On February 17, 2026, Anthropic released Claude Sonnet 4.6 and made it the default model for free and Pro users outright. CNBC’s coverage also noted the version’s knowledge cutoff: May 2025.
It arrives twelve days after Claude Opus 4.6 (February 5) — Anthropic refreshed both its flagship and its workhorse within two weeks.
The Default Slot Is the Real Battlefield
Most users never switch models. Whatever sits at the default is what Claude means to them. Putting the newest Sonnet straight into that slot upgrades the entire free and Pro population at once — the experience, the word of mouth, and the usage data all change generations together.
It is also the most expensive slot in a model launch. The free tier is not getting a downgraded older version; it is getting the same new release as Pro subscribers. That signal is itself an acquisition strategy: the lowest-friction entrance gets the newest capability.
Default placement is also the fastest feedback loop in the industry. Every free-tier conversation becomes same-day telemetry on the new model’s behavior across every domain users bring to it. A flagship launch tells you what a lab believes; a default swap tells you what it is willing to let every user feel immediately.
A Flagship and a Workhorse, Divided Cleanly
February’s two releases assemble a clear product matrix. Opus 4.6 takes the flagship lane — agent teams, a 1M-token context window, Claude in PowerPoint — aimed at heavy and enterprise use. Sonnet 4.6 is the daily workhorse carrying most conversations and coding workloads.
For developers, that layering simplifies selection: heavy agent workflows go to Opus, everyday tasks go to Sonnet. The real decision variable shifts from “which one is stronger” to “how to structure costs” — which is what a maturing model market looks like.
It is also a load-bearing week for Anthropic’s operations: swapping the default model for all free and Pro users in one move means capacity, rate limits, and cost curves all reset at once. Doing it twelve days after a flagship release is a statement about confidence in the pipeline.
A May 2025 Knowledge Cutoff in Practice
The knowledge cutoff sits roughly nine months before the release date (February 2026). The world the model “knows” ends in mid-2025 — later events, package versions, and API changes are simply not in its training data.
The practical implications:
- Time-sensitive work (new library versions, recent events) needs retrieval or grounding tools on top of the model, not the model alone
- When evaluating performance, separate “doesn’t know” from “does it poorly” before judging capability
- Teams migrating to Sonnet 4.6 should include knowledge-dense scenarios in regression testing
The cutoff also argues for pairing the model with fresh sources by default rather than by exception — treating retrieval as part of the product, not a fallback. None of this is unique to Sonnet 4.6 — every model ships with a cutoff — but a default model raises the stakes, because time-sensitive questions will now be answered at population scale by a model whose world ends in May 2025 unless something fresher is wired in.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.