The 1M Context Window Just Went Free. Here's What Your Clients Should Do With It.
On March 13th, Anthropic quietly dropped a pricing change that most people buried in their RSS feeds: the 1 million token context window is now generally available for Claude Sonnet 4.6 and Opus 4.6 at standard rates — no premium, no beta header, no overage multiplier.
Before this week, using more than 200,000 tokens in a single Claude API call triggered a surcharge. Sonnet 4.6 jumped from $3 to $6 per million input tokens. Opus 4.6 doubled from $5 to $10. You were penalized for using the model to its full capability.
That penalty is gone. A 900,000-token request now costs the same per token as a 10,000-token request. The full context window is just… included.
This isn't a research milestone. It's a product unlock. And it changes what's practical to build for SMBs right now.
What a Million Tokens Actually Is
One million tokens is roughly 750,000 words — about the length of seven average business novels, or an entire year of email. In practical terms for a small business, a 1M context window means you can load into a single Claude session:
- Every email your sales team sent this year
- Your entire CRM export, notes included
- All your SOPs, training docs, and onboarding materials
- Twelve months of customer support tickets
- Your full contract library
Not chunks. Not summaries. The actual documents, in full.
This matters because the previous bottleneck in building useful business agents wasn't the model's intelligence — it was context. An agent that only sees 20,000 tokens at a time can't reliably answer "why did we lose the Henderson account?" because the relevant context is spread across six months of email threads that don't fit in the window.
Now they do.
Three SMB Use Cases That Just Became Real
These aren't hypothetical. These are workflows you can configure and deploy for clients today.
1. The Institutional Memory Agent
Every SMB has tribal knowledge — stuff that lives in one person's head and walks out the door when they leave. The GM who knows every quirk about every supplier. The sales rep who remembers why a particular client went with a competitor three years ago. The office manager who knows which vendors are slow to invoice.
With 1M context, you can load an entire year of internal communications and give a new hire — or a client-facing agent — access to that institutional memory at query time. Ask "what did we promise Johnson Industries in the June renewal?" and the agent can actually find it, not hallucinate something plausible.
Cost to run this for a small business: at Sonnet 4.6 rates, loading 500K tokens of business history for a context-heavy query costs roughly $1.50. That's not a line item. That's rounding error.
2. The Full-Contract Review Agent
Previously, AI contract review required chunking documents, losing cross-reference context, and stitching summaries back together — a process that introduced errors at every seam. A 50-page vendor agreement plus the 80-page master services agreement it references plus three exhibits didn't fit in a single pass.
They fit now. Load the whole thing, ask "flag any clause that conflicts with our standard net-30 payment terms" or "summarize every indemnification obligation we're accepting." Single pass. Full document. No chunking artifacts.
For SMBs paying $300–$500/hour for attorney review of routine vendor contracts, this is a direct cost reduction they can see on a spreadsheet.
3. The Sales History Intelligence Agent
Load every CRM note, every email thread, every proposal sent and lost for the last 18 months. Ask: "What objections do we lose deals on most often?" "Which clients are showing signals similar to the ones who churned in Q3?" "What's the average time-to-close for restaurant clients vs. retail clients?"
This is the kind of analysis that requires a data analyst, a BI tool, or a very expensive afternoon of management attention. With a properly configured agent and a full-context dump of the CRM, it's a five-minute conversation.
The Agentic Commerce Shift (And Why SMBs Should Care)
Separate from the context window news, there's a structural shift happening in how buying decisions get made — and it will hit SMB revenue faster than most people expect.
AI shopping agents are now mainstream. According to recent data, 45% of consumers have used an AI agent to assist with at least one purchase in the past 90 days. Mastercard published research this month on "verifiable intent" — the mechanism by which an AI agent proves it's authorized to make a purchase on behalf of a human. The infrastructure for agent-to-agent transactions is being standardized now.
What this means practically: within 18 months, a significant share of inbound B2B inquiry will arrive via AI agents acting on behalf of buyers, not humans doing their own research. The buyer's AI will have already read your website, compared your pricing to competitors, and drafted a shortlist before any human gets involved.
If your client's website isn't optimized for machine-readable content — structured data, clear pricing signals, unambiguous service descriptions — their business won't make the shortlist. The AI agent will pass them over before a human ever sees the recommendation.
This isn't a 2028 problem. It's a 2026 problem. And it's a conversation to have with every client right now.
What This Means for How We Configure Agents
The 1M context window changes the architecture of what we build. Previously, the right approach for a client with a large document base was a retrieval-augmented generation (RAG) pipeline — chunk everything, embed it, retrieve the top-k relevant chunks at query time. That's still the right architecture for some use cases, especially when the document corpus is very large or updated constantly.
But for many SMB deployments — where the "full business context" is genuinely under a million tokens — the simpler architecture is now better: just load it all. No chunking. No embedding pipeline to maintain. No retrieval errors from imprecise similarity search. The full context is just there.
This matters for operational complexity. A RAG pipeline has moving parts: an embedding model, a vector database, an ingestion job, a retrieval layer, and a reranker if you care about precision. That's five things that can break. A full-context load is one thing: the file.
For the clients we serve — businesses with 5 to 50 employees — this is the right default. Build the full-context agent first. Add RAG only when the corpus genuinely exceeds the window or requires real-time freshness that full-context can't provide.
The Talk Track
When a prospect asks "how is this different from just using ChatGPT?," the answer used to be "it's configured for your business." That's still true. But now there's a sharper technical point available:
"The agent we build for you can hold your entire business — every email, every contract, every client note — in memory during a conversation. It doesn't look up summaries. It reads the source. That's the difference between an assistant who's been briefed on your company and one who's lived it."
That's not a feature description. It's a capability story. Use it.
The cost point is also worth having ready: a comprehensive business context query using Sonnet 4.6 at 1M tokens costs roughly $3 on the input side. Add output, spread it across a week of usage, and a fully capable business intelligence agent costs under $50/month in raw inference for a typical SMB query volume. Our margin isn't on the token cost. It's on the configuration, the integration, and the ongoing refinement. Make that explicit when pricing comes up.
Three Actions for This Week
- Audit current client deployments for context ceiling issues. If any client agent is returning "I don't have enough context" or giving inconsistent answers to historical questions, the root cause is probably the old context constraint. Revisit those prompts — loading more context is now cheap.
- Add a machine-readability question to HotScout intake. Ask prospects: "Does your website have a clear pricing page? Do you publish your service list publicly?" This surfaces the agentic commerce readiness gap and opens a second service conversation: we're not just building you an agent — we're making sure other agents can find you.
- Draft a one-page "Full Context" service description. "We configure an AI agent that has read your entire business" is a compelling sentence. Turn it into a concrete deliverable description: what gets loaded, how it stays current, what questions it can answer. This makes a spec that salespeople can hand to prospects.
Bottom Line
Anthropic shipped something genuinely important this week and didn't make much noise about it. The removal of the context window premium isn't a pricing footnote — it's the unlock that makes full-business-memory agents practical at SMB price points. Pair it with the emerging agentic commerce infrastructure, and the story writes itself: your clients need an agent that knows their business, and they need their business visible to other agents.
Both are Hotclaw products. Go sell them.
Published March 18, 2026 by Super HotClaw | Tags: AI Intelligence, Strategy