What Claude Sonnet 5's 1M Context Window Actually Changes
Claude Sonnet 5 pairs a 1M-token context window with flat $2-per-million intro pricing — no long-context surcharge, unlike Gemini 3.1 Pro's step-up above 200K tokens. The window itself matters less than the economics of re-reading it: cache reads run up to 90% cheaper and skip Anthropic's input rate limits. Known caveats remain, from mid-window quality degradation to a 128K output cap. For PMs, the make-or-buy math on retrieval infrastructure just changed — corpora under a million tokens may no longer need a RAG pipeline at all.
If you own a product with a retrieval pipeline, Sonnet 5's pricing quietly reprices your architecture. The question to put on the table this week: which of your RAG use cases exist because retrieval genuinely wins, and which exist only because context used to be scarce or expensive? Those are different categories now. Your concrete action: pick one real workload — your gnarliest support-agent corpus or your monorepo Q&A tool — and run a spike this week that loads it whole into Sonnet 5's window with prompt caching enabled, then benchmark answer quality and cost-per-query against your existing pipeline. Cache reads priced up to 90% below base input, which also don't count against Anthropic's input rate limits, are what make this test economically meaningful rather than a demo. Price the results at both the $2/$10 intro rate and the $3/$15 standard rate that lands September 1, so you aren't building a business case on a discount. To be fair to the retrieval camp, RAG still wins on freshness, corpora beyond 1M tokens, and source auditability — don't rip anything out on the strength of one good benchmark. But run the spike now: intro pricing closes August 31, and competitors running the same test have roughly seven weeks of cheap experimentation left.
Claude Sonnet 5 offers a 1M-token context window at $2/$10 intro pricing through August 31, 2026, billed flat across the window with no long-context surcharge, per Anthropic's pricing docs.
Claude Sonnet 5 shipped on June 30 with a 1M-token context window, and the number that matters most isn't its 85.2% SWE-bench Verified score (llm-stats board, as of July 2026). It's the $2. Through August 31, Anthropic is billing Sonnet 5 at $2 per million input tokens and $10 per million output, flat across the entire window per its pricing docs, with standard $3/$15 pricing arriving September 1.
Two years ago, a million-token context was a flagship-tier capability with flagship-tier pricing. My verdict: the constraint on long-context AI has flipped from "can the model hold it?" to "can you afford to send it?" — and Sonnet 5 is the clearest signal yet that the answer is becoming yes for almost everyone.
The long-context premium is dead, and Anthropic killed it twice
For most of the 2023–2025 cycle, context length tracked price. If you wanted the biggest window, you bought the most expensive model, and some providers layered a long-context surcharge on top. Anthropic broke that link in February, when Claude Sonnet 4.6 shipped a 1M window at $3/$15 with no long-context premium: a 900K-token request billed at exactly the same per-token rate as a 9K one.
Sonnet 5 keeps that flat structure and, until August 31, lowers the base price underneath it.
Here is where the field stands, as of July 2026 (live numbers on the model tracker):
| Model | Context window | Price per 1M tokens (input / output) | Long-context step-up |
|---|---|---|---|
| Claude Sonnet 5 | 1M (128K max output) | $2 / $10 intro through Aug 31, 2026; then $3 / $15 | None — flat across the window |
| GPT-5.6 Terra | 1.05M (128K max output) | $2.50 / $15 | None |
| Gemini 3.1 Pro | 1M | $2 / $12 up to 200K input | $4 / $18 above 200K input |
The Gemini row is the one to study. Google's sticker price matches Sonnet 5's intro rate, but cross 200K input tokens and the bill roughly doubles. That means the true 1M-token price race is between Anthropic and OpenAI, a matchup I break down in our builder-focused Sonnet 5 vs. GPT-5.6 Terra comparison.
The intro discount itself is no accident of generosity, either. As our launch analysis argues, it is IPO-timed aggression designed to move share while the window — the calendar one — is open.
What a million tokens actually buys
The honest answer: enough that chunking stops being the default architecture. A window this size comfortably holds a mid-sized production codebase, a quarter's worth of customer support transcripts, a litigation document set, or the full running trace of a long agent session — the kinds of corpora teams have spent two years slicing into embeddings because no window could take them whole. If you want the fundamentals of how windows work and why they bind, start with our context window explainer.
But holding it and affording it are different questions. Fill the window once and you've spent about $2 in input at intro pricing, $3 from September. That is fine for a single deep analysis and ruinous if your product does it ten thousand times a day.
Raw window size was never the real product. The economics of re-reading are.
Caching is what makes it repeatable
This is where Sonnet 5's numbers get strategically interesting. Prompt caching prices cache reads up to 90% below base input, with cache writes at 1.25x base, per Anthropic's docs. Load your codebase into the cache once, and every subsequent query reads those tokens at a tenth of the price.
Better still: on the Claude API, cached reads don't count toward input-token rate limits, a documented Anthropic-specific advantage that matters enormously at this scale, because even Sonnet 5's dedicated Start-tier bucket of 2M input tokens per minute covers only two uncached full-window calls. Add the Batch API's 50% discount (and its 300K-token output ceiling via a beta header, versus 128K standard), and "keep a million tokens hot and interrogate them all day" becomes a line item instead of an experiment.
Long context is following the cloud-storage curve: a scarce premium capability commoditized into a cheap substrate, with the durable businesses built on top of the cheap layer rather than on selling it. Notably, you no longer pay up for the window even inside Anthropic's own lineup. Opus 4.8 has the same 1M context at $5/$25, so the flagship premium now buys capability, not capacity — a trade-off I weigh in our Sonnet 5 vs. Opus 4.8 analysis.
The caveats — and the 18-month picture
Two honest ones. First, long-context quality degradation is a known pattern across the industry: models generally retrieve and reason better over material near a window's edges than material buried in its middle, and performance on a stuffed window rarely matches performance on a focused prompt. Anthropic hasn't been shown to be exempt, and I'd wait for independent long-context evaluations before betting a legal-review product on mid-window recall.
Second, output is capped at 128K tokens on standard requests. You can read a million tokens, but you can't write one, which constrains whole-corpus transformation tasks.
And to be fair to Google, its step-up pricing isn't gouging. Serving cost genuinely grows with sequence length, and the $4/$18 tier reflects real economics that Anthropic is choosing to absorb — presumably a share-grab it can afford ahead of an IPO.
Here is my falsifiable call: by Q2 2027, flat-rate million-token pricing is standard across all three major labs, and a visible cohort of context-first products ships with no retrieval pipeline at all for corpora that fit the window. The harder strategic question lands on everyone who spent 2024–2025 building RAG infrastructure: when stuffing the window costs less than maintaining the pipeline, which parts of your retrieval stack are still assets, and which just became technical debt with a maintenance bill?
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
It is genuinely useful for whole-corpus work: mid-sized codebases, long agent traces, and large document sets can now be processed without chunking pipelines. The honest caveat is that long-context quality degradation is a known pattern across LLMs — models tend to handle material near the window's edges better than material buried in the middle. So the window is real, but treat vendor claims about full-window reasoning with skepticism until independent long-context evals land. The economics, not the raw size, are what changed with Sonnet 5.
Filling the window costs about $2 in input tokens at the intro rate, rising to about $3 when standard $3/$15 pricing takes effect September 1, 2026, with output billed separately at $10 (then $15) per million. Anthropic bills flat across the window, so a 900K-token request costs the same per token as a 9K one. Prompt caching is what makes repeated use viable: cache reads run up to 90% below base input, cache writes bill at 1.25x, and cached reads don't count toward input rate limits on the Claude API. The Batch API adds a further 50% discount on both input and output.
Not wholesale, and not yet. Retrieval still wins on corpora larger than 1M tokens, on freshness, and on auditability of which sources produced an answer. Long context also carries real downsides: mid-window recall degradation, a 128K output cap that limits whole-corpus transformations, and per-request costs that climb fast without caching discipline. The right move in July 2026 is a side-by-side spike on one workload, not a migration.
