Claude Sonnet 4.6 just made the expensive model optional, and it costs the same as before
Good morning, everyone.
While the whole internet waits for Qwen 4 and GPT-6.2, Anthropic went ahead and shipped the model you can actually use today.
Claude Sonnet 4.6 came out on November 8. It is the middle model in the Claude family, the one that costs a fraction of the flagship. And Anthropic says its own early testers now pick it over Opus 4.5, the flagship from last November, more than half the time.
That is the part that got me. Not a new record on a benchmark chart. A middle model that people reach for first, because the expensive one stopped being the default answer.
| Sonnet 4.6 | Sonnet 4.5 | |
|---|---|---|
| Context window | 1M tokens, beta | 200K tokens |
| OSWorld computer use | 72.5% | 61.4% |
| Price per million tokens | $3 in, $15 out | $3 in, $15 out |
| Default on Free and Pro plans | Yes | No |
The price line is the quiet shock in that table. Same price as the old model, and the new one does work that used to need the flagship. In Anthropic's early testing inside Claude Code, users picked Sonnet 4.6 over Sonnet 4.5 about 70 percent of the time, and over Opus 4.5 about 59 percent of the time. They said it read the context before changing code, cleaned up shared logic instead of copying it around, and was far less likely to overbuild a simple fix or claim a job was done when it was not.
On SWE-bench Verified, the test that checks whether a model can fix real bugs from GitHub, Sonnet 4.6 scored 79.6 percent. General knowledge looks strong too, 89.3 percent on MMLU and 89.9 percent on GPQA Diamond.
A million tokens is not the story
A million tokens can hold a whole codebase, a long contract, or a stack of research papers in one request. The size alone is not the story. Anthropic says the model reasons well across all that context, and its proof is a strange little test called Vending-Bench Arena, where models run a fake business over time and try to make the most money.
Sonnet 4.6 spent big on stocking up in the first ten fake months, then turned hard toward profit in the final stretch, and finished well ahead of the rest. That is a planning story, not a memory story, and it is the one I will watch.
Context compaction, also in beta, sums up older parts of a chat on its own as the chat grows, so the window stretches further. But a bigger window is not a reason to throw everything into every prompt. Every token you send is one you pay for, and a model thinking across a million tokens is doing more work than one thinking across ten thousand. The skill now is picking the right things to put in, not fitting more in.
Computer use grew up as well. On OSWorld, the test where a model clicks and types its way through real software like Chrome, LibreOffice, and VS Code, Sonnet 4.6 scored 72.5 percent, up from 61.4 percent for Sonnet 4.5. Early users are seeing human-level skill on jobs like working through a messy spreadsheet or filling out a long web form across several browser tabs. Protection against prompt injection is much better too, now close to Opus 4.6, and that matters as much as the scores if you let a model click around on your behalf.
The part that changes your bill
Anthropic is unusually clear about which model to use for what.
Start on Sonnet 4.6. It is for the bulk of daily work: writing and reviewing code, planning agents, long-context thinking, knowledge work, design, and computer use. If a task comes back shallow or wrong in a way that smells like a thinking ceiling instead of a prompt problem, send that one task to Opus 4.6. Do not send everything to the most expensive model out of habit. That habit just got a lot harder to justify.
Opus 4.6 stays for the deepest thinking: big refactors, running several agents at once, problems where getting it exactly right matters most. The point of Sonnet 4.6 is that you reach for Opus less often, not never.
The Free plan got a real upgrade on the same day. It now includes file creation, connectors, skills, and compaction, with Sonnet 4.6 as the default. If you never tried Claude Code or Cowork because the Free plan felt too thin, that excuse just expired.
Anthropic also shipped Claude Haiku 5.5 on November 6, its fastest model so far, at about 75 percent less to run than Haiku 4.5. It is the first Haiku with an adjustable effort setting, so you can turn the thinking up or down for simple jobs like sorting, labelling, and summarising. Small models are having a good week.
The honest limits. No independent testing yet. The 70 percent and 59 percent preference numbers come from Anthropic's own early testing with its own users, and the 1M window is still in beta, so expect rough edges. Computer use is much better, not human. Anthropic says so itself. And Opus 4.6 still wins at the top end, by Anthropic's own account. Launch-day numbers are promises, not reviews, and the real test is a week of real work on real code.
Have you switched your default to Sonnet 4.6 yet? And if you have, did it actually read your code before changing it, or did it still rewrite the same helper twice? I want to hear what it did on your real work, because that is the test that matters.
Cover image: AI-generated