Claude Sonnet 5 vs Opus 5: A Real-World Comparison (2026)

Originally published on the Cosmic blog.

Anthropic shipped Claude Sonnet 5 on June 30, 2026 and Claude Opus 5 on July 24, 2026. Both are excellent. The expensive mistake is picking one, wiring it into every code path, and then either overpaying on trivial work or under-resourcing the work that actually matters.

This is the practical breakdown: what each model is measurably good at, what the same job costs on each once you account for tokenization, and a routing rule you can ship this week.

If you want one sentence to take away: run Sonnet 5 by default and escalate to Opus 5 on the specific paths where judgment matters more than throughput.

There is a wrinkle worth knowing before you compare scores. Anthropic evaluated these two models on different suites, so there is no clean head-to-head table, and anyone who publishes one has filled in gaps with guesses.

Here is what was actually published for Sonnet 5:

And here is what was published for Opus 5:

Notice what is missing: Anthropic did not publish a SWE-bench Verified number for Opus 5. So the honest comparison on classic coding benchmarks is Sonnet 5 at 72.7% against Opus 4.8 at 79.4%, with Opus 5 positioned above Opus 4.8 on the newer agentic and knowledge-work suites. Treat any "Opus 5 scores X% on SWE-bench" claim you find elsewhere as unsourced.

The practical read: Sonnet 5 closed most of the gap to the previous Opus generation on straightforward coding. Opus 5's advantage shows up in long-horizon agent work, self-verification, and problems where the model has to decide what the task even is.

List prices per million tokens:

Headline math says Opus 5 costs 1.67x Sonnet 5 at standard pricing, and 2.5x during the introductory window. That understates Sonnet 5's real cost, because Sonnet 5 ships a new tokenizer that counts the same text as 1.0x to 1.35x more tokens. You pay per token, not per word, so that inflation lands on your invoice.

Work a concrete example. Take a job that a baseline tokenizer counts as 1,000,000 input tokens and 200,000 output tokens:

That 1.23x row is the one that should change your thinking. On text-heavy workloads that tokenize badly, at standard pricing, Opus 5 costs about 23% more than Sonnet 5 rather than 67% more. If judgment quality matters on that path at all, a 23% premium is easy to justify.

Two caveats so you use this honestly. The inflation range is 1.0x to 1.35x depending on your content, and where you land is an empirical question about your own data. And the September 1 price change means any cost model you build this month needs revisiting. Run 1,000 real requests through both models, compare your actual billed token counts, and decide from your numbers instead of this table.

Most teams do not need a clever classifier. A static route based on task type captures nearly all of the savings:

Point four is where most of the long-term pain lives. If your model identifiers are compiled into your application, every model release becomes a pull request, a review, and a deploy.

Store your routing configuration as content instead. Then swapping Sonnet 5 for Sonnet 5.1 is a field edit that takes effect immediately, with no rebuild.

Install the SDK:

npm install @cosmicjs/sdk

Define the routing table as an object in Cosmic and read it at runtime:

import { createBucketClient } from '@cosmicjs/sdk';

const cosmic = createBucketClient({
  bucketSlug: process.env.COSMIC_BUCKET_SLUG!,
  readKey: process.env.COSMIC_READ_KEY!,
});

type ModelRoute = {
  default_model: string;
  escalation_model: string;
  escalate_after_failures: number;
};

export async function getModelRoute(): Promise<ModelRoute> {
  const { object } = await cosmic.objects
    .findOne({ type: 'model-config', slug: 'production' })
    .props('metadata')
    .depth(0);

  return object.metadata as ModelRoute;
}

Use it at the call site:

const route = await getModelRoute();

let model = route.default_model;        // "claude-sonnet-5"
if (isArchitectural(task) || failures >= route.escalate_after_failures) {
  model = route.escalation_model;       // "claude-opus-5"
}

const result = await runTask({ model, task });

When the next model lands, a non-engineer updates one field in the dashboard and production picks it up. No deploy, no pull request, no engineer in the loop. Cosmic exposes this over a REST API and the TypeScript SDK, so the same config is readable from any framework or runtime you use.

For the previous generation's version of this decision, see Sonnet 4.5 vs Opus 4.5.

Is Opus 5 better than Sonnet 5?

On ambiguous reasoning, long autonomous agent runs, and self-verification, yes. On cost per acceptable output for high-volume tasks, Sonnet 5 wins clearly. They are built for different jobs.

What does Claude Sonnet 5 cost?

$2 per million input tokens and $10 per million output tokens through August 31, 2026, moving to $3 and $15 on September 1, 2026.

What does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8.

Did Anthropic publish a SWE-bench Verified score for Opus 5?

No. Sonnet 5 scored 72.7% and Opus 4.8 scored 79.4%. Opus 5's published results cover Frontier-Bench, GDPval-AA, Zapier AutomationBench, and ARC-AGI 3.

Which Claude model should I use for coding?

Sonnet 5 for the majority of day-to-day work. Escalate to Opus 5 for architecture, cross-cutting refactors, and anything where you cannot cheaply verify the result.

Why do my Sonnet 5 token counts look higher than expected?

Sonnet 5 uses a new tokenizer that counts the same text as 1.0x to 1.35x more tokens than the previous generation. Compare billed tokens, not word counts.

Anthropic shipped two frontier models in under a month. The teams handling that well are the ones that never put a model name in a source file.

Cosmic is an AI-powered headless CMS where your content and your configuration both live behind a REST API and a TypeScript SDK. Model choice becomes a content decision, editable by anyone on your team, live in production the moment it is saved.

Start free on the Free plan with 1 Bucket, 2 team members, and 1,000 Objects. No credit card required. If you want to talk through a specific architecture, book 20 minutes with our CEO.