull request, a Slack thread or a schedule, hold a long-lived /goal, and run subagents on their own VMs. No price is on the changelog. Budget for the spend and the extra review load.
Sources: Cursor Origin, Cursor builds, Cursor changelog
Consumer Copilot lost Group Chat, Podcasts and Deep Research
Account updates began on 18 August and are rolling out gradually. Podcasts become inaccessible. Deep Research reports can be opened in Word; Microsoft 365 Premium keeps creating them via Researcher. Group Chat threads become one-to-one with Copilot; other people’s prompts and images do not come across. The work web app moves from m365.cloud.microsoft to copilot.cloud.microsoft. Users redirect unless the new host is blocked. Unified Windows and Mac desktop is preview now, broader from mid-September.
Sources: Microsoft Support, Partner Center
DeepSeek’s cheap flat rates ended on 16 August
Peak and off-peak API prices took effect at 17:00 London on 16 August. Off-peak is half of peak, and both are a rise on the old flat rates. UK morning (07:00–11:00) sits in a peak window. V4-Pro output went from about $0.87 per million tokens to $1.98 off-peak and $3.96 peak. Data-residency and China-API risk still apply. If you routed agents through DeepSeek to save money, rebuild the cost model.
Sources: DeepSeek pricing, 13 August note
GPT-5.6 Sol is 50 per cent off on two gateways only
Vercel AI Gateway cut Sol in half through 18 September: default $2.50 / $15 per million tokens (was $5 / $30). The cut covers requests billed through Gateway’s OpenAI provider. Bring-your-own-key still pays full price. OpenRouter is running a matching cut. OpenAI’s own API list is unchanged. Do not bake the promo into next quarter’s budget.
Sources: Vercel changelog, OpenRouter Sol
Grok 4.6 is in GitHub Copilot and generally available on Bedrock
Copilot Pro through Enterprise are getting Grok 4.6 across VS Code, Visual Studio, Copilot CLI, JetBrains, Xcode and Eclipse. Business and Enterprise: the policy is off by default. Billed at provider list rates. Separately, Grok 4.6 is now generally available on Amazon Bedrock at xAI’s list of $2 / $6 per million tokens. Capability is unchanged. What is new is where you can buy it.
Sources: GitHub changelog, xAI on Bedrock, AWS What’s New
Claude Code auto mode is the default on Pro, Max and Team
New sessions start in auto mode unless someone already pinned another default. A classifier auto-approves tool calls and is meant to block irreversible actions; after repeated blocks it falls back to manual. Enterprise, API, Bedrock and Foundry stay opt-in. Anthropic’s own tester study: humans caught 13.6 per cent of planted dangerous commands, auto mode 89 per cent. Still review high-stakes production changes yourself. /design is a separate research preview for paid seats: generate a few artboards, then implement. The Decoder reports it is token-heavy.
Sources: Anthropic on auto mode, The Decoder on /design
Anthropic took Files, Skills and Enterprise admin to general availability
The Files API no longer needs a beta header (1 TB per organisation, 500 requests per minute). Agent Skills are generally available. The Admin API for Claude Enterprise members, invites, groups and custom roles is generally available. Managed Agents can set allowed and blocked domains on web tools. Team (non-Enterprise) admin coverage is not stated.
Source: Claude platform release notes
ChatGPT ads reach 31 European markets on 24 August
From 24 August, ads appear on Free and Go in Germany, France, Spain, Italy, the Nordics, Netherlands, Ireland, Poland and others. Plus, Pro and Enterprise stay ad-free. GDPR: opt-in for personalised ads; contextual is the default. The UK has had ads since 6 June. ChatGPT Computer History, the Mac click-and-keystroke timeline, is still not available in the UK, EEA or Switzerland.
Sources: Digiday, OpenAI Computer History docs
Qwen released Qwen3.8-27B weights under Apache 2.0
The 27 billion parameter multimodal model is on Hugging Face and ModelScope, with a native 262,144-token context. Qwen’s own benchmark claims are vendor numbers. The much larger 2.4T Max-class release uses a different commercial licence: MaaS or AI-assistant businesses above $50 million trailing-12-month revenue need a separate licence. The 27B is the version most teams can run locally.
Sources: The Decoder, Qwen3.8-27B
Quick Hits
Higgsfield raises $400 million at $5.4 billion: DST Global led a Series B for the video and image platform. The company claims $700 million annualised revenue; treat that as a company figure.
Groq raises $350 million at $3.5 billion: Planned NVIDIA participation. The valuation is down from $6.9 billion last September, after NVIDIA’s licensing and talent deal. Groq is selling inference cloud on NVIDIA GPUs, not a new chip story.
Linear: agents now author about half of issues. In paid workspaces, pull requests per workspace are up 111 per cent versus June 2024. Time on the old tasks did not fall.
Cerebras launched CS-4, a three-wafer rack. First shipments this quarter. Speed claims are vendor-reported.
Microsoft patched CoSnitch (CVE-2026-24301) in Copilot Personal: a one-click link could pull connected Gmail, Drive or OneDrive. Cloud-side fix; still disconnect unused connectors.