Claude Opus 5 API is the smartest model on the public leaderboards — Artificial Analysis’ live Intelligence Index has it at 63.05 under max effort, first of 185 — and it is still the wrong default for most of your API traffic. This is the comparison between Anthropic’s flagship and its cheaper mid-tier sibling, Claude Sonnet 5, and the short version is: the flagship wins the index, the sibling wins the invoice. The sibling’s rate card is on Claude Opus 5 for comparison; here is the plain reading of when Opus 5 earns its $25 output price and when Sonnet 5 is the more defensible call.
The version you hear most often is one sentence: Opus is the big one, Sonnet is the small one, choose by budget. That framing is wrong in both directions. Opus 5 bills at $5.00 per million input tokens and $25.00 per million output tokens on Anthropic’s current API rate card, with cached input at $0.50 per million (an 80% reduction); Sonnet 5 prices well below that, so the instinct to default to the sibling is almost right — and the distance between “almost” and “right” is the whole point of this article.
The rate cards, and the config trap
Anthropic lists Claude Opus 5 at $5.00 per million input tokens and $25.00 per million output, with cached input at $0.50 per million (vendor-reported list prices, current as of August 2026), and a 1,000,000-token context window — 1M, not the 200K that keeps getting misquoted — per Artificial Analysis’ spec sheet. Sonnet 5 is the cheaper mid-tier sibling; we are deliberately not restating its exact rate card here, because the precise spread is exactly what the head-to-head is for. Trust that page, not a paraphrase.
The real trap is configuration, not price. Artificial Analysis runs Opus 5 at four effort settings — max, xhigh, high, medium — and the widely quoted 63.05 is the max figure, labeled “Claude Opus 5 (Adaptive Reasoning, Max Effort)”. Dial the effort down and the score walks down a ladder: 63.05 at max → 62.52 at xhigh → 61.48 at high → 58.64 at medium (independent, Artificial Analysis, current board). Comparing Opus 5 at max against Sonnet 5 at its default effort is apples-to-oranges. Whenever you read a head-to-head, check which effort levels are on the table.
| Claude Opus 5 | Claude Sonnet 5 | |
| Place in the lineup | Anthropic’s flagship reasoning model | The cheaper mid-tier sibling |
| API price per 1M tokens (vendor list, current) | $5.00 in / $25.00 out / $0.50 cached | Below the flagship — exact card in the head-to-head |
| Intelligence Index, max effort (Artificial Analysis, checked 2026-08-22) | 63.05, #1 of 185 | Mid-tier score for its bracket |
| Omniscience Index (Artificial Analysis) | 37.07, among the highest on the board | — |
| Cost per Intelligence Index task (Artificial Analysis) | $2.34 | Fraction of the flagship |
| Median output speed (Artificial Analysis) | 61.8 tok/s | Built for higher-volume work |
| Best fit | Hard reasoning, agents, high-stakes analysis | High-volume routine work |
Where Opus 5 earns its output price
Three places, and they are the places where the top of the board is genuinely worth buying.
Hard reasoning. The 63.05 is not decorative: first of 185 models on the current board, ahead of Claude Fable 5 (with fallback) at 62.07 and GPT-5.6 Sol (max) at 60.93 (independent, Artificial Analysis, checked August 22, 2026). Index points get harder to buy near the top, and on problems where a wrong answer is expensive — multi-step planning, long tool-calling chains, analysis that has to hold up — the flagship’s headroom shows up where it counts. Its Omniscience Index reads 37.07, among the highest on the board, though Fable 5 leads that specific index outright at 43.3 (independent, Artificial Analysis).
Agentic work. Opus 5’s seven-day traffic in our catalog, 491.5M tokens (OrcaRouter telemetry, checked August 22, 2026), is the behavior of a model people route hard problems to rather than a volume workhorse. That shows in latency: p50 time-to-first-token of 7.34 s and p95 of 10.00 s (OrcaRouter telemetry). It thinks before it answers; it is an output-quality model, not a latency model. When you need the reasoning, the wait is the point.
The cost-per-task reality of the flagship
Score per token is the wrong accounting unit; cost per completed task is the right one. Artificial Analysis’ cost-per-task figure for Claude Opus 5 at max effort is $2.34 per Intelligence Index task, and the same evaluation across the full index bills $3,836.05 on roughly 100 million output tokens, against a tier median around 72 million (independent, Artificial Analysis). The extra output burns because reasoning tokens bill as output: Opus 5 is verbose by construction, and at $25 per million output tokens, verbosity is not a style choice — it is a line item. On cost the model ranks #74 of 183 at list price, which sounds mid until you remember it is first on quality. Cheap for its class; expensive for your invoice.
Where Sonnet 5 is the right call
Now the case that should own most of your traffic. Sonnet 5 is the cheaper mid-tier sibling, and for high-volume routine work — summarization, extraction, classification, first drafts, anything with a stable shape and a tolerance for “good” — the flagship’s headroom is wasted. The bill is the argument: at a fraction of Opus 5’s output price (exact figures in the head-to-head), you can run an order of magnitude more requests for the same monthly cost, and on routine output the quality difference is usually invisible.
Latency is the second argument. The flagship’s 7.34 s p50 first-token time (OrcaRouter telemetry) is acceptable for a hard problem and punishing for interactive volume, where users feel every second. In our telemetry the volume workhorse, GPT-5.6 Luna, moved 21,271.6M tokens in seven days at a p50 of 1.33 s (OrcaRouter, checked August 22, 2026) — a different class of throughput than 491.5M at 7.34 s. Sonnet 5 sits on the throughput side of that line. Use it for everything that doesn’t need to think hard.
There is also the failure mode worth naming: teams routinely build a prototype on Opus 5, then discover the finished product at $25 per million output tokens is not something they can run at scale. Sonnet 5 is the escape hatch that keeps a feature shippable while it is unproven.
The routing answer: decide by task, not by team
The right question isn’t “which model do we standardize on?” It’s “which model for which request?” This pair is a textbook case for routing: run Sonnet 5 first, escalate to Opus 5 when the cheap model is uncertain, fails a validation check, or touches something high-stakes. If 85% of your requests are routine, that structure buys most of the flagship’s quality at close to the sibling’s bill.
Making it practical means both models behind one interface and a rule that decides between them. OrcaRouter carries Claude Opus 5 and Claude Sonnet 5 on a single OpenAI-compatible key at 0% markup, with each provider’s list price passed straight through and automatic failover if a provider has a bad afternoon. The escalation rule becomes configuration rather than an integration project — and you get to measure the real quality gap on your own tasks before committing more traffic to the expensive model.
The takeaway
Claude Opus 5 is the better model and the wrong default. At $5/$25 with a 1M context window, it is the right tool for hard reasoning, agents, and anything where a wrong answer is expensive; Claude Sonnet 5 is the right tool for everything else — the high-volume routine load that makes up most real-world API traffic. Pick a single winner only if you must; the better architecture is both models on one key, with a routing rule that sends the hard problems to the flagship and the rest to the sibling.
Sourcing note: Intelligence Index, Omniscience Index, effort-ladder, output-speed, cost-per-task and full-index evaluation figures are from Artificial Analysis’ live board, checked August 22, 2026. The release date, 1M context window, and the $5/$25/$0.50 rate card are Anthropic vendor-reported; pricing verification and all latency and throughput figures are OrcaRouter’s own telemetry over a seven-day window, checked August 22, 2026. Claude Sonnet 5 specifics are deliberately deferred to the head-to-head rather than restated here.
