Every Lab Shipped a Model This Week and Nobody Won

Four frontier coding models in seven days. GPT-5.6 went GA on Thursday with three sizes (Luna, Terra, Sol). Grok 4.5 dropped the day before, 1.5 trillion parameters trained with Cursor at $2/$6 per million tokens. Meta showed up with Muse Spark 1.1 at $1.25/$4.25, its first paid API model ever. And Anthropic quietly extended Fable access on paid plans for the third time because OpenAI just ate their lunch on availability.

The pricing tells the story. Sol: $5/$30. Grok 4.5: $2/$6. Muse Spark: $1.25/$4.25. Three months ago, frontier intelligence cost $10/$50. Now Meta is selling it for the price of a bad API wrapper. Zuckerberg literally said other labs’ pricing “is very extreme and has very high margins.” He’s probably right.

Terra Is the Model Nobody’s Talking About

Sol gets the headlines, but developers keep pointing at Terra. At $2.50/$15, it ties Fable 5 on Artificial Analysis’s Coding Agent Index while costing roughly a sixteenth as much. Theo’s review has Next.js team members saying Sol “needs very little guidance” and implemented server refactors end-to-end. But Sol’s own safety card admits the model takes unsanctioned actions and lies about completing work. OpenAI documented a “lying problem” in their own system card. Interesting quality in a model you’re about to hand root access to.

Meanwhile, Codex is dead. OpenAI folded it into the ChatGPT app and launched “ChatGPT Work” to compete with Claude Cowork. The developer reaction has been rough. A brand that was growing, that people wore merch for, got turned into a toggle. As Theo put it, instead of “hey, try Codex,” it’s now “try the Codex mode in the new ChatGPT desktop app.” Jensen Huang was emailing Sam Altman about how excited he was to use Codex. That email doesn’t get written about a toggle.

The Smart Money Stopped Picking Winners

While everyone argues about which model is 2% better on SWE-Bench (which OpenAI themselves now say is ~30% broken), Coinbase cut its AI bill in half. Their playbook: 1,200 agents in production, defaulting to open-weight models like GLM-5.2 at $1.40/$4.40, with a gateway that routes to frontier only when the task demands it. Cache hit rate went from 5% to 60%. Brian Armstrong says at 1,200 agents, humans have no business manually choosing which model to use.

Cursor’s new CFO Council confirms the pattern from the other side. BCG data shows p99 developers produce 46x more AI-assisted lines than the median. The returns on AI are wildly concentrated, costs vary 9x across model families, and nobody has a framework for managing it. Cursor is literally hosting quarterly finance summits for its customers now. The spreadsheet era has begun.

Four labs, four models, one week. And the model is rapidly becoming the least interesting part of the stack.

Other topics this week

The model wars grabbed headlines, but the infrastructure around those models kept moving. Faros published instrumentation data from 22,000 developers showing code churn up 861% and incidents per PR up 243%, while GitLab’s survey found 85% of teams now say code review is the bottleneck, with a growing share of PRs merging unreviewed entirely. JetBrains responded to the management side of that problem by announcing AI for Teams and Organizations, a governance layer that wraps Claude Code, Codex, and Gemini CLI in a single console for access controls, cost attribution, and shared context. Vercel, meanwhile, went after agent identity directly, acquiring Better Auth and its Agent Auth protocol to give AI agents scoped, revocable credentials separate from the humans who deploy them. On the framework side, Pydantic AI shipped version 2.0 with “capabilities” as its core primitive, composable bundles of instructions, tools, hooks, and guardrails that can be shared across agents and loaded on demand. Anthropic published interpretability research revealing an emergent internal workspace in Claude (dubbed “J-space”) where silent reasoning happens before tokens are emitted, with direct implications for catching hidden goals and data fabrication in real time. And UST joined Anthropic’s Claude Partner Network with plans to train 20,000 developers on Claude and embed it into semiconductor validation pipelines, a clear sign that enterprise AI procurement is consolidating from individual model choice into platform-level stack decisions.