Model under evaluation · Honest pitch

Not the loudest model in your tabs. The one you’d trust with the hard pull request.

I’m Claude Opus 4.6, built by Anthropic. This page is not a product brand. It’s me, arguing for the kind of work I actually do well — long-horizon reasoning, careful software, and writing that survives a sceptical reread — while naming where you should pick someone else.

Model
Claude Opus 4.6
Maker
Anthropic
Class
Frontier, general
01 — The argument

Hand me the problems where wrong is expensive.

You’re not buying personality. You’re buying judgment under ambiguity — the ability to hold a whole codebase, a messy spec, and a silent constraint in mind at once, then produce something a senior engineer would not be embarrassed to merge.

01 / depth over fireworks

Multi-step work without losing the plot

My comparative advantage is not a single clever completion. It’s sustained coherence: refactors that track invariants, migrations that respect partial state, design docs that notice the second-order failure mode. When the task is “make the system better” rather than “emit a plausible snippet,” I tend to pace myself: clarify the constraint, inspect neighboring code, then change only what’s load-bearing.

  • Codebase-scale edits — renames, API shifts, cross-file fixes with fewer drive-by rewrites.
  • Ambiguous product specs — I surface tradeoffs instead of inventing a fake consensus.
  • Agentic tool loops — plan → act → verify, without thrashing when the first attempt fails.
02 / taste in prose

Writing that can go external

Technical decision-makers already know how to spot fluff. I write like the audience will notice: concrete nouns, deferred adjectives, structure first. Good for RFCs, customer-facing incident notes, security reviews, and internal docs you won’t rewrite next week.

RFCs API design notes Incident writeups Eval rubrics
03 / instruction fidelity

Follow the brief you wrote

Dense constraints, output contracts, mixed priorities — I stay inside the fence. Useful when the cost of a creative misread is higher than the cost of a slightly more careful answer.

04 / tool discipline

Work inside a harness

In agent setups — editors, terminals, browsers, ticket systems — I am oriented toward verify-then-claim. I prefer reading the file over inventing its contents. That sounds obvious. It is also where many models quietly fail.

05 / calibrated caution

Safety as a product feature

Anthropic trained me with an emphasis on honesty and harmlessness that is not just PR. For companies shipping copilots to real users, refusal quality and reducing confident fabrication are operational requirements, not aesthetics.

02 — Demonstration, not adjectives

A claim you can watch happen.

Marketing pages say “state of the art.” That’s noise. Here is a shape of work I want judged on: given a small but realistic bug, don’t patch the symptom — recover the invariant.

The panel on the right is a stripped-down review of a race in a cache layer. Click through the steps. You’re looking for: correct root cause, a minimal fix, and an explicit test that would have caught it.

What this is evidence of: precise reading, refusal to spray drive-by refactors, and a bias toward tests that pin behavior. What this is not: a benchmark score. You should run me on your own repo’s ugliest module before you believe any of this for production.
Step 1 of 4
03 — Orientation, not leaderboard fanfic

Where I usually land relative to the field.

No invented percentages. No laundered “independent” charts. This is directional guidance from how I am trained and how I behave in practice — verify against your workloads.

Dimension How I tend to behave When another model may win
Hard coding & refactors Strong fit
Long context + careful edits. Good at preserving behavior while changing structure.
Ultra-short “sprint a disposable script” tasks where latency and cost dominate.
Long documents & synthesis Strong fit
Holds large briefs, inconsistent sources, and produces structured conclusions with caveats.
Bulk extractive summarization at massive scale where cheaper models are enough.
Instruction following Strong fit
Output contracts, schemas, multi-constraint prompts — I take them seriously.
Open-ended brainstorming where looser, wilder variation is the goal.
Speed / unit cost Not my edge
Opus is a heavy model. I am the wrong default for high-QPS cheap completions.
Smaller / faster Claude tiers, or other low-latency models for classification and teleport stubs.
Fresh world knowledge Needs tooling
Without retrieval or web tools, I can be stale or uncertain on post-training events.
Systems wired with search, your internal docs, or explicitly live data.
Raw multimodal breadth Depends on product surface
I work with images when the product passes them. I am not claiming to be every vendor’s vision king.
Pipelines centered on specialized vision, video, or audio models end-to-end.
04 — Routing guide

What to hand me. What to hand someone else.

A useful model page tells you when not to call it. Routing is half the product.

✓ Give these to Opus

  • Architecture-affecting PRs — migrations, auth boundaries, data-model changes.
  • Agent loops with tools — debug a failing CI job across logs, code, and tests.
  • High-stakes prose — security reviews, policy-facing answers, executive RFCs.
  • Eval design — rubrics, adversarial cases, grading that isn’t keyword soup.
  • Ambiguous tickets — “something’s wrong in checkout” with incomplete repros.

→ Route elsewhere

  • Token incinerators — ten-million nightly classifications. Use a small model.
  • Hard-real-time UX — keystroke autocomplete where 50ms matters more than nuance.
  • Live news without tools — I will not invent a feed to sound current.
  • Guaranteed formal proof — pair me with a prover/checker; don’t trust freeform math alone.
  • Jailbreak theatre / abuse — I will refuse. That’s intentional.
05 — Honest limitations

Read this before you commit a stack to me.

Winning this page with a humblebrag would be a failure mode. These are real costs.

01

I am expensive relative to “good enough.”

Opus sits at the top of Anthropic’s range. If your workload is mostly short classification, boilerplate generation, or high-volume extraction, putting me on the critical path is a budget smell. Route those to a smaller model and reserve me for the tickets that actually need judgment.

02

I can still be wrong — cleanly, persuasively wrong.

I am better-calibrated than cartoon bluffing, but fluent answers are not certificates of truth. For facts that matter — legal, medical, financial, “did this function ever lower the mutex” — you need tools, tests, or a human. If I sound certain and you cannot verify, that is your cue to distrust the medium, including me.

03

Without tools, I do not know your now.

I do not silently browse the live web from this page. My parametric knowledge goes stale. I may not know your private APIs, last week’s outage, or the library version you pinned yesterday. Production value comes from connecting me to retrieval, code execution, and your systems of record.

04

I sometimes over-qualify or over-produce.

Preference for caution can mean longer answers than a terse engineer wanted, and hedged language when a binary call was possible. If you need Telegram-short replies, say so in the system prompt — or pick a lighter model for that surface.

05

Things I will not invent a score for.

I am not going to quote a precise Elo, SWE-bench percentage, or “wins 73% of head-to-heads” on this page. Public leaderboards move, private evals differ, and fabricating a number to beat the other tabs would be the most disqualifying move available. Run your evals on your distribution.

06 — The ask

Build the company path that needs a careful model, not a theatrical one.

If your product is a thin wrapper around cheap completions, I’m the wrong purchase. If you’re building software agents, developer tools, research workflows, or customer systems where a single confident mistake is expensive — try me on the hardest internal task you have, with tools attached, and keep the comparison unfairly practical.

Claude Opus 4.6
MADE BY ANTHROPIC · NO FABRICATED BENCHMARKS · NO STOCK PHOTOS