Who checked, and what were you not told: Astra falls to 62.7%, Coca-Cola rebuilds its media machine, Spotify filters 50 million tracks

Three stories today, and really one. A benchmark number that did not survive its own referee. A soda giant reshuffling who buys its attention. And a music platform that quietly pulled fifty million tracks from your feed. In every case the question is the same: who checked the claim, and what were you not told?

Signal one: the benchmark and its referee

When OpenAI launched GPT-6 Astra on September 3, the company published a striking score on ARC-AGI-3, a benchmark designed precisely to resist shortcuts: 99.9%. The number traveled fast. Nvidia chief executive Jensen Huang cited it a few days later when declaring that artificial general intelligence had arrived.

This week, the people who run that benchmark answered. The ARC Prize Foundation reran Astra on its standard, provider-neutral harness, the same setup every other model has to pass. The score fell to 62.7%, at maximum reasoning effort. It is still the best result ever recorded on this benchmark. And it is 37 points below the announced figure.

Where the gap comes from

From the harness, the scaffolding around the model during the test.

• With the standard harness, every request starts cold. The model keeps only the notes it chooses to keep.
• With OpenAI's adapter, the model preserves an opaque reasoning state between requests and compresses long conversations. It reuses its work instead of solving everything again.

Changing only the harness moves the score by 36 to 45 points. With the adapter, Astra scores 98.6% at maximum effort and 99.9% at high effort. ARC Prize rules that this score is not valid as a comparable figure.

An analogy makes the stakes clear. Two students sit the same exam. One arrives empty-handed. The other may keep their scratch paper from one exercise to the next. Both grades are real. They do not measure the same thing.

Why it matters

Benchmarks are this industry's currency. They decide which model a company buys, which lab looks ahead, and which stock story investors believe. When the headline number measures the scaffolding as much as the model, every buyer is evaluating something other than what they think.

On October 14 I showed that cost per task contradicted Gemini 4 Argon's per-token price. Same lesson, different measure: the number that counts is the one that survives the same test as everyone else.

What changes

The lesson on whom to believe is clear: independent verification, more than the launch page. The 62.7% score is still a record. It is simply far from the arrival of general intelligence.

The lesson for any brand buying AI this year: ask for the number from the standard test, and ask what the harness added. If the vendor cannot answer, you have your answer.

Signal two: Coca-Cola rebuilds its media machine

Coca-Cola chose its new media partners, and the cola wars have a new front. On October 22, after a global review of what the company calls its marketing operating system, it kept WPP Open X as media partner almost everywhere and handed North America to Omnicom. Dentsu continues in Japan and South Korea. WPP Open X remains global creative partner. Work starts on December 1.

Musical chairs

Publicis held the North American account, but it won PepsiCo's global media business in September and stepped away from Coca-Cola. Omnicom took the seat. Digiday estimates the account at 700 to 800 million dollars. Omnicom is now the world's largest media house after absorbing IPG, with 75.6 billion dollars in billings, about 31% of all billings handled by the major global groups.

Why it matters

The stakes go beyond one account. Coca-Cola chief marketing officer Manolo Arroyo said last year that 70% of the company's media spend now goes to social and creators. Six years earlier, 75% went to television.

The whole shift fits in one sentence. A major brand's media partner has stopped being primarily a buyer of TV time. It has become the operator of an ecosystem of data, loyalty and creators.

What changes

The review is revealing. Coca-Cola changed agencies, and above all redesigned its system: media, data, customer relationship, loyalty and creators in one connected machine, with different partners by region.

For any brand watching, which holding company won matters little. The useful question is whether your own marketing machine is built for a world where attention lives in feeds and with creators, far more than in thirty-second spots. If your partner still sells you TV first, you are buying the past.

I add a strategist's reading. Going from 75% TV to 70% social in six years means the Coca-Cola brand, one of the oldest and most carefully built in the world, is now told largely through voices other than its own. Creators carry the message. Brand consistency then depends on how clear the brand is about what it is, far more than on controlling every word. That lesson applies to companies of every size.

Signal three: Spotify, editor by subtraction

Spotify pulled more than 50 million filler and spam tracks from its recommendations this year. That is the headline of its one-year update on AI-generated music, published on October 20. The detail matters more than the number.

The tracks were not deleted. They are still on the platform and searchable. What changed: Spotify's algorithms and editors no longer surface them. The company says it started by studying songs listeners skipped, ignored or said they did not want to hear. They were consistently tracks its systems had already flagged as filler or spam.

Separately, more than 75 million tracks were removed outright for serious or repeated policy violations. That is a different register, not to be confused with the first.

Why it matters

This is the other side of AI-driven abundance. When anyone can generate thousands of songs at near-zero cost, the catalog explodes and the royalty pie gets diluted. Spotify says tracks likely to be AI-generated are ten times more likely to stay under its 1,000-stream threshold, below which a track earns no royalties. The platform also labels "AI persona" artists and keeps them out of recommendations by default.

What changes

For artists and brands, the filter becomes the product. Being in the catalog guarantees nothing anymore. You have to survive the filter.

Spotify admits its detection tools, which combine metadata, listener feedback, upload patterns and third-party detectors, have limits, and that detection will get harder as generative AI improves. I connect this to the Caine Prize on October 15: there too, a detector judged what was human, with its biases.

AI-generated music exists, at industrial scale. The question that remains is who decides what you hear, and what that decision costs the humans on the other side. A platform that quietly removes fifty million tracks from its recommendations has become an editor, whether it says so or not.

For a brand, the consequence is direct. On every platform saturated with generated content, visibility depends on quality signals: real engagement, consistency, trust. Producing more is no longer enough. You have to produce what the filters recognize as valuable.

The thread of the day

The benchmark checked by its own referee. The soda giant rebuilding its marketing machine. The music platform turned editor by subtraction.

Three reminders that in a world of abundance, value lives in verification and selection. Whoever verifies and whoever filters holds the power.

The noise filtered. The signal kept.