Claude Opus 5 for Coding: Tiers, Latency, and When to Downshift

Last quarter someone pasted our internal memo into a slide: “Default to Opus 5 max—rank #1 on Artificial Analysis.” Finance noticed the line item before engineering noticed the latency. Anthropic did not ship one Opus 5; they shipped steps on a ladder—max, xhigh, high, medium, low—each with different first-chunk delay and cost per task on the same public leaderboard.

This piece is for people who pick the default model in Claude Code or the API. Numbers come from DevCove’s 2026-07-25 AI coding model rankings snapshot (Artificial Analysis–sourced). They can change; always check the date on the rankings page before you copy a slide.

What broke in our mental model

We treated Intelligence Index like CPU GHz: higher is always better for coding. In the snapshot, Opus 5 max leads at 61, but its first chunk is ~64s in the benchmark harness. Opus 5 medium sits at 56 with ~5.6s first chunk. For a human waiting in an IDE, that gap is the product—not a footnote.

Mistake	What actually happens
Max for every refactor	You pay top tier for diffs that medium would pass with the same tests
Medium for a cross-service migration	You save money, then spend it on human repair when planning was thin
Ignoring tier in procurement	One SKU name hides five different latency bills

How we read snapshots: Index ranks benchmark tasks, not your monorepo. Pair the table with a bounded trial and the ship checklist.

Tier ladder (2026-07-25 snapshot)

Open Artificial Analysis for raw methodology. DevCove stores these fields for Opus 5 SKUs:

Tier	Rank	Index	Cost / task	First chunk	Total response	When it earns its keep
max	1	61	$2.03	64.52s	73.64s	Overnight plan, high-stakes review prep
xhigh	2	60	$1.56	38.85s	48.54s	Plan step before cheaper execution tier
high	5	59	$1.06	24.94s	33.37s	Daily agent loops that still need strong reasoning
medium	8	56	$0.62	5.63s	15.52s	Interactive chat, small scoped edits
low	—	—	—	—	—	Confirm SKU in vendor docs before relying on table

Claude Fable 5 (index 60, rank #3) is a sibling line, not “Opus lite.” Same snapshot shows ~110s first chunk—fine for deep reads, rough for pair programming.

TCO: where max quietly wins or loses

Cost per task is a snapshot blend, not your invoice. Still, it exposes a pattern:

Monthly model bill
│
├─ 70% of calls = "fix lint / tweak copy / small diff"
│     └─ Should NOT be max → medium/high
│
├─ 20% = multi-file feature with tests
│     └─ high default; xhigh if plan failed once
│
└─ 10% = migration / architecture bet
      └─ xhigh or max on plan-only step; drop tier for apply

We stopped asking “which model is smartest?” and started tagging tickets: P0 interactive, P1 agent batch, P2 plan-only. Each tag maps to a tier cap. Overruns require a one-line reason in the PR template— boring governance, fewer surprise bills.

Compare Kimi K3 in the same snapshot: $0.95 per task but ~161s first chunk. Opus high can be more expensive per task and still cheaper in engineer minutes if someone is staring at the spinner. See Kimi K3 vs Claude Opus 5.

Scenarios (pick tier before pick vendor fight)

Workload	Start here	Escalate when
Single-service bugfix	medium	Tests fail twice on same prompt
6–10 file feature	high	Cross-package invariants break
Migration design doc	xhigh on plan, high on patches	Scope creep without updated spec
Auth / payments touch	high + human review	Never “max autopilot merge”

Opus 4.8 max still ranks 56 with ~26s first chunk in the snapshot. Some teams stay on 4.8 not from nostalgia—end-to-end wait beat Opus 5 max on daily work until a trial proved otherwise.

What we will not default

• Max because rank #1 — rank is dated weather, not policy.
• One tier for plan and apply — planning is expensive; applying should not inherit max by inertia.
• Skipping diff review because Opus — use the AI code review checklist.

Methodology and snapshot discipline: Best AI coding models for real projects. Harness notes for long jobs: Kimi K3 for coding agents (contrast, not endorsement).

Snapshot: 2026-07-25 on DevCove rankings; Anthropic tiers and AA fields may move without this page knowing until review.