September 23, 2026
OpenAI released two new models this week, Sol and Luna, pitching them as cheaper and more reliable versions of the Astra lineage. Anthropic answered on the same day with Opus 5.5, which it says is the strongest model it has tested, also at a lower price than its predecessor. Neither company led with a wild new capability. Both led with cost. That's the tell. When two labs spending billions on training start competing on price instead of pure benchmark bragging rights, it means the gap between 'best available model' and 'good enough model' has narrowed to the point where it's not the deciding factor for most buyers anymore.
For a business evaluating AI tools, this is good news dressed up as a rivalry. Whichever model a vendor plugs in behind the scenes, you're increasingly paying less per token for comparable output than you were even a quarter ago. We wrote about this dynamic in more depth in our take on the GPT-6 and Opus 5.5 price war, and the pattern is consistent: raw model access is becoming a commodity, which means the actual value in AI products is shifting to what's built around the model -- the workflows, the data, the integrations, the guardrails. That's a healthier place for buyers to be than an arms race where every quarter you're locked into whoever has the single smartest model.
If model quality is converging, what's left to fight over? Data. Snorkel AI just tripled its valuation to $3.5 billion on a $350 million Series E, on the strength of demand for AI training data. That's a striking number for a seven-year-old company, and it tells you where the smart money thinks the next bottleneck sits. Training a frontier model used to be the hard, expensive part. Increasingly, the hard part is getting clean, well-labeled, domain-specific data to fine-tune and evaluate models for a specific task -- legal review, medical coding, customer support tone, whatever the buyer actually needs.
I think this is the more important story of the day, even though it will get less attention than the model releases. It confirms something we've argued before: the companies that win with AI aren't necessarily the ones with the fanciest model, they're the ones with the best data about their own operations. That's precisely the argument for building on a platform that already holds your business's structured data -- your CRM records, your support tickets, your project history -- rather than bolting AI onto a pile of exported spreadsheets. A CRM or helpdesk that already knows your customers is worth more paired with a cheap, capable model than an expensive frontier model paired with messy data.
Meta acknowledged that its Muse assistant was 'heavily inspired' by OpenClaw, right down to workspace filenames, while insisting it was built from scratch. That's a strange thing to admit if it's true, and a worse thing to be caught saying if it isn't. Either way, it's a preview of a fight that's coming for every company shipping AI agents: as these tools converge on similar architectures, where's the line between 'inspired by' and 'copied'? We covered the specifics of the Muse situation separately, and it's worth reading alongside this piece because it's really the same underlying story -- as models commoditize, the fight moves to what surrounds them, whether that's proprietary data, workspace design, or, apparently, filenames.
Buried under the product launches was a more candid comment from Greek Prime Minister Kyriakos Mitsotakis, who told TechCrunch that no government is actually ready for what AI is about to do -- that leaders are still fighting yesterday's battle. It's rare to hear a sitting head of government say that plainly instead of reciting an innovation-strategy talking point. I take him at his word, and I'd extend the point to plenty of boardrooms too: most organizations, public and private, are still writing AI policy for the models of a year ago, not the cheaper, faster, more capable ones landing this week.
If model prices keep falling and data keeps getting more valuable, where's your business actually building its edge -- on the model you pick, or on the data you already have sitting in spreadsheets and disconnected tools?
Sources