Claude Opus 5.5 vs Claude Sonnet 5: Half the Price, Twenty Index Points, and a Five-Month Blind Spot

If you only want the rule: run Claude Sonnet 5 by default and escalate to Claude Opus 5.5 when a task fails, because the cheaper model is genuinely half the price and genuinely a different capability class. The two numbers behind that are 38 and 58 — Sonnet 5's and Opus 5.5's scores on the Artificial Analysis Intelligence Index, on the same evaluator, in the same configuration family. A twenty-point gap on that index is not a tiebreak, and it is the largest separation in any Claude-versus-Claude matchup you can currently build: Claude Opus 5.5 is ranked #1 of 210 models and Claude Sonnet 5 is ranked #54, and the price difference between them is $2 per million input tokens.

That is the easy half. The hard half is that escalating selectively requires knowing which tasks fail, and the two models disagree about the world after January 2026 in a way no eval will surface as an error.

The twenty-point gap, in context

Artificial Analysis runs each model in one published configuration, and these two share a family label — "Adaptive Reasoning, Max Effort" — which makes the comparison unusually clean:

• Intelligence Index — Claude Opus 5.5 58, ranked #1 of 210 vs Claude Sonnet 5 38, ranked #54

• Price — Claude Opus 5.5 $4.00 in / $20.00 out per million vs Claude Sonnet 5 $2.00 / $10.00

• Output speed — Claude Sonnet 5 79.3 tokens/sec vs Claude Opus 5.5 not yet published on the board

• Time to first token — Claude Sonnet 5 150.02s vs a board median of 3.83s

• Default effort — Claude Opus 5.5 medium vs Claude Sonnet 5 high

• Knowledge cutoff — June 2026 for Claude Opus 5.5 vs January 2026 for Claude Sonnet 5

• Context and output — 1M context and 128K max output on both

The rank column is the part worth sitting with. Twenty index points is not four places on a leaderboard; it is forty-nine. On a board of 210 models, Opus 5.5 and Sonnet 5 are not neighbours with a price difference — they are in different bands, and the $2-per-million gap is buying a real capability step rather than a badge. Anyone who read this week's coverage as "Anthropic shipped a cheaper Opus" and assumed the Sonnet tier absorbed it should look at that rank pair: Anthropic's own pricing page still lists Claude Sonnet 5 as a current product, and its position on the independent board has not moved.

Two caveats keep this honest. Both scores were run at max effort, and neither model's shipping default is max — Sonnet 5 defaults to high, Opus 5.5 to medium. And the speed rows complicate the story in Sonnet 5's favour: 79.3 tokens per second with a 150-second first token is a real, measured latency profile, and it is the only one of the two that exists on the board. Opus 5.5's speed and latency have not been published by Artificial Analysis, which means the "faster" claim in this matchup currently rests on vendor material rather than a measurement.

The footnote that changes Sonnet 5's price

The most concrete thing to come out of the past month for Claude Sonnet 5 is a sentence most coverage skipped. When the model launched on June 30, 2026, its $2 / $10 rate was announced as introductory pricing through August 31, with a scheduled increase to $3 / $15 on September 1. On Anthropic's current pricing page, the footnote reads that the $2/$10 pricing "is now the standard price" and that the increase "will not occur."

That matters for two reasons. The obvious one is that a 50% increase that was announced and then cancelled is a cost line you can now plan against — Sonnet 5 is not a promotional rate you have to model an expiry for. The less obvious one is that it fixes the escalation arithmetic. With Sonnet 5 permanent at $2/$10 and Opus 5.5 at $4/$20, the ratio between the two tiers is exactly 2x on both directions, and it is stable. An escalation rule that sends, say, a fifth of your traffic to Opus 5.5 has a cost you can compute once and stop recomputing.

The batch and cache rates move together, which is what makes the whole-table comparison useful: Sonnet 5 batches to $1.00 / $5.00 against Opus 5.5's $2.00 / $10.00, and both bill cache reads at $0.20 per million. That last figure is a genuine tie — Opus 5.5 cut its cache read from $0.50 to $0.20, and Sonnet 5 was already there. If your workload is dominated by re-reading a large cached context rather than generating fresh output, the price advantage of the cheaper model is much smaller than the headline ratio suggests, because the one line where they agree is the line you spend most on.

The cutoff is the part that fails quietly

Claude Opus 5.5 carries a knowledge cutoff of June 2026. Claude Sonnet 5's is January 2026. That is five months of difference, and it is the one dimension in this comparison that does not announce itself. A model that does not know about an event is not slower or less accurate on it in a way your error handling catches — it answers confidently from a world that has moved on, and the failure looks like a plausible wrong answer rather than a refusal.

For some workloads this is irrelevant: refactoring a codebase whose conventions are stable, summarising documents you supply, transforming structured data. For others it is disqualifying on its own, regardless of the twenty index points or the price. Anything touching software released in the second half of 2026, anything answering questions about recent events, anything depending on an API or a library version that changed after January — those tasks cannot be escalated to Opus 5.5, they have to start there. The five months are not a quality difference; they are a scope restriction, and it belongs in the routing rule rather than in the evaluation.

Both models share the rest of the envelope, which limits how much the cutoff can be worked around. Both take 1M tokens of context and return up to 128K, both run adaptive thinking that cannot be switched off, and both expose depth only through the effort parameter. There is no configuration in which Sonnet 5 is handed a longer document to compensate.

Running the split

The practical version of this matchup is a default and an exception, and the exception has to be defined by something other than "hard tasks." The pattern that holds up is by task class rather than by difficulty: Sonnet 5 for the high-volume work where its cutoff is irrelevant and its latency profile is acceptable, Opus 5.5 for anything current, anything long-horizon, or anything where a failed attempt costs more than the token difference.

That is a routing configuration rather than a rewrite, and it is where a single endpoint earns its place. Claude Sonnet 5 is on OrcaRouter at Anthropic's own $2.00 / $10.00, and Claude Opus 5.5 is there at $4.00 / $20.00 — provider list price passed through with 0% markup, so the cost ratio in your escalation rule is the cost ratio Anthropic publishes, with no markup layer moving it. Both sit behind one key, which means the escalation path is a rule edit instead of a second integration, automatic failover keeps a failing primary from taking the request down with it, and the comparison that decides your thresholds is something you can run inside your own traffic rather than reconstruct from two vendor tables.

Who should move, and who should wait

If you are already on Claude Sonnet 5 and it is working, the case for replacing it is weak: the model is half price, its rate is now permanent, and the index gap only matters on the tasks it is failing. The case for adding Claude Opus 5.5 above it is strong, provided you can name the trigger — and the trigger is usually either a cutoff failure or a task that Sonnet 5's max-effort runs still get wrong.

If you are choosing your first Claude model, the answer is less balanced than the price ratio makes it look. A twenty-point index gap and five extra months of knowledge for double the tokens is a trade most production workloads will take on a minority of calls and almost none will take on all of them. The mistake to avoid in either direction is treating the two as interchangeable and switching on cost alone, because the failure that follows is not a worse answer — it is a confident answer about a world that ended in January.