The harness, the watermark and the billion

On 12 August Jeremy Berman published the code of a harness that has an agent play the 25 games of the ARC-AGI-3 benchmark. The repository reports 30.2% for the model on its own, 96.2% with the harness — 24 games won out of 25, 99.3% pass@2 with all 25 won, for roughly $540.

The harness is almost nothing ARC-specific: Claude Code, six pre-approved tools (Bash, Read, Write, Edit, Grep, Glob), no MCP server, no subagents, no plugins. In a single pass the model wrote 269 programs, about 12,700 lines, all discarded when their game ended. The same harness with Codex: 73.7%, and 26,971 actions against 8,789.

↑ Back to the summary

Anthropic confirmed it in a help page published on 11 August: content produced by Claude will carry a machine-readable mark, in application of article 50 of the AI Act. The requirement is European, but the company plans to extend the marking to content produced worldwide.

The watermark is woven into the text itself: it survives a copy-paste into Gmail, Slack or Word and, Anthropic says, “may persist even after modification”. Models launched in the EU from 2 August 2026 carry it from day one; earlier models get a transition period. It applies in Claude, the API, Claude Code, Claude Cowork and Claude Tag, including through AWS, Google Cloud and Microsoft Foundry; PNG, JPEG and SVG files get C2PA provenance metadata instead. How it works technically has not been detailed yet.

↑ Back to the summary

Sundar Pichai announced it on X: more than a billion people use Gemini every month, making it the fourteenth Google product to cross that line. The figure covers the app only — the AI features built into Google Search are not counted.

Les Numériques points out how much of that growth is owed to pre-installation on Android, and notes 100 million users on iOS. ChatGPT keeps the lead, having crossed a billion in July — but in weekly visits. Nothing is said about paying subscribers, while Alphabet is planning between $195bn and $205bn of investment in 2026.

↑ Back to the summary

The AI code-review platform announced on Wednesday a Series C of $143 million at a $1.5 billion valuation — less than a year after a $60 million Series B. The round is co-led by Atomico and Smash Capital, with BMW i Ventures, Datadog and Hirtle Callaghan taking part.

CodeRabbit reports more than 2 million code reviews a week and over 17,000 customers, among them Nvidia, BMW, JFrog, Adyen and Indeed. Atomico's Luca Eisenstecken, who joins the board, told Reuters that as AI becomes critical infrastructure, organisations will increasingly need independent governance layers able to validate software whatever model produced it.

↑ Back to the summary

The Swedish company closed a Series C of $400 million — above the $300 million expected — taking its valuation to $13.3 billion, double the $6.6 billion of December 2025. The round is led by Menlo Ventures, alongside the Scale Europe Fund, Balderton, Carmignac, Tencent, World Innovation Lab, Kaszek Ventures, LTS Growth and Regent; Accel, CapitalG, DST Global, HubSpot Ventures and Salesforce Ventures put in again.

Lovable claims more than 60 million projects created in under two years (25 million in December 2025) and said in June it had passed $500 million of ARR; the Wall Street Journal, quoted by Maddyness, expects it to reach $600 million by the end of the month. Headcount target: 450 by year end. Net income, on the other hand, is not disclosed.

↑ Back to the summary

On 19 August, Merck and Moderna announced that the phase III trial of intismeran autogene + Keytruda met its endpoints — recurrence-free and distant-metastasis-free survival — in patients with completely resected stage IIB-IV melanoma. The trial covers roughly 1,000 patients. It is the last step before a possible market authorisation.

AI's role here is not discovery. Keytruda has been on the market since 2014 and comes out of the work of Allison and Honjo, 2018 Nobel laureates; the personalised cancer vaccine was grounded in 2012 and proven in 2017 by Catherine Wu and Ugur Sahin. AI is what makes it scale: sifting the biopsy's genomic data to pick the 34 mutations most likely to trigger an immune response — possible without it, but at a cost and a delay that would rule out treating anyone this way. What the announcement does not say: by how much the pair beats Keytruda alone.

↑ Back to the summary