Sonnet 4.6 Is Now the Default — Even on the Free Plan: 1M Context at the Same Price
On November 8 Anthropic released Claude Sonnet 4.6: a 1M-token context window in beta, now the default model on the Free and Pro plans, at exactly the same price as Sonnet 4.5. Early testers preferred it over Sonnet 4.5 about 70% of the time — and over the previous flagship Opus 4.5 59% of the time.

On November 8, Anthropic released Claude Sonnet 4.6. The official announcement puts it plainly: this is the most capable Sonnet model yet, a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design — plus a 1-million-token context window in beta.
But the numbers on the spec sheet are not the part vibe coders should pay attention to. Two sentences matter more. First: from this day on, Sonnet 4.6 is the default model in claude.ai and Claude Cowork on the Free and Pro plans. Second: pricing is exactly the same as Sonnet 4.5 — $3 per million input tokens and $15 per million output tokens, not a cent more. A stronger model, at the same price, becoming the default everyone gets when they open the app — that is the most concrete change in this release.

1. A 1M context window: holding the whole codebase is only the first half
What does 1 million tokens actually mean? According to Anthropic, the window can hold an entire codebase, a very long contract, or dozens of research papers in a single request. For vibe coders, that means a big project that used to be fed to the model in several pieces can now, in theory, be handed over all at once: no more reading half the code first and then chasing the other half, no more stitching context together in your own head.
But the more important half of the sentence comes after: Sonnet 4.6 doesn't just hold more, it "reasons effectively" across all that context. A bigger window is not the rare part — the rare part is whether long-horizon planning keeps up. Anthropic's example is Vending-Bench Arena, an evaluation that asks a model to run a simulated business over time. Sonnet 4.6 took a striking route: it invested heavily in capacity for the first ten simulated months, spending far more than its competitors, then pivoted sharply to focus on profitability in the final stretch, finishing well ahead of the competition. Knowing when to spend, when to pull back, and when to change course is closer to being able to carry a long task than simply remembering more.
One caveat: the 1M window is still in beta. Anthropic also put context compaction into beta at the same time: as a conversation approaches its limits, older content is automatically summarized, stretching the effective context further. Read together, the intent is clear — long sessions should not break. But beta means rough edges remain, and hallucination and attention drift inside long contexts are still yours to catch at the key checkpoints, not something you can ignore by dumping a whole codebase in and walking away.
2. 70% and 59%: Anthropic's own early testing — don't read them as independent benchmarks
The most eye-catching numbers in the announcement come from early testing in Claude Code: users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time — and even preferred it to the previous frontier model from November 2025, Claude Opus 4.5, 59% of the time. The specific strengths users mentioned: it reads the context carefully before modifying code, and it consolidates shared logic rather than duplicating it everywhere. That made it less frustrating to use over long sessions.
Both numbers need their qualifier attached: they come from Anthropic's own early testing, not independent evaluation. Anthropic says so in the announcement itself. As a reader, the qualitative descriptions behind the numbers are worth more — "significantly less prone to overengineering and laziness," better instruction following, fewer false claims of success, fewer hallucinations, more consistent follow-through on multi-step tasks. These are exactly the most grinding problems in vibe coding: a model that swears it fixed something and, when you run it, plainly didn't; or one that rebuilds half the project to complete a small request. If Sonnet 4.6 genuinely converges on these, that matters more to people who pair-program with agents every day than any benchmark score.
Computer use is the other improvement with numbers behind it. OSWorld is the standard benchmark for computer use, putting a model in a simulated computer to operate real software — Chrome, LibreOffice, VS Code — with no special APIs. According to same-day coverage that checked the announcement, Sonnet 4.6 scored 72.5% on OSWorld, up from 61.4% for Sonnet 4.5. Anthropic also says early users are seeing human-level capability in tasks like navigating a complex spreadsheet or filling out a multi-step web form before pulling it all together across multiple browser tabs. But the announcement keeps a hand back: the model still lags behind the most skilled humans at using computers, and prompt injection remains a real risk — Sonnet 4.6's resistance to prompt injection is a major improvement over Sonnet 4.5, performing similarly to Opus 4.6, but that is not the same as being safe to let it click around any web page.
3. The free tier upgrade: files, connectors, skills, and compaction, all included
The easiest change to overlook in this release — and the one with the biggest effect on independent developers — is the free tier upgrade. As Sonnet 4.6 becomes the free plan's default model, the free tier also gains file creation, connectors, skills, and context compaction. In other words, Anthropic has dropped the workflow that used to require payment down to zero cost.
For vibe coders, that lowers the cost of trying things by another notch: without paying $20 first, you can run a real task in Claude Cowork with file creation and skills, and only upgrade when it feels right. Read together with unchanged pricing, Anthropic's strategy is clear — make money by keeping more people on the default option, not by charging more for it. For paying Pro users, the change is a stronger default model with an unchanged bill; for people still wavering between tools, the free tier's new bundle makes "just try Claude" much easier.
There are a few more product updates worth noting. On the API, the web search and fetch tools now automatically write and execute code to filter and process search results, keeping only relevant content in context — improving both response quality and token efficiency. Code execution, memory, programmatic tool calling, tool search, and tool use examples are now generally available. The Claude in Excel add-in now supports MCP connectors, letting Claude pull context from tools you use day to day — S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet — without leaving the spreadsheet, on Pro, Max, Team, and Enterprise plans. The API model ID is claude-sonnet-4-6, available now on all Claude plans, Claude Cowork, Claude Code, the API, and all major cloud platforms. For people already using Claude Code, there is one more immediate change: early testers say the model makes fewer false claims of success and fewer hallucinations, and follows through more consistently on multi-step tasks — improvements you get just by switching the default model, with no configuration changes at all.
4. How to read it: mid-tier models are eating flagship work
Put this release back into the model rhythm of 2026 and the picture is clear: in early testing, Sonnet 4.6 beats the previous flagship Opus 4.5 59% of the time. Anthropic draws the line itself in the announcement: Opus 4.6 remains the strongest option for tasks that demand the deepest reasoning, such as codebase refactoring, coordinating multiple agents in a workflow, and problems where getting it just right is paramount. In other words, the flagship's job is narrowing — only the tasks that truly need the deepest thinking are left to it.
That is good news for vibe coders, with one warning attached. The good news is that most everyday development work — writing features, fixing bugs, building frontends, running analysis — is being taken over by mid-tier models at mid-tier prices, so the flagship budget can be saved for the sharp end. The warning is that Sonnet 4.6, however strong, is still the Sonnet 4.6 of Anthropic's own testing: the 1M context window is in beta, the preference percentages come from early testing, and the computer-use numbers come from vendor-cited benchmarks. The only way to really test it is to put it into your own project for a week and see whether it actually lies about success less often and copies less duplicated logic. The numbers at a launch event are only the ticket in — the answer is in your own codebase.
5. The other side of a 1M context: the more you load, the more carefully you must count the bill
Once the window gets bigger, the first thing you hit is not a capability problem — it is a cost problem. At the announced pricing, 1 million input tokens cost $3, with output billed separately at $15 per million. In a session where the whole codebase sits in context, the input token bill quietly grows over dozens of turns, because every turn re-reads everything accumulated so far. Anthropic shipping context compaction at the same time is, in part, a patch for this: when a conversation nears its limit, older content is summarized automatically, which helps control the size of the context.
For vibe coders, there are three sums to learn to do. First, whether "handing over the whole codebase" is worth it: for architecture reviews, finding duplicated logic, and cross-module refactors, feeding the full codebase at once saves rounds of explanation and is worth the money; but for fixing one isolated feature, there is no need to drag the entire repository into context — the more the model reads, the more chances it has to get distracted and make mistakes. Second, caching: when the same codebase shows up in many turns of one conversation, prompt caching pushes the cost of the repeated parts down, so it is usually cheaper to keep talking in a long session than to keep opening fresh ones. Third, output: the easiest cost to overlook in long-context tasks is on the output side — a model spitting out thousands of lines of code at $15 per million output tokens, five times the input price, so asking it to explain the plan first and only then write the code saves more than telling it to "just write everything."
There is one more detail that is easy to miss: Anthropic says Sonnet 4.6 "offers strong performance at any thinking effort, even with extended thinking off." That sentence is written for API users — tasks that used to need high thinking effort to stay stable can now start at lower effort, which is friendlier on both speed and cost. The official migration advice is to "explore across the spectrum to find the ideal balance of speed and reliable performance" rather than cranking the thinking budget to maximum on day one. For independent developers, the subtext is: don't rush to buy the most expensive configuration for every request — the default effort may already be enough for most everyday tasks.
6. In the 2026 model landscape: the mid-tier is repricing what "good enough" means
The timing of this release is also worth noting. About two weeks earlier, Anthropic released the flagship Opus 4.6; and looking further back, GitHub has announced that on November 16 it will retire three models from Copilot — Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra — with the suggested replacements being Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna. Model generations are turning over so fast that even the "default option" has to be re-chosen every year, and a developer's model list has a shorter and shorter shelf life.
Against that background, Sonnet 4.6's strategy looks deliberately restrained: no price hike, no new tier, no limited-time discount — just a stronger model put into the same price and the same default slot. For individual developers, that restraint is friendlier than any spec: you don't have to redo the math or re-pick your stack; you just open the app and get a stronger default. Anthropic's bet is that as the mid-tier keeps raising the bar for "good enough," users will stop agonizing over which flagship to use for every task. Upgrading the free tier with file creation, connectors, and skills is the same move: get people into the ecosystem first, and paying becomes the natural next step rather than something forced by a paywall.
Of course, competitors are not standing still. In the same week, OpenRouter shipped a rebuilt Auto Router and a coding-specialized Auto Code, selling "you don't pick the model yourself — the router picks the most cost-effective one per task." And the model vendors — Anthropic, OpenAI, Google — are swapping default models every month. For vibe coders, that means two things. First, model selection is turning from a one-time decision into ongoing maintenance — the default you settle on today may not be the best one three months from now. Second, precisely because things change this fast, it is worth keeping your prompts, skills, and workflows in your own hands rather than tying them to one model's quirks. Models will keep turning over; your project structure, tests, and acceptance criteria are the long-term assets.
7. As agent planning gets stronger, the pairing posture should change too
One passage in the announcement is easy to skim past, but it may matter most to people who use Claude Code every day: Sonnet 4.6's improvement in agent planning and long-context reasoning. The two things early testers mentioned — "reads the context before modifying code" and "consolidates shared logic rather than duplicating it" — sound unremarkable, but they map to the two most common ways agents fail: changing code without thinking, and copy-pasting their way into new tech debt. If both genuinely improve, long sessions will feel much better: a session that used to need replacing after two hours might last longer, and tasks that needed constant human correction might be left to run a few more steps on their own.
But that also asks more of the user. The better a model plans, the more clearly you have to state the goal. The Vending-Bench example is a reminder: Sonnet 4.6 decided on its own to invest heavily in capacity for ten months and then pivot to profitability — the strategy was right, but only because the evaluation gave it a clear objective function (make the most money). In your project, the objective function is the requirements and acceptance criteria you write: the vaguer they are, the more room the model has to improvise, and the harder it is to predict which direction it will go wrong in. A habit worth building alongside this release: before a long task starts, ask the model to restate the goal and plan it understood, and only let it start once you have confirmed it; halfway through a long session, ask it to summarize progress and remaining steps — a human version of context compaction.
There is also the computer-use angle. Sonnet 4.6 went from 61.4% to 72.5% on OSWorld, and early users are already seeing near-human-level performance on complex spreadsheets and multi-step forms. For vibe coders, that means some of the chores that used to be unavoidably manual — copying data from one system to another, filling out a series of forms from a template — are starting to be delegable. But the warning about prompt injection in the announcement deserves a careful read too: malicious instructions hidden on a web page can hijack a model that is operating a computer. Before letting an agent operate a computer, think about which pages it will open and whether those pages contain content you don't control. The stronger the capability, the sooner the guardrails need to go up.
8. Three practical tips for vibe coders
First, switch the default model without switching your workflow. Sonnet 4.6 is already the default on the Free and Pro plans, available directly in Claude Code, with the API model ID claude-sonnet-4-6. If you used to switch back and forth between Sonnet 4.5 and Opus to save money, try running on 4.6 as the default for a week and see whether long-session stability feels different before deciding whether to bring in Opus 4.6 for specific hard tasks.
Second, treat the 1M context window as a tool for "handing over the whole codebase," not as an excuse to stop breaking tasks down. Feeding the entire codebase at once fits architecture reviews, cross-module refactors, and finding duplicated logic; but when writing a specific feature, still give the model one clear task boundary at a time, and let context compaction keep the long session unbroken. A big window does not mean unlimited attention — the clearer the task is broken down, the less room the model has to mess up.
Third, don't rebuild your workflow around computer use just yet. An OSWorld score of 72.5% is a handsome number, but the announcement itself admits the model is still behind the most skilled humans. Repetitive computer work like filling in forms and tidying spreadsheets is worth trying on a small scale; for anything involving payments, deleting data, or sending things out, keep a human in the loop to click confirm. It is far steadier to let go after you have run it successfully a few times yourself than to hand everything over based on launch-day numbers.
In one sentence: Sonnet 4.6 is not another routine "here's a stronger model" update, but a combination move — same price, upgraded default, and a completed free tier. Its real test is not on launch day but in the weeks that follow: when millions of free and Pro users are all on it by default, can Anthropic make "the mid-tier model is good enough" feel like everyday reality for most people?
Comments (0)
Sources
Related articles
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.
On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.