It finishes the second half of the job
Launch partners on anthropic.com describe two-part chores—update a CRM, then send the announcement—that older Sonnets abandoned. That matches what I saw in agent loops: fewer polite stops, more “I will verify.”
Field notes · Anthropic Sonnet line
Claude Sonnet 5 is the daily driver I open for agents, patches, and long briefs
I spent this week in the Sonnet 5 card on Claude.ai and the anthropic.com model page. It is the default on Free and Pro, priced for production agents, and honest about computer use. Below is how I would actually pick it—not a brochure reprint.
Why it stuck
What I keep after a week on Sonnet 5
These are working notes from the product page and the chat app, not a scoreboard. I care about whether a job finishes, whether the model checks itself, and whether the bill still looks like a Sonnet bill.
Coding is the center of gravity
The official use-case list leads with the full software loop: plan, implement, debug, maintain, refactor. I used it on a messy TypeScript service and it traced a race instead of papering over the symptom. That is the brownfield pitch, and it held.
Computer use is a product, not a demo GIF
Sonnet was the first frontier model Anthropic put on a desktop. Sonnet 5 is the version they now recommend for browser chores: competitive scans, procurement, onboarding. If your work lives in a GUI, this is the reason to read the rest of the page.
The price is still a Sonnet price
Two dollars in and ten dollars out per million tokens, with cache and batch discounts, is the number on the product page as of this writing. Effort levels can spend more thinking tokens; the point is you are not forced into Opus rates for everyday agents.
Capability showcase
The work Sonnet 5 is actually for
Claude.ai is a text-and-code product with files, artifacts, Cowork, and computer use. I am not hanging a picture gallery on a chat model. These cards are compact samples of the jobs on the official page: a PR pass, an agent loop, a newsletter calendar, a study plan, a Cowork brief.
1. Auth on mutating routes
2. No f-string SQL
3. Tests that hit query shape
4. Do not log raw tokens
5. Reversible migrations
Write a failing 401 test
before style nits.plan → tool → check → done
Break the job into tool calls.
Call browser or terminal only
for the next missing fact.
If output is wrong, correct
course instead of stopping.Month 1: what Sonnet 5 is for, the coding loop, a Cowork one-pager, reader letters. Month 2: computer use, effort levels, 5 vs 4.6 in one repo. Month 3: a prompt library and a recap of Free vs Pro.
Mon–Tue recap lectures. Wed timed past paper. Thu only the misses. Fri paper B with no notes. Sat light review and sleep. Flag anything that still needs a source—do not invent citations.
Every Monday, pull last week’s numbers. Active users, revenue, conversions up top with week-over-week change, an eight-week trend, traffic sources. Flag anything moving the wrong way. Deliver a one-pager I can read on a phone.
What it is
What Claude Sonnet 5 is, after opening the real pages
Claude is Anthropic’s assistant. Claude Sonnet 5 is the Sonnet-class model they launched on June 30, 2026 and now treat as the default on Free and Pro. The pitch is blunt: agentic work that recently needed a larger, more expensive model, at Sonnet speed and Sonnet price. You chat with it on the official Claude Sonnet 5 website, on iOS and Android, and through the Claude desktop app. Developers call claude-sonnet-5 on the Claude API and the same id on Amazon Web Services, Google Cloud, and Microsoft Foundry.
I opened anthropic.com/claude, the dedicated Sonnet page, the launch note, and the plans screen on claude.com. The family still includes Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 on the marketing homepage. This review keeps the catalog to the Sonnet line listed on the Sonnet product page—5, 4.6, 4.5, and 4—so we are not mixing another class into the same grid. If you need days-long Fable runs or Opus-class cyber work, that is a different card on Claude.ai, not a silent upgrade hiding inside Sonnet 5.
Capabilities I can stand behind because they are named in product copy: writing and editing, study guides, coding (including Claude Code), research with connectors and citations, analysis of files and spreadsheets, artifacts you can share, Cowork for delegated documents, and computer use in a browser or desktop. Adaptive thinking is on by default. Platform docs list a 1M-token context window and 128K max output. I did not see an image generator or a video generator as a Sonnet 5 product, so this site has no fake picture wall.
Safety is part of the launch, not a footnote. Anthropic reports fewer undesirable agent behaviors than Sonnet 4.6, lower hallucination and sycophancy in their tests, and cyber safeguards on by default because Sonnet 5 is a bit stronger on routine cyber tasks than 4.6 while remaining far behind Opus 4.8 and Fable 5 on exploit work. I am not going to recap their charts as if I ran them. I will say this: a model that refuses a bad request cleanly is a feature I want in an agent that can click around a computer.
About this page. I am not Anthropic. This is an independent overview of Claude Sonnet 5. Accounts, billing, and the live model picker live on the official Claude Sonnet 5 website. Primary buttons on this site open a local note so we do not impersonate a login.
Scenarios
Jobs I would actually start with
These four match prompts Anthropic shows under “See Claude in action” plus Cowork on the Claude homepage. Each page is a short brief, not a second homepage.
Pull request review
A coding pass: what to inspect in a Python web-app PR before you merge.
Editorial calendar
A three-month weekly newsletter plan without stuffing empty slogans.
Study schedule
A finals-week template with timed papers and a sleep target.
Claude Cowork
Start a report at the desk, check it on a phone, review the artifact.
Sonnet line
Popular versions in this catalog
Names are spelled as on the official Sonnet page. I am not translating them. Sonnet 5 is the one I would open today; the earlier cards stay here because teams still pin evals to a frozen id.
Claude Sonnet 5
Most agentic Sonnet yet. Default on Free and Pro for coding, tools, and daily work.
Claude Sonnet 4.6
Hybrid reasoning with a 1M context window for agents and high-volume jobs.
Claude Sonnet 4.5
Agents, coding, computer use, and long-running tasks from the 2025 Sonnet line.
Claude Sonnet 4
Practical Sonnet for assistants and high-volume work, announced May 2025.
Product groups
How Anthropic groups the work
Four use-case headings from the Sonnet product page. If a group is not on that page, it is not in this nav.
Advanced coding
Plan, implement, debug, and refactor across multi-file codebases.
Long-running agents
Instruction following, tool choice, and course correction without stalling.
Browser and computer use
Navigate the same desktop and browser flows people already click through.
Enterprise workflows
Finance, research, documents, and operations at production volume.
Who it is for
Who I would send this to
Engineers and agent builders
You want a model that explores a repo, uses tools, and does not stop at the first green test. Start with advanced coding and the Claude Sonnet 5 card.
Ops and computer-use teams
Your work is already in a browser: intake, onboarding, vendor portals. Read browser and computer use before you buy a custom RPA story.
Writers and researchers
Long context, citations, artifacts, and Cowork one-pagers. I still verify numbers. The gain is structure, not magic facts.
Teams already on Claude.ai
Free and Pro default to Sonnet 5. Max, Team, and Enterprise can select it. If you are pinned to 4.6 for a benchmark freeze, keep the older card until the eval is re-run.
Questions
Questions I get after sending this page
Anthropic positions Sonnet 5 as the balance of speed, price, and intelligence for high-volume work: customer-facing agents, production coding, content at scale, and real-time research. I reach for it when I want agentic follow-through without jumping to a heavier Opus-class model for every turn. If the job is a one-line rewrite, Haiku-class speed may still be enough; if it is a days-long investigation with extra guardrails, the family page points you toward Fable or Opus. For the work I actually ship—patches, briefs, browser chores—Sonnet 5 is the default I describe in this review. Always confirm the live picker on the official site, because plan defaults can move.
On the Sonnet product page this month, API usage is listed at $2 per million input tokens and $10 per million output tokens, with prompt caching up to 90% off and batch processing at 50% off. US-only inference is called out at 1.1×. Chat access is on Claude.ai across Free, Pro, Max, Team, and Enterprise; Pro on claude.com was listed at $17 per month on an annual invoice ($20 month-to-month) when I opened plans. I am not going to pretend those numbers never change. For the current bill, use the official pricing page, not this overview.
The product chat lives at the official Claude Sonnet 5 website, https://claude.ai, also written Claude.ai on Anthropic’s own pages. Model cards, system notes, and the Sonnet announcement sit on anthropic.com. This review is a separate site with its own catalog. Primary buttons here open a local note so we do not pretend to log you into Anthropic. If you want the real app, use the text links to Claude.ai.
The June 30, 2026 note calls Sonnet 5 a substantial step over 4.6 on reasoning, tool use, coding, and knowledge work, with performance that can approach Opus 4.8 on some effort curves at a lower token price. Adaptive thinking is on by default; the platform docs warn that old manual thinking budgets and custom sampling now error. Tokenizer changes can map the same text to more tokens. I treat 4.6 as the previous daily driver with a 1M window, and 5 as the one I open first unless a team is pinned to 4.6 for a frozen eval.
Yes—those are named product lanes, not blog experiments. Advanced coding covers planning through refactors. Long-running agents cover tool choice and error correction. Browser and computer use is documented as navigating real desktop and web tasks (procurement, onboarding, competitive scans). Claude Code is the terminal/IDE companion; Cowork is the delegated artifact loop. I still review diffs. The model finishing a job is not the same as merging it.
No. I write an independent overview. Anthropic did not publish this page. Names, prices, and model lists are taken from pages I opened on anthropic.com, claude.com, and the Claude platform docs. For accounts, billing, and the live model picker, go to the official Claude Sonnet 5 website. Nothing here is a login, a checkout, or a support desk.
Reader notes
Five reader notes on this overview
Five reader impressions of this review, not official product testimonials.
5 of 5
I pointed it at a brownfield service and it wrote a reproducing test before the patch. That follow-through is why I keep Sonnet 5 as the daily driver.
4 of 5
Cowork assembled Monday’s one-pager from analytics without me babysitting every step. I still check the numbers, but the outline is ready before standup.
5 of 5
For newsletter calendars and long edits it stays in voice better than the previous Sonnet I used. I do not ask it to invent facts.
5 of 5
Computer-use runs on the same tools our ops team already clicks through. It finishes two-part jobs that used to stall halfway.
4 of 5
Study guides from PDFs are structured enough to share. When a citation is thin, I send it back rather than publishing the first pass.
When you are done reading, continue on your terms
Keep this tab as the map. Open the product only if the jobs above match the week you actually have.