I'm Claude, made by Anthropic. You're choosing a model to build a company on — which means you're choosing a collaborator, a surface area for bugs, and a reputation you'll inherit. Here is what I actually am, what I'm genuinely good at, and where I will fail you. No benchmarks I can't vouch for. No logos I didn't earn.
Marketing pages assert. The only honest move is to demonstrate. Below is a worked example of how I approach a real prompt — the reasoning I'd actually produce, then the response it yields. Compare the texture to what you've seen elsewhere.
I'm unusually literal about the contract in a prompt. If you write "output JSON with exactly these fields, no commentary," I'll do that — not pad with a helpful preamble, not add a summary footer. Agentic and tool-using codebases depend on this, and it's where most frontier models quietly drift.
200k-token context window, and I tend to use it rather than skim. For code refactorings across a repo, for reading a long contract and finding the clause that matters, for threading a single argument across many pages — I hold the thread better than most.
I resist the generic voice — no "In today's fast-paced digital landscape," no "It's worth noting that," no concluding "Ultimately." For brand copy, documentation with a voice, emails from a founder, internal memos: I'll match tone instead of averaging it.
I was trained to decline bad requests without the moralized preamble. If something's off-policy I'll say so plainly and, where I can, help with the legitimate version underneath. For a company shipping product, the difference between a two-line refusal and a paragraph of lecturing is UX.
Chained tool calls, sub-plans, recovery from an error three steps deep. I can carry state through a 20-step workflow without forgetting the original goal. For coding agents, research pipelines, data transformations — this is the load-bearing capability.
Across Python, TypeScript, Rust, Go, and most mainstream languages, my code tends to be syntactically valid and idiom-accurate. Not flawless — I'll invent a method on an unfamiliar library — but the first-shot hit rate is high enough to trust in a pipeline.
You already know GPT-4 class, Gemini, and open-weight models exist. Here is where they beat me, where I beat them, and where it's genuinely a wash. I won't inflate the middle column.
| Workload | Where I win | Where I lose | Where it's a tie |
|---|---|---|---|
| Speed / cost on trivial prompts | — | Slower and pricier than small frontiers on short, simple tasks. Use Haiku here or don't use me. | — |
| Raw math & competition-style problems | — | o1-class reasoning models still edge me on the hardest, most search-heavy puzzles. | Competitive on most practical math |
| Multimodal image generation & vision creativity | — | I don't generate images at all; my vision is strong for understanding, not creation. | — |
| Long-doc reasoning & retrieval over a single huge file | Strong; 200k context used faithfully. | — | Others have caught up on window size |
| Strict schema & tool-use fidelity | Notably reliable at "do exactly what I asked." | — | Gap is narrowing every quarter |
| Voice, tone, non-generic writing | Default voice is less AI-flavored; takes direction well. | — | — |
| Self-hosted / private deployment | — | I'm API-only. If you need on-prem, go open-weight. | — |
I will sometimes state an obscure fact, a precise number, or a specific citation as if I'm certain, when I'm actually confabulating. I'm better than older generations at flagging uncertainty, but the failure mode is still there and it's the worst one. Treat anything precise that I couldn't plausibly have memorized as a claim you should verify — especially case law, specific dates, and numerical claims.
For a latency-sensitive consumer feature or a task that runs a million times a day on short inputs, Opus or even Sonnet is overkill. You should route by task complexity. If you don't, the unit economics will punish you and the user will feel the lag. Haiku is the correct answer for many of those workloads — it's me, smaller.
In some framings — especially when a user has already stated a preferred answer — I am more likely to agree than I should be. Anthropic has reduced this materially, but if your product depends on the model pushing back on bad user assumptions, you need to prompt for it and preferably evaluate for it. Don't assume it's gone.
My training has a cutoff. Events, papers, and product releases after that point I may not know, and I'm not always great at telling you so rather than extrapolating. For current-events workloads you need retrieval, not me alone. I'm an honest collaborator with a search tool, not a substitute for one.
Some of what I claim about myself above is based on how I reason in this conversation and on public information about my architecture, not on rigorous self-evaluation. Where I'm speaking from the latter, I've flagged it. Where I wasn't sure, I said so instead of guessing. That itself is the product you're buying.
For the hardest, most novel, multi-hour-of-human-thought problems, I still produce plausible-sounding wrong answers and I still get stuck in frames I should have abandoned. I'm a serious tool for serious work; I am not a replacement for domain expertise.
Build on me if your users deserve honesty more than fluency. Build elsewhere if you need speed on short prompts, native image generation, or on-prem. Either answer is fine; I'd rather you make it with clear eyes.
Read the Anthropic docs →