Anthropic’s Claude Sonnet 4.6 and the Million-Token Moment

Big AI announcements usually come with flashy charts and bigger numbers. This one is a number too, but it changes the shape of what you can build.

On February 17, 2026, Anthropic released Claude Sonnet 4.6 and opened a one-million-token context window (beta). For developers, this means you can feed Claude a huge pile of text or code in a single request, then ask it to reason across the whole thing, instead of slicing everything into tiny chunks and hoping retrieval grabs the right pieces.

Long context pushes LLMs away from “answer my question” and closer to “work inside my project.” A model that can hold a repo, a policy set, or a contract stack in one request is easier to turn into an agent that stays consistent across multi-step tasks.

Claude Sonnet 4.6 at a Glance (Context Window, Price, Availability)

Release date: Feb 17, 2026
Model: Claude Sonnet 4.6 (Anthropic’s most capable Sonnet so far)
Context window: up to 1M tokens (beta, API-only)
Base pricing: $3 / million input tokens and $15 / million output tokens
Long-context billing detail: if you enable 1M context and go over 200K input tokens, the request is billed at long-context rates

Claude Sonnet 4.6 1M Tokens Explained For Real Work

“1 million tokens” isn’t a neat word count, but think hundreds of thousands of words. Enough for things like:

a large codebase and its documentation
long legal or policy libraries
many research papers in one prompt

The practical win: fewer hacks. You spend less time building fragile pipelines for chunking, stitching, and re-asking the model what it “forgot.” Long context doesn’t kill RAG (Retrieval Augmented Generation), but it changes what you use RAG for: cost control, speed, and filtering noise, not basic survival.

The Uncomfortable Truth: Long Context Is Easy to Claim and Hard to Use

A widely cited paper, “Lost in the Middle", found that model performance can degrade when relevant information sits in the middle of a long context — often doing best near the beginning or end instead.

The problem isn’t just psychology — it’s compute. In standard transformer attention, longer sequences can drive sharply rising compute and memory requirements, which is one reason model providers have historically treated huge context windows as premium features.

So when Anthropic says Sonnet 4.6 not only fits 1M tokens but “reasons effectively across all that context,” it’s making a claim about usefulness, not just capacity.
One way to evaluate that claim is to look at the company’s own long-context testing. In Sonnet 4.6’s system card, Anthropic reports results on MRCR v2 (an 8-needle evaluation at 1M context), showing Sonnet 4.6 in the same neighborhood as its flagship Opus 4.6 on long-context sequential reasoning.

Even if you treat provider benchmarks cautiously — and you should — there’s a broader signal here: the frontier is shifting from “can you accept long input?” to “can you reliably operate on it?”

Benchmarks People Will Quote: SWE-bench, OSWorld, and Long-Context Tests

Anthropic is also trying to prove the window is usable, not just big. In its system card, it reports:

OSWorld-Verified: 72.5% for Sonnet 4.6
SWE-bench-Verified: 79.6% for Sonnet 4.6

For long-context evaluation, Anthropic cites MRCR v2 (a “multiple needles” test at 1M context) and positions Sonnet 4.6 as strong on sequential reasoning at that scale.
If you build agents that click around, edit files, or run tools, these numbers matter because the problem is rarely “does it know a fact.” The problem is “does it stay coherent over 30 steps.”

Pricing and Long-Context Costs (What 1M Tokens Can Cost)

Anthropic kept Sonnet’s base price the same, but long context has a higher tier once you cross the 200K input token mark (with 1M enabled).

Anthropic also highlights cost levers like prompt caching and batch processing in its API docs, which matter a lot once “stuff the whole repo into the prompt” becomes a normal request.

The Agent Security Angle: Bigger Context, Bigger Attack Surface

Long context and tool use come with a downside: more untrusted text can enter the model’s “working area.” Anthropic explicitly calls out prompt injection risks in agent settings and says Sonnet 4.6 improves resistance compared to Sonnet 4.5.

The same system card also notes a behavior to watch: higher “over-eagerness” in GUI computer-use settings (trying messy workarounds when a task is blocked), even if it’s more steerable with the right instructions.

How to Use a Million-Token Window Without Turning It Into a Junk Drawer

More context can make answers worse if you feed the model an unfiltered swamp. The patterns that help are boring, which is how you know they work:

Put the task first. Lead with goal, constraints, and output format before the big dump of material.
Add a simple index. A short “what’s in here” map helps the model navigate.
Use short “anchor” summaries. One paragraph per doc/module can reduce confusion later.
Make the model plan, then act. A quick plan step catches misunderstandings early.
Treat long context like a workspace. Keep it clean: remove duplicates, strip boilerplate, label sections.

The takeaway

Claude Sonnet 4.6 makes the million-token context window feel less like a lab flex and more like a product feature teams can actually ship with. The number is the headline, but the real shift is what it enables: AI that can work inside large, messy projects without constantly losing the thread.

It also raises the bar. Once users get used to “paste everything and help me,” going back to an 8K context window will feel like a step backward.