The cost of AI is dropping fast enough that "which model should I use" has become a real engineering problem, not a rhetorical one. Three stories today touch this thread from different angles: a $7B acquisition building the plumbing for model switching, a benchmark showing exactly when cheaper beats premium, and a product team publicly checking off a to-do list. The infrastructure layer is getting its own infrastructure.
01
Model routing is the new must-have for enterprise AI
Latent Space ran a deep piece on Glean, the enterprise search company now hitting $300M in annual recurring revenue after tripling in 15 months. CEO Arvind Jain's pitch: Glean functions as a "superset of ChatGPT, Claude, Gemini, Grok" that routes each query to whichever model handles it best. Crucially, the goal includes *not* using a language model at all when a simpler tool will do. "Sometimes you'll see queries where people are adding two numbers. They could have used a calculator," Jain told Latent Space. Stripe's $7B acquisition of OpenRouter is the same bet at the infrastructure level. ---
Why it matters: If your company is paying frontier model prices for every query, including the ones that don't need frontier models, you're leaving real money on the table. Glean's growth suggests large enterprises are already waking up to this. The next CFO conversation about your AI budget will probably include the phrase "routing strategy."
DeepSeek V4 Pro beats GPT-5.6 Sol on coding, at 1/35th the cost
Together AI ran 904 coding benchmark rollouts comparing DeepSeek V4 Pro 0813 against GPT-5.6 Sol. Sol wins on single-attempt accuracy by 10 points. But Pro wins when you give each model multiple attempts, and a cascade approach that tries Pro first hits 83% accuracy overall. The cost difference: 35x. ---
Why it matters: This is the routing story made concrete. If your coding agent runs thousands of tasks a day, you probably can't afford to send all of them to Sol. The benchmark gives you a practical decision tree: use Pro by default, escalate to Sol only when the task genuinely demands it. That's not a theoretical architecture choice anymore. It has a number attached.
Google's Gemini team posts a public progress report
Josh Woodward, who leads Google Labs, circled back on a prior list of Gemini improvements with a status update. Seven of ten items are done. Workspace tools are being tested in one to two weeks. Gemini 3.7 Flash improved tool calling. A new Projects design is in implementation. Connector count is at 49 and rising. ---
Why it matters: The transparency is worth noting. Gemini has had a credibility problem with developers who felt promises didn't land on schedule. A public checklist doesn't fix that, but it does create accountability. If the Workspace tools don't ship in two weeks, everyone will notice.
Garry Tan shares a quick-start template for AI agent setup
Y Combinator president Garry Tan posted a setup tip: create a new directory, start Claude Code or Codex, paste in a specific image, and your agent repo configures itself automatically. Works with existing subscriptions, slightly better on Desktop. ---
Swyx highlights Trajectory's work on continual learning
Latent Space founder Swyx flagged a talk by Ronak Pradeep of Trajectory at the Continual Learning track, calling it a "very thoughtful overview" of the data problems that still need solving when you want models to keep improving after initial training. The key point: standard reinforcement learning techniques weren't sufficient, pushing the team to develop on-policy approaches and then fix the new problems those introduced.
Why it matters: Continual learning is the unsexy problem that determines whether your fine-tuned model stays useful six months from now or quietly degrades. Teams investing in this infrastructure now will have a meaningful advantage over teams that assume their model is "done" after the first training run.