Claude Sonnet 4.6: What Service Business Owners and Founders Actually Need to Know

Released yesterday. Already inside my workflow. Here's what changed — and why it matters for building software and running AI-assisted operations.

Anthropic released Claude Sonnet 4.6 on 17 February 2026. It's a meaningful step forward for anyone using AI to run a business — better coding, near-Opus computer use, 1M token context, and smarter office task handling. Here's what actually changed and how I'm using it.

I'm writing this post from inside a Claude Sonnet 4.6 conversation.

That sentence is doing more work than it looks like. The blog post you're reading was researched, fact-checked against my live MCP API, structured, written, and published — all from a single Claude conversation. No copy-paste. No manual CMS work. Just a conversation that ends with a live post on hellocrossman.com.

That workflow is possible because of how significantly AI tooling has improved. And Sonnet 4.6, released yesterday (17 February 2026), is the latest — and most substantial — step forward.

Here's what actually changed, what the benchmarks mean in plain English, and how I'm already using it.

What Is Claude Sonnet 4.6?

Claude is Anthropic's AI model. The "Sonnet" tier sits in the middle of their model range — faster and cheaper than Opus (their most capable model), more powerful than Haiku (their fastest, smallest model).

Sonnet 4.6 is the first upgrade to the Sonnet model since version 4.5 arrived in September 2025. According to Anthropic, it's a "full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design."

That's not marketing language. The benchmark improvements are real, and the practical implications for service business owners are significant.

The Five Things That Actually Changed

1. Coding Got Meaningfully Better

This is the headline improvement. Sonnet 4.6 is more consistent when modifying existing code, better at following precise instructions, and less prone to the "overengineering spiral" where AI coding tools add complexity when you ask for something simple.

Early testing data is striking: developers preferred Sonnet 4.6 over Sonnet 4.5 in 70% of head-to-head comparisons. More telling — they preferred it over the previous Opus 4.5 in 59% of comparisons. Opus is Anthropic's premium, higher-cost tier. The fact that a Sonnet model is outperforming last year's Opus on coding tasks is a genuine shift.

For founders using tools like Replit Agent, Cursor, or Lovable — or anyone working with Claude Code for production builds — this matters. The consistency improvements mean less time debugging AI-introduced regressions and more time shipping.

I've written before about the final 10% of software development — the gap between an AI-generated prototype and a production-ready product. Sonnet 4.6's improvements in instruction-following and code reasoning push that gap in the right direction. It won't close it. But it narrows it.

2. Computer Use Is Now Near-Opus Level

"Computer use" means Claude can navigate software interfaces the way a human would — clicking, typing, filling in forms, navigating spreadsheets — without needing special APIs or integrations.

Sonnet 4.6 scores 72.5% on OSWorld-Verified, the industry benchmark for computer use. Opus 4.6 scores 72.7%. The gap between the budget model and the premium model is now 0.2 percentage points.

For practical context: tasks like navigating a complex spreadsheet, filling out a multi-step web form, or working through a structured process in a web application are now within reach at Sonnet pricing. That's $3 per million input tokens versus Opus's significantly higher rate.

For service businesses exploring AI agents, this is important. Many automation workflows don't need reasoning depth — they need reliable interface navigation. Sonnet 4.6 can handle those workflows without paying for Opus.

3. 1M Token Context Window (Beta)

The context window is how much information Claude can hold in a single conversation. Sonnet 4.5 had a 512K token window. Sonnet 4.6 doubles that to 1M tokens in beta.

1M tokens is roughly 750,000 words. In practical terms, that's:

For service businesses, this changes what's possible in a single working session. You can load your complete methodology documentation, your client history, your process playbooks — and work with all of it simultaneously without the conversation "forgetting" earlier context.

Paired with context compaction (which automatically summarises older conversation when approaching limits), conversations can effectively run indefinitely without losing critical information.

4. Office Task Performance Leads All Models

This one surprised me. On GDPval-AA, the benchmark for real-world office productivity tasks — financial analysis, document comprehension, data extraction — Sonnet 4.6 scores 1,633 Elo. That leads all Claude models, including Opus.

Specific numbers from enterprise testing: Sonnet 4.6 achieved 89% accuracy on tasks requiring mathematical calculation (up from 62% in Sonnet 4.5), and 88% on government and compliance-related tasks. For a compliance consultancy, a financial advisory firm, or any service business where precision in document work matters, these aren't abstract improvements.

This connects directly to why I built the AI blog pipeline described in How I Built an AI Blog Pipeline That Researches, Writes, and Publishes Itself. Content work — research, cross-referencing, fact-checking against live data sources, structured writing — is exactly the kind of office task where Sonnet 4.6 shows the most meaningful improvement over its predecessor.

5. Instruction Following Got More Reliable

This is less headline-grabbing than the benchmark numbers but arguably the most important improvement for anyone building production workflows.

Agentic AI workflows — sequences of tasks that Claude performs autonomously — live or die on instruction following. If Claude reliably executes step 3 of a 10-step process but occasionally interprets step 7 differently, your workflow has a reliability problem. You end up adding error handling, retry logic, and manual checkpoints that eat up the time you were trying to save.

Sonnet 4.6's improvements in consistency directly improve the reliability of multi-step workflows. For the MCP server architecture I use across my own business operations — blog publishing, lead tracking, analytics, contractor placement — this translates to fewer edge cases and more predictable behaviour.

How I'm Using It Right Now

This post is the most literal example: Sonnet 4.6 is running the entire content production pipeline for hellocrossman.com. It queries the MCP API for canonical statistics (so numbers are never invented), checks existing posts for internal linking opportunities, verifies case study metrics, and publishes directly via the API.

But the applications extend well beyond content.

Contractor placement operations. My cybersecurity contractor placement business uses Claude to search and match contractors against job requirements, draft personalised outreach emails, and manage shortlisting. The consistency improvements in Sonnet 4.6 mean the matching logic runs more reliably across edge cases — contractors with unusual certification combinations, roles with non-standard requirements.

Analytics interpretation. Rather than reading dashboards manually, I pipe analytics data from hellocrossman.com into Claude and ask for interpretation and recommendations. The office task improvements mean more accurate pattern recognition across longer data windows.

Specification drafting. Better instruction following means more reliable generation of the structured specs used in my BuildKits tool — the AI-powered specification generator I built for founders who want to stop wasting development budget on vague prompts.

What This Means If You're Building with AI

If you're using AI coding tools to build your first software product, Sonnet 4.6 is the best general-purpose model available at its price point. The coding consistency improvements reduce the frequency of the "AI introduced a regression" problem — where fixing one thing breaks two others.

If you're further along — managing agentic workflows, building MCP servers around your methodology, or exploring what an AI-powered version of your service business actually looks like — the instruction-following and office task improvements are where you'll feel the difference most.

If you're not yet using Claude at all and you're running a service business, the straightforward starting point is this: try loading your most complex recurring analysis or reporting task into a conversation, give Claude the data, and ask it to do what you'd normally do manually. The 1M token context window means you can give it far more context than you might expect.

The Pricing Stays the Same

Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens. Same as Sonnet 4.5.

Given the performance improvements — especially on tasks that previously required Opus — this represents a meaningful increase in value at the same price point. Anthropic has simultaneously upgraded the free tier to use Sonnet 4.6 as the default model, which means the model I'm describing here is what free users now access by default.

For context: Opus 4.6 costs significantly more. If your workflows were previously on Opus because Sonnet wasn't reliable enough, Sonnet 4.6 is worth re-evaluating now.

Is This Relevant If You're Not a Developer?

Yes. The most important capability improvements in Sonnet 4.6 — office tasks, computer use, instruction following — don't require any coding to benefit from.

If you're a service business owner who wants to understand whether AI agents could help your operations, or if you're a founder exploring what software built from your methodology might look like, the starting point is a conversation — not a codebase.

That's the core proposition of what I do at hellocrossman.com. I take 18 years of product development experience across companies like Habito (where products I built processed £3B in mortgages) and apply it to service businesses who have a methodology worth productising. The AI tools — Claude included — handle execution. The human judgment handles everything that makes the product actually work.

If you want to explore what that looks like for your business, the discovery sprint is the place to start.

FAQ

What is Claude Sonnet 4.6?

Claude Sonnet 4.6 is Anthropic's mid-tier AI model, released on 17 February 2026. It's a significant upgrade over Sonnet 4.5 with improvements to coding consistency, computer use, long-context reasoning, and office task performance. It's now the default model for free and Pro Claude users.

How does Claude Sonnet 4.6 compare to previous versions?

Sonnet 4.6 outperforms Sonnet 4.5 in developer preference in 70% of comparisons and even beats the previous Opus 4.5 in 59% of comparisons on coding tasks. On mathematical accuracy, it jumped from 62% to 89%. On computer use benchmarks, it's within 0.2% of the current Opus 4.6.

Can service businesses use Claude without coding skills?

Yes. Claude's most practically useful capabilities for service businesses — document analysis, structured writing, data interpretation, computer use — don't require coding knowledge. The 1M token context window means you can work with substantial volumes of your own business data in a single conversation.

What's the difference between Claude Sonnet and Claude Opus?

Opus is Anthropic's most capable model, optimised for complex multi-step reasoning and deep coding tasks. Sonnet is the mid-tier model — faster, more cost-effective, and with Sonnet 4.6, now near-Opus performance on computer use and office tasks. For most service business applications, Sonnet 4.6 is the right choice.

How much does Claude Sonnet 4.6 cost?

$3 per million input tokens and $15 per million output tokens via the API. Pricing is unchanged from Sonnet 4.5. For most founders and service business owners using it conversationally through Claude.ai, it's included in free and Pro plans.