Sonnet 5 is 12 points behind Opus 5. That gap is bigger than every argument people have about Luna.
Opus 5 leads Sonnet 5 by 12 points on Artificial Analysis v4.3, at 2.5 times the price. The biggest step among Claude's top tiers, and when paying it is wrong.
GTM Architect & Growth Operator · Now · 16 September 2026
TL;DR · Key insights
- On the Artificial Analysis Intelligence Index v4.3, Opus 5 scores 50.70 and Sonnet 5 scores 38.36. That 12-point gap is the widest step among Anthropic's top three tiers, Sonnet 5, Opus 5 and Fable 5.1.
- Opus 5 costs $5 / $25 per million tokens and Sonnet 5 costs $2 / $10. The ratio is exactly 2.5 on both sides, and Sonnet 5's price is now Anthropic's standard one.
- Sonnet 5 is the faster model on every measure that matters to a waiting user: 79.5 tokens per second against 49.4, and labelled fast against moderate by Anthropic itself.
- The mistake is treating Sonnet 5 as a cheaper Opus. On this index it scores within a point of GPT-5.6 Luna, which costs a tenth as much. Its case is speed, not intelligence per dollar.
Most of the arguments I see about cheap models are fights over two index points. Luna against Sonnet 5. Terra against Sonnet 5. Sol against Terra.
Meanwhile there is a 12-point gap inside Anthropic’s own lineup that almost nobody talks about.
| Info | Claude Sonnet 5 | Claude Opus 5 |
|---|---|---|
| Input / output per 1M | $2 / $10 | $5 / $25 |
| Batch input / output per 1M | $1 / $5 | $2.50 / $12.50 |
| Intelligence Index v4.3 | 38.36 | 50.70 |
| Output speed, tokens/s | 79.5 | 49.4 |
| Anthropic's latency label | Fast | Moderate |
| Blended price per 1M | $1.54 | $3.85 |
| Context / max output | 1M / 128K | 1M / 128K |
| Reliable knowledge cutoff | Jan 2026 | May 2026 |
Prices and latency labels from Anthropic's pricing and models pages. Index, speed and blended price from Artificial Analysis, read on 16 September 2026 against Intelligence Index v4.3. Blended price uses Artificial Analysis's 7:2:1 cache-hit, input and output ratio.
Twelve points for 2.5 times the price. On most comparison pages that would be the headline.
Put the 12 points next to the other gaps
On the same index, the steps look like this:
- Sonnet 5 to Opus 5: 12.3 points for 2.5 times the price
- Opus 5 to Fable 5.1: 2.7 points for twice the price
- Sonnet 5 to GPT-5.6 Luna: 0.9 points, both displayed as 38
Read those in order and the lineup stops looking like a smooth ladder. The top step is short and expensive, which I went through in Opus 5 against Fable 5.1. The bottom step is long and comparatively cheap.
That makes the step up to Opus 5 the best-value upgrade in Anthropic’s lineup. It also matches Anthropic’s own advice, which is to start with Opus 5 for most workloads.
Sonnet 5 is not a cheaper Opus
The tempting reading of the price table is that Sonnet 5 is Opus 5 at 40% of the cost, with some capability shaved off. The index says otherwise.
At 38.36, Sonnet 5 lands within a point of GPT-5.6 Luna at 37.50, which lists at $0.20 / $1.20. That is a tenth of Sonnet 5’s price for nearly the same score. If intelligence per dollar were the only axis, Sonnet 5 would have no case at all, and I made a version of that argument in Luna against Sonnet 5.
Sonnet 5 does have a case. It is just a different one:
- Speed. 79.5 tokens per second against Opus 5’s 49.4, and a fast latency label against moderate. A faster model shortens every loop a person sits through, and in interactive coding that adds up across a day.
- Price stability. Its $2 / $10 was launched as introductory and is now Anthropic’s standard price, with the scheduled rise cancelled.
Neither of those is intelligence. Both are real reasons to pick it.
When Sonnet 5 is the right call
| The work | Model | Why |
|---|---|---|
| Ambiguous architecture or planning | Opus 5 | Twelve points is the gap that shows up as a wrong plan. The expensive part is being wrong. |
| Complex agentic coding | Opus 5 | Anthropic's own description of what Opus 5 is for. The price ratio is the best in the lineup. |
| A well-specified implementation task | Sonnet 5 | Scope is fixed, execution is the job, and the faster model finishes sooner. |
| Anything a person waits on | Sonnet 5 | Fast against moderate, felt on every request. |
| Huge volume of simple classification | Neither | Luna scores within a point of Sonnet 5 at a tenth of the price, and Haiku 4.5 costs half as much for work that needs no reasoning. |
Default to Opus 5, route well-specified and latency-sensitive work down to Sonnet 5, and do not pay Sonnet 5 prices for work a volume model can do.
The last row is the one I would check first in any existing setup. A pipeline that sends bulk classification through Sonnet 5 because it is “the cheap Claude” is paying a premium for speed it does not need and intelligence it does not use.
What would change my mind
A Sonnet 5 point release. The 12-point gap is the largest step among the top three tiers, and it is the one a point release would most visibly move.
A published cost-per-completed-task comparison. Index run costs measure the whole evaluation. On short, well-specified coding tasks the faster model with fewer steps can finish cheaper than the ratio suggests, and I have seen no clean measurement of that for this pair.
A price change on Opus 5. The 2.5 ratio is what makes the step up such good value, and a cut would widen the case further while a rise would narrow it.
Questions people asked
Is Claude Opus 5 worth the price over Sonnet 5?
For complex agentic coding and anything ambiguous, yes. Opus 5 scores 50.70 and Sonnet 5 scores 38.36 on the Artificial Analysis Intelligence Index v4.3: 12 points for 2.5 times the price. That is a far better trade than Fable 5.1 over Opus 5, which costs twice as much for under three points.
How much cheaper is Sonnet 5 than Opus 5?
60% cheaper on both sides: $2 and $10 per million tokens against $5 and $25. Batch keeps the ratio at $1 / $5 against $2.50 / $12.50, and $2 / $10 is now Sonnet 5’s standard price.
Is Sonnet 5 faster than Opus 5?
Yes: 79.5 against 49.4 output tokens per second, and labelled fast against moderate by Anthropic.
Which Claude model should I use by default?
Opus 5, which is also Anthropic’s own recommendation. Route well-specified, latency-sensitive work down to Sonnet 5 and long-horizon work that Opus 5 fails up to Fable 5.1.
If you want this mapped onto what your team actually runs, the contact page is the fastest route to me.
