AI Daily Digest September 09, 2026: Meta Scraps 'Tokenmaxxing', OpenAI Embroiled in Navier-Stokes Co-Author Controversy

Good morning, tech enthusiasts! The AI Daily Digest for September 9, 2026 brings an unfiltered look at internal corporate friction, high-stakes academic drama, monumental financial valuations, and emerging cybersecurity threats. Leading today’s coverage, Meta has officially abandoned AI utilization metrics in engineer performance reviews, shutting down a perverse incentive structure known internally as “tokenmaxxing” that wasted millions in compute. Simultaneously, the mathematics and AI communities are grappling with startling allegations from an NYU mathematician who claims an OpenAI researcher threatened his career after he refused to strip an Anthropic-affiliated co-author from a breakthrough paper on the Millennium Prize Navier-Stokes equations. In venture capital news, Cognition - creator of the Devin AI software engineer - has closed a fresh $2 billion funding round at a mind-boggling $48 billion valuation, while France’s Mistral AI secured €3 billion in the largest equity round in European tech history. Finally, a pressing warning for paid Claude subscribers: infostealer malware is actively siphoning session keys to mint unauthorized tokens and drain account quotas on the black market. Let us delve into each story!

🛑 Meta Drops AI Usage From Engineer Reviews After “Tokenmaxxing” Backfires

Meta leadership has officially eliminated AI dashboards and token consumption metrics from software engineer performance reviews. According to an internal memo authored by vice presidents Maher Saba and Santosh Janardhan and obtained by The Information, performance appraisals will immediately revert to evaluating foundational engineering outcomes: overall code quality, execution velocity, and the inherent complexity of delivered systems.

The policy shift directly addresses an internal gaming phenomenon dubbed “tokenmaxxing.” When Meta initially mandated AI adoption and tracked engineer usage on public internal leaderboards, employees began artificially inflating their numbers by firing off massive, unnecessary prompts, generating redundant boilerplate code, and running automated query scripts purely to climb the rankings. This unchecked utilization has driven Meta’s internal AI inference expenditures toward billions of dollars in 2026 alone, forcing executives to announce strict departmental budgets and a centralized monitoring dashboard set to launch in 2027. Concurrently, Meta’s newly piloted desktop automation agent Hatch has encountered internal employee resistance, with engineers hesitating to connect personal accounts over unresolved data privacy and governance concerns.

Source: The Decoder

📐 Mathematical Drama: OpenAI Accused of Pressuring Mathematician to Drop Anthropic Co-Author From Navier-Stokes Breakthrough

The mathematics and AI safety communities are reeling from public allegations made by NYU mathematician Tristan Buckmaster against OpenAI. Buckmaster and his collaborator Levent Alpöge had spent months deploying frontier models - including OpenAI Codex running GPT-5.6 Sol and Anthropic’s Claude - to achieve major breakthroughs on the Navier-Stokes existence and smoothness equations, one of the seven historic Clay Millennium Prize Problems carrying a $1 million bounty. Throughout the investigation, the mathematicians regularly uploaded working proofs and scratchpads into Codex sessions.

According to Buckmaster’s statement, after word of their mathematical progress leaked in early September, OpenAI researcher Sébastien Bubeck contacted him repeatedly and demanded that Alpöge be dropped from the paper’s authorship because it was “so annoying” that Alpöge is employed by Anthropic. When Buckmaster refused to compromise academic integrity, Bubeck allegedly countered: “Why would you ruin your career? If you don’t want me to be nice, then I don’t have to be nice.” Shortly thereafter, OpenAI announced its own Navier-Stokes breakthrough, claiming an internal swarm of 10,000 AI agents spent 88 hours and millions of dollars in compute to produce a Lean-formalized proof along an identical mathematical path. While OpenAI insists it did not access private user data, the company conceded it could not rule out that de-identified customer telemetry helped train its models, sparking fierce debate over whether frontier labs are leveraging customer sessions to front-run their own users.

Source: The Decoder

💻 Cognition Hits $48B Valuation: The AI Coding Race Is Far From Winner-Take-All

Cognition, the startup behind the autonomous software engineering agent Devin, has raised $2 billion in a fresh funding round that values the company at an astonishing $48 billion. The fundraise comes just four months after Cognition was valued at $26 billion in May, with blue-chip venture firms Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir leading the capital injection.

The soaring valuation demonstrates that institutional investors reject the notion that AI-assisted software engineering will consolidate into a single monopoly. Cognition reported that its annualized run-rate revenue (ARR) rocketed from $492 million to $900 million over the summer, positioning the company to target between $4 billion and $5 billion by the end of 2026. Strikingly, Cognition’s current revenue multiple exceeds that of rival Cursor prior to Cursor’s $60 billion acquisition by SpaceX in April. With enterprise clients such as Mercedes-Benz, NASA, Goldman Sachs, and Citigroup onboarded, Cognition is currently burning an estimated $800 million in cash this year - driven largely by high-end Nvidia cluster leases. To achieve long-term unit economics, the startup is aggressively training proprietary models built upon open-source architectures to eliminate its reliance on third-party APIs from OpenAI and Anthropic.

Source: TechCrunch

🇪🇺 Mistral AI Secures Historic €3B Round: Europe’s Largest Tech Fundraise Defends AI Sovereignty

Three years after its inception, Parisian artificial intelligence champion Mistral AI has closed a €3 billion Series D funding round at a post-money valuation exceeding €21 billion. The transaction stands as the largest equity fundraise ever completed by a European technology enterprise. Samsung Electronics led the syndicate alongside the Scaleup Europe Fund (managed by EQT) and PSG Equity, with participation from BlackRock, the Grand Duchy of Luxembourg, a16z, Nvidia, and ASML.

Mistral’s market valuation has effectively doubled since ASML’s €1.3 billion strategic investment in September 2025. Although Mistral’s latest open models like Mistral Medium 3.5 trail leading open-weight Chinese alternatives such as Qwen and Kimi, and remain behind frontier closed models from the United States, Mistral has carved out an exceptionally lucrative enterprise stronghold. Driven by European corporate demand for digital sovereignty and reduced dependence on American hyperscalers, Mistral now supports over 125 major institutional clients across 20 countries, including Airbus, ASML, and HSBC. Backed by its new equity capital and an $830 million credit facility secured in March, the company is building proprietary European data centers to anchor continental AI infrastructure.

Source: The Decoder

🕵️ Hackers Are Siphoning Claude Tokens From Paid Subscribers via Infostealer Malware

A concerning security vulnerability has come to light as subscribers of Anthropic’s premium tiers - including the $200-per-month Claude Max 20x plan - report massive unauthorized token consumption while their systems are entirely dormant. In one documented incident, an independent UK consultant discovered his token allocation jumped from 45% to 55% during a strict control window where all local Claude Code commands, scheduled tasks, and cloud executions had been completely paused.

Following internal inquiries, Anthropic acknowledged that cybercriminals are utilizing widespread infostealer malware to compromise local workstations. After stealing active Claude browser session cookies, adversaries generate unauthorized server-side Claude Code OAuth tokens, effectively tunneling the victim’s paid compute allowance into illicit third-party proxy networks. The vulnerability has provoked widespread frustration across developer channels because Anthropic does not offer itemized usage logs or granular token auditing. Without visibility into which requests or IP addresses are exhausting quota limits, subscribers remain unable to detect compromises until accounts are drained or preemptively frozen by fraud algorithms, prompting multiple developers to migrate toward open-source multi-model tooling.

Source: TechCrunch