Claude Sonnet 5.5 promises lower AI bills without a price cut

Anthropic has launched Claude Sonnet 5.5, saying that the mannequin can full work sooner whereas utilizing fewer tokens than its predecessor. In its personal testing, the corporate says that may cut back the price of a process by up to 30%.

The discount isn’t by way of value per token because the prices for tokens stay unchanged: $2 per million enter tokens, $10 per million output tokens, $0.20 per million cache-read tokens and $2.50 per million cache-write tokens.

Enhancements on Sonnet 5.5

Anthropic says token effectivity is one in all Sonnet’s 5.5 benefits in the case of finishing the identical work in comparison with Sonnet 5. In coding workflows, the corporate says early testers additionally noticed fewer steps and gear calls.

On Anthropic’s reported benchmarks, Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, in contrast with 10.3% for Sonnet 5. Anthropic describes Terminal-Bench as an analysis of agentic coding efficiency

It demonstrates higher judgment throughout long-horizon work, catches information errors by re-checking supply supplies, and avoids over-relying on internet search instruments in comparison with Sonnet 5.

There are robust enhancements in picture understanding and visible reasoning. One instance Anthropic lists is taking part in Pokémon Purple working solely from screenshots.

It’s also the primary Sonnet mannequin built-in with Anthropic’s elevated cybersecurity safeguards and automatic fallbacks beforehand reserved for flagship Opus fashions.

I examined Sonnet 5.5 towards Sonnet 5

To see whether or not Anthropic’s velocity and effectivity claims had been seen in a easy advertising workflow, I gave Sonnet 5 and Sonnet 5.5 the identical 150-word promotional-email temporary for a fictional coastal lodge.

Each accomplished the duty in two seconds, and neither produced a materially totally different end result: each integrated the core data equipped within the temporary. Claude’s free interface didn’t expose token utilization, so the check couldn’t independently confirm Anthropic’s declare that Sonnet 5.5 makes use of fewer tokens or prices much less per process. Neither mannequin reached the free-tier utilization restrict throughout the check.

Anthropic gives its personal instance of the mannequin’s knowledge-work efficiency. The corporate says it gave Sonnet 5.5 a public firm’s quarterly earnings supplies and name transcripts, along with a slide template, and requested it to create a 10-slide working overview. From the response, two specialists judged the primary draft able to ship as is.

On external-company testing, Anthropic cites that Slack says it carried out higher than Sonnet 5 on virtually all of its offline Slackbot evaluations whereas utilizing about 14% fewer output tokens. Zendesk says its assist tickets had been processed 20% sooner, whereas Field studies that Sonnet 5.5 was 2.4 instances sooner and used 12% fewer complete tokens in its testing. These are company-reported outcomes, slightly than impartial SEW checks.

Key takeaway

As a result of our consumer-interface check couldn’t confirm token utilization, the actual query isn’t simply whether or not Sonnet 5.5 is quicker. It’s whether or not it makes use of fewer tokens whereas producing an output that also meets the identical high quality bar for the duty you truly run.

For entrepreneurs considering the upgrade, the helpful measurement is subsequently not merely the mannequin’s listed token value. Evaluate the fee and token consumption of the particular duties you run, alongside output high quality and completion time, in case your interface or API exposes these measurements.