Anthropic Claude
Why I use it for technical work where reasoning depth matters more than speed
I use Claude every day. This site was built with Claude Code. The newsletter system runs on it. This very page was written with it. That’s not a marketing claim — it’s the context you need to calibrate this evaluation. What follows is what I’ve actually learned from daily production use, not a feature comparison table.
Why Claude works
Four reasons hold up across daily use.
Reasoning depth and extended thinking
Opus and Sonnet handle multi-step problems coherently. Extended thinking mode (available on Opus) shows you the chain of thought as the model works through a problem. That visibility matters: you can watch it reason toward an answer, catch a wrong assumption mid-path, and correct it before it propagates into a code change or a document section you then have to undo.
The extended thinking chain on a complex multi-file refactor caught a circular dependency I had missed across three sessions of review. The model found it before the code ran. That's the kind of thing extended thinking earns its keep on.
Long-context handling that's actually usable
200K tokens on Sonnet, 1M on the Opus 1M context variant. That isn't just a spec sheet number. Claude maintains coherence and recalls details from early in a long context without the "lost in the middle" failure mode that hits other models. Feed it an entire codebase, a long technical spec, or a transcript of a week of design discussions: it holds the thread.
I've fed the full build script plus template files plus the last ten session log entries into a single context window and asked Claude to reason about an architecture decision that touched all of them. The answer accounted for constraints buried 80K tokens earlier. No hallucination, no dropped context.
Claude Code: AI as a real coding partner
Claude Code is the CLI tool that turns Claude from a chat interface into a competent pair programmer. It reads your codebase, proposes and makes coordinated multi-file edits, runs tests, and commits cleanly. The key word is "coordinated": it understands that changing a function signature in one file means updating every caller, and it does that without being told explicitly.
The packetloss.tech build system, the newsletter publication pipeline, and this vendor page were all worked on through Claude Code. A typical session covers 5–15 file edits in one pass with no broken cross-references. That's a different category of tool than a chat window where you paste snippets.
Honest tradeoffs in the responses themselves
Claude pushes back when prompted with bad ideas, flags uncertainty, and admits when it doesn't know something. The behavior is engineered: it comes from Constitutional AI training, not from tuning a chat interface to be agreeable. The practical effect is fewer confident-but-wrong answers, and less cleanup after an AI-assisted session that went confidently in the wrong direction.
Twice in the past month Claude refused a refactor I asked for and explained why it would create a race condition in a different module. Both times it was right. An AI that tells you no when you're wrong is worth more than one that does what you ask.
Where it fits best
Not every use case. The fit is sharpest when one of these describes your work:
The sharpest entry point. Claude Code turns an AI subscription into a coding partner that knows your codebase, not just the snippet you pasted. Teams doing active feature development, code review, or refactoring work get the most out of it.
Legal contracts, technical specs, compliance documentation, RFP responses. The long-context window means you don't have to chunk the document and stitch answers back together. Feed the whole thing, ask the real question.
Where reasoning matters more than speed. A model that can hold a full policy document plus the relevant control framework plus the exceptions log in a single context window, and reason across all three, is genuinely useful for audit prep work.
Network engineers, technical consultants, independent analysts: anyone whose output is words and decisions rather than raw code volume. Claude functions as a thinking partner that improves with the quality of the question, not just a faster search engine.
If you need image generation, Google Gemini or other multimodal tools cover that better. If you need the cheapest token cost for high-volume conversational use, OpenAI’s lower tiers compete.
The honest tradeoffs
Marketing won’t print these. I’ve run into all of them. Tap to expand.
CostSonnet and Opus API pricing is real money at scale
Sonnet 4.x API runs approximately $3/MTok input, $15/MTok output. Opus is roughly 5x that. The Pro subscription at $20/mo is a reasonable on-ramp for exploration, but serious daily use with extended thinking or long-context workloads will hit the subscription ceiling fast and push you to the API, where a busy engineering team can easily clear $100+/mo per engineer. Set up cost monitoring before you build production workflows. The Anthropic console's usage dashboard is your first line of visibility; it's not as polished as AWS Cost Explorer, but the numbers are accurate.
Rate limitsAPI tier limits constrain throughput at scale
Free tier is generous for exploration; the paid API tiers scale with your spend history, but Opus is meaningfully slower per request than Sonnet or Haiku. For production workloads that need low-latency responses on complex prompts, the choice between Opus depth and Sonnet speed is a real architecture decision, not a checkbox. Plan the model tier into your design, not as an afterthought. The failure mode is building on Opus for the quality, discovering the latency in production, and having to rewrite the integration to use Sonnet with prompt engineering to compensate.
Image and audio gapClaude is text-first: no image generation, no audio
Claude reads images as input (vision capability exists on Sonnet and Opus). It does not generate images. Audio is not supported in either direction. If your workflow requires image generation, diagram creation from scratch, or voice interfaces, Claude is not the right tool for those specific jobs; you need a second tool for those outputs. This isn't a criticism; it's a scope decision Anthropic made deliberately in favor of reasoning depth over multimodal breadth. Know it going in so you don't build an integration that hits the wall six weeks later.
Knowledge cutoffTraining data has a cutoff; recent accuracy requires verification
Like every large language model, Claude's training data has a cutoff date. Web search exists via Claude Code tool-use integrations and via the API's tool-use capability, but it isn't built into the core conversation by default. For time-sensitive work (recent CVEs, latest firmware versions, a vendor's current pricing), the model will tell you its knowledge boundary if you ask, but it won't always volunteer the caveat. Treat Claude as extremely deep on reasoning and broad on knowledge up to its cutoff, and pair it with a live source for anything where recent accuracy matters.
Claude is the model I reach for when the answer matters more than the speed. If you need quick autocomplete or pattern-matching, lower-tier models are cheaper. If you need reasoning that holds up under review, this is the conversation.
Is it right for your team?
Four dimensions to check before you commit:
- Size: Solo through enterprise. Per-seat economics on the Pro subscription work for individual professionals; the API scales to teams without per-seat overhead. The model tier decision (Haiku for volume, Sonnet for daily work, Opus for deep reasoning) is more important than the team size question.
- IT maturity: Any technical team can self-onboard via claude.ai with no infrastructure required. Engineering teams get the most value from Claude Code, which requires someone comfortable with CLI tools and willing to iterate on how they prompt and constrain the tool.
- Existing stack: Tool-agnostic. Claude integrates via API into any codebase or workflow. The API is clean, the SDKs are well-maintained, and the documentation is honest about what the model can and can’t do. No lock-in at the infrastructure level. The lock-in that does exist is cognitive: once your team builds workflows around Claude, moving to a different model means re-tuning all the prompts.
- Geography: claude.ai and the API are available globally with some country restrictions. LATAM access works across most countries; the console and API are reachable from Colombia, Mexico, Peru, Argentina, and most of the region without workarounds. Check the Anthropic country availability page for the current list.
Who implements it
Anyone technical can self-onboard at claude.ai. The free tier is a real product, not a bait-and-switch. It’s genuinely usable for daily work and gives you an accurate read on what Claude is good at before you commit to a subscription or API integration.
For engineering teams adopting Claude Code: designate someone to own API key management, usage monitoring, and usage policies. There’s no formal certification program. The community resources and the official Claude Code documentation are extensive enough that most teams get productive without external help. Where external help is worth it is in prompt engineering for specific production workflows: getting consistent output from a complex, long-context pipeline is a skill that benefits from iteration and expertise.
If you’re evaluating Claude for a technical consulting or security workflow and want a second opinion on how to structure the integration, let’s talk. I run Claude on real work daily and can tell you in 30 minutes whether your use case fits what the model is actually good at.
First steps
- Start at claude.ai (free or Pro $20/mo) before committing to the API. Get a feel for what Claude is good at on your actual work: not generic demos, but the documents, code, and reasoning tasks that show up in your week. The free tier is honest about its limits; the Pro tier gives you enough usage to build real intuition before you architect integrations.
- For engineering teams: install Claude Code. It's the sharpest entry point — turns Claude from a chat tool into a coding partner that knows your codebase. The install is straightforward, the CLI is well-documented, and the first real multi-file session will tell you more about the tool's fit than any amount of reading about it:
- Install via npm:
npm install -g @anthropic-ai/claude-codeand authenticate with your Anthropic API key. - Start in a real project. Not a toy repo. A real codebase with real complexity is where Claude Code earns its keep or doesn't.
- Read the CLAUDE.md convention. Adding a CLAUDE.md to your repo root tells Claude Code about project conventions, banned patterns, and context it should carry across sessions. That one file changes the quality of sessions significantly.
- Install via npm:
- Set up cost monitoring before production workloads. API usage scales fast. The Anthropic console's usage page shows spend by day and by model. Set a monthly budget threshold before you build anything that calls the API automatically. Think of it as the AWS Budgets equivalent for token spend. Monthly limits before production workloads land are not optional.
Beyond first steps: I use Claude on consulting work, on this site, and on the newsletter system daily. If you want a read on whether your specific use case is a Claude job, a Gemini job, a GPT job, or a “fix your workflow before adding AI” job, talk to me. Thirty minutes is enough to find out.