Claude Sonnet 5 vs Opus 4.8: Is the Flagship Worth 2.5×?
Claude Opus 4.8 costs 2.5× Claude Sonnet 5's intro rate for a 3.4-point edge on SWE-bench Verified — 88.6 vs 85.2, as of July 2026. On most workloads that premium is wasted; on long-horizon agentic work with expensive retries, it pays for itself. After September 1 the multiple compresses to roughly 1.7×, which flips the math for borderline tasks. For PMs, the answer is a routing policy — Sonnet 5 by default, Opus 4.8 on logged escalation — not a single model choice.
Your budget decision this quarter isn't "which Claude model" — it's "what's your escalation policy." Treat Opus 4.8 as a paid upgrade you grant to specific task classes, not a default you inherit from whoever set up the API keys. Here's the concrete action for this week: pull 30 days of usage from your Anthropic console, split it by task type, and price each type at both rate cards — Sonnet 5 at $2/$10 (and again at the $3/$15 standard rate that lands September 1) versus Opus 4.8 at $5/$25. Ten minutes in a cost calculator gives you a per-task-class delta. Then check your eval or retry data: any task class where Sonnet 5's failure rate is materially higher is an Opus candidate, because retries plus human review usually cost more than the 2.5× premium. Everything else defaults to Sonnet 5 with batching and prompt caching turned on — those two levers alone can cut the Sonnet bill by half or more. To be fair, if your team runs a handful of genuinely hard agentic workloads and nothing else, paying for Opus everywhere is a defensible simplification; routing has its own maintenance cost. But most teams are not that team. The intro window closes August 31, and teams that instrument routing now get weeks of discounted tokens to gather the data. Do the split this week, not after the price resets.
Claude Opus 4.8 at $5/$25 per 1M tokens costs 2.5× Claude Sonnet 5's intro $2/$10 rate for +3.4 points on SWE-bench Verified — 88.6 vs 85.2, as of July 2026.
Every Claude shop is running the same spreadsheet this month. Claude Opus 4.8 costs $5 per million input tokens and $25 per million output, per Anthropic's pricing docs. Claude Sonnet 5's intro rate is $2 and $10 through August 31, 2026.
That's 2.5× the price for 3.4 extra points on SWE-bench Verified — 88.6 versus 85.2, as of July 2026, per the llm-stats board. My verdict: make Sonnet 5 your default, route the hardest 10-20% of work to Opus, and stop paying flagship rates for commodity tokens.
One framing note before the math. Opus 4.8 isn't the top of the family. Claude Fable 5 sits above it at $10/$50 with a 95.0% SWE-bench Verified score, as of July 2026.
But Opus is the flagship most teams actually price against, so that's the comparison that decides budgets.
The arithmetic, without hand-waving
| Model | Input / output (per 1M tokens) | SWE-bench Verified | Notes |
|---|---|---|---|
| Claude Sonnet 5 (intro) | $2 / $10 | 85.2% | Through Aug 31, 2026 |
| Claude Sonnet 5 (standard) | $3 / $15 | 85.2% | From Sep 1, 2026 |
| Claude Opus 4.8 | $5 / $25 | 88.6% | Released May 28, 2026 |
| Claude Fable 5 | $10 / $50 | 95.0% | The actual top rung |
Prices per Anthropic's pricing docs; scores from the llm-stats board (July 2026, self-reported). Live numbers sit on our model tracker.
Now make it concrete. Say your coding agents burn 50M input and 10M output tokens a month. On Sonnet 5's intro rate that's $100 in, $100 out: $200.
On Opus 4.8 it's $250 plus $250: $500. You're paying an extra $300 a month, per workload, for 3.4 benchmark points.
After September 1, Sonnet 5 moves to $3/$15 and that same workload costs $300. The Opus multiple compresses from 2.5× to roughly 1.7×. Budget against the 1.7×, not the discount — the intro rate is a promotion, not a price.
We covered the deadline mechanics in the pricing breakdown, and you can run your own token volumes through the cost calculator in about two minutes.
Where the 3.4 points actually show up
Benchmark averages hide the distribution. On routine work, Sonnet 5 and Opus 4.8 are hard to tell apart. The gap lives in the tail: long-horizon agentic runs, cross-repo refactors, messy migrations, and high-stakes reasoning where a wrong answer costs real money.
Here's the operational math that matters more than the sticker price. A failed agent run doesn't just waste its own tokens. It costs the retry, plus the engineer who reviews the wreckage.
If a task class fails three times in ten on Sonnet and once in ten on Opus, Opus is cheaper on that task class even at 2.5× — the retries and the humans ate the difference. That's the honest case for the flagship. It's a case for Opus on specific work, not across your whole stack.
And if you're shopping outside the family: OpenAI's GPT-5.6 Terra sits at $2.50/$15 with a 1.05M-token context, though OpenAI no longer self-reports to the llm-stats SWE-bench board. We put it head-to-head with Sonnet 5 in the builder comparison.
Where Opus is wasted spend
Summarization. Extraction. Classification.
Ticket triage. First-draft codegen with a human in the loop. On this work the 3.4-point gap is invisible and the 2.5× is pure loss.
Routing your standup summaries through Opus is like couriering interoffice mail by chartered jet: it arrives no sooner, and finance will eventually ask questions.
Both models carry the same 1M-token context window, so context size is never a reason to escalate. And Sonnet 5's cost levers stack. The Batch API takes 50% off both input and output.
Prompt caching makes cache reads up to 90% cheaper than base input — and on the Claude API, cached reads don't count toward input-token rate limits at all, per Anthropic's docs. A cached, batched Sonnet 5 pipeline runs at a small fraction of naive Opus spend. None of this pricing is accidental: the intro rate reads as IPO-timed land-grab pricing, a case our launch analysis makes in full.
The routing playbook
What to do Monday morning:
- Default everything to Sonnet 5. New workloads start there. No exceptions without data.
- Escalate on failure, not on vibes. When a Sonnet run fails eval or needs a retry, re-run it on Opus 4.8. Log every escalation.
- Pin Opus to named task classes. After two weeks of logs you'll know exactly which tasks earn the flagship. Route those to Opus directly and skip the failed Sonnet pass.
- Re-run the math on September 1. At roughly 1.7×, some borderline task classes flip toward Opus. Put it in the calendar now.
To be fair to Anthropic, 2.5× is the distorted number. Sonnet's discount is doing the distorting; the durable multiple is roughly 1.7×, and Opus 4.8's rate card hasn't moved since its May 28 launch. As flagship premiums go, that's restrained.
The harder question isn't Sonnet versus Opus. It's that the ladder now has three rungs — Sonnet 5 at $2/$10, Opus 4.8 at $5/$25, Fable 5 at $10/$50 and 95.0% — and the prices reset faster than most teams re-run their comparisons. Was your current model a decision, or a default?
If swapping rungs takes you a quarter, you don't have a routing strategy. You have a dependency. Build the escalation layer this month, while the discount is still paying for it.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
On averages, barely — Sonnet 5 and Opus 4.8 are hard to tell apart on routine work. The gap concentrates in tail tasks: long-horizon agentic runs, cross-repo refactors, and high-stakes reasoning where failures are expensive. Both scores on the llm-stats board are self-reported by Anthropic, so treat them as directional. The only number that settles it for your team is an eval on your own task classes, which the intro-priced weeks make cheap to run.
Start crude: it's one model parameter. Default every call to claude-sonnet-5, catch eval failures and retries, and re-run those on Opus 4.8 while logging the escalation. After two weeks, pin the task classes that consistently escalate straight to Opus so you stop paying for the failed Sonnet pass. Turn on prompt caching and the Batch API for the Sonnet tier — cache reads run up to 90% cheaper and batch jobs get 50% off, which widens the gap the router is exploiting.
Two real risks. First, hard agentic work that quietly fails more often on Sonnet costs you retries and engineer review time, which can exceed the 2.5× premium you saved — that's why escalations must be logged, not guessed. Second, the $2/$10 rate expires August 31, 2026, so any budget built on intro pricing is wrong by September; model your spend at $3/$15 from the start. And if even Opus 4.8 struggles on your extreme tail, Claude Fable 5 at $10/$50 exists above it.
