SIGNALDIGITAL.COM

Pioneer digital agency established in 1996

Claude Sonnet 5.5: The Price Tag Didn’t Move. The Bill Did.

Claude Sonnet 5.5 is a lesson in how AI actually gets cheaper — and why Europe’s small firms should stop shopping for the flagship

FROM THE EDITORS

Every car buyer knows the trick. The sticker price is the least interesting number on the window. What matters is what it costs to get from A to B, every day, for years.

Anthropic just pulled the same move with AI. On 28 September it released Claude Sonnet 5.5, the second model in its 5.5 family, a week after Opus 5.5. The headline promise: more than 30% faster than Sonnet 5, and up to 30% cheaper for most work.

Here’s the part most coverage will skip past: the price per token didn’t change at all. Sonnet 5.5 still lists at $2 per million input tokens and $10 per million output tokens — identical to Sonnet 5. The savings come from somewhere else. The model simply uses fewer tokens and fewer tool calls to finish the same job.

Same fuel price. Better mileage.

For a 12-person accounting firm in Lyon, a webshop in Aarhus or an engineering consultancy in Graz, that distinction is everything.

Europe’s small firms don’t buy tokens. They buy finished jobs.

The AI industry prices its products like a utility: per unit consumed. But nobody running a small business thinks in units. They think in outcomes. What did it cost to turn 40 supplier invoices into a clean spreadsheet? To draft the tender response? To fix the bug in the booking page before Monday?

Measured that way, Sonnet 5.5 is a genuine price cut — even though no price was cut. Anthropic says that on several evaluations, the model running at low or medium effort beats Sonnet 5’s best score at roughly a tenth of the cost per task.

A sceptic will say “up to 30%” is doing a lot of heavy lifting, and they’d be right. “Up to” is marketing’s favourite phrase. Your mileage — literally — will vary with how you use it. But the direction of travel is unambiguous: the same money now buys more finished work, and it arrives faster.

For EU firms there’s one more wrinkle. The list price is in dollars. A weaker euro can quietly eat a chunk of any efficiency gain, and a stronger one amplifies it. Budget in outcomes, but watch the exchange rate.

The two-point gap

Look closely at Anthropic’s own benchmark table and one comparison stands out.

On GDPval-AA — a test of real-world knowledge work across dozens of occupations — Sonnet 5.5 scores 1844. The flagship Opus 5.5 scores 1846. Two points. On AA-Briefcase, another office-work benchmark, the gap is 11 points. On computer use, it’s under two percentage points.

That is the whole story for small business in one number. The mid-tier model now does office work at essentially flagship level. Documents, summaries, spreadsheets, presentations, clearing an inbox of routine requests — the everyday grind that eats a small team’s week.

The obvious objection: these are vendor-published benchmarks, run on vendor-chosen tests. True. Even the footnotes admit that the independent knowledge-work evaluations were run on a pre-release deployment with a bug, which Anthropic believes slightly understated Sonnet’s results. Take every number with salt. But note that the flagship is also in the same table, graded by the same people. When the vendor’s own premium product barely beats its cheaper one, that’s not a gap worth paying for.

The lesson for a Mittelstand business or a Nordic SME: stop defaulting to the most expensive model on the menu. The flagship still wins on the hardest problems — complex coding, deep multidisciplinary reasoning, the work that needs careful judgment. For most of what a small firm does, it’s a sports car for the school run.

The coding jump nobody should ignore

One row in the table looks like a typo. On Terminal-Bench 4.0, an agentic coding test, Sonnet 5 scored 10.3%. Sonnet 5.5 scores 70.6% — higher than Opus 5.5.

For the thousands of small European firms that run on a patchwork of WordPress sites, Shopify stores, homegrown scripts and one overworked freelancer, this matters. Fixing bugs is one of the tasks Anthropic explicitly says the model is built for. The cost of “can someone just fix this?” keeps falling.

A caveat woven in rather than tacked on: Sonnet 5.5 is the first Sonnet to ship with cybersecurity safeguards borrowed from Anthropic’s most capable systems. Routine development and patching are meant to work normally, but some higher-risk security requests will be quietly routed to the older Sonnet 5. For the small IT consultancy doing penetration testing for local clients, that’s a limitation worth knowing about before it surprises you mid-project.

What it means under European rules

Three practical points for firms operating under GDPR and the AI Act.

Data retention. Sonnet 5.5 is offered with zero data retention, as its predecessors were. For a firm feeding customer records, contracts or HR documents through an AI tool, that’s the first question your data protection adviser will ask — and it has a clean answer. It’s not a GDPR compliance certificate; you still need your own data processing agreement and a lawful basis. But it removes one of the most common objections.

Where it runs. The model is available through Anthropic directly and through Amazon Web Services, Google Cloud and Microsoft Azure. Many European firms already live inside one of those clouds, often with contracts and regional settings their IT partner has already negotiated. Adopting a new model through infrastructure you already trust is far easier than onboarding a new vendor.

Accuracy on sensitive work. One early customer in financial services and healthcare reported that the model rechecks figures against source documents and catches errors its predecessor missed. For an EU business, a confidently wrong AI summary isn’t just embarrassing — under the AI Act’s direction of travel, it’s the kind of failure you’ll increasingly be expected to guard against. A model that double-checks itself is a better starting point, not a substitute for human review.

Upgrade now, or wait for Haiku?

Anthropic says a third model, Haiku 5.5, built for high-volume, cost-sensitive work, is coming in the coming weeks. Some firms will be tempted to wait.

Don’t overthink it. If you already use Sonnet 5 — through the Claude apps or an integration a developer built for you — moving to 5.5 is a model-name change at the same list price. There’s little to lose. If you run genuinely high-volume, simple jobs — tagging thousands of product listings, sorting support emails — Haiku may be the better fit when it lands. Test both on your own work. Your own tasks are the only benchmark that pays your invoices.

The bottom line

The AI race is usually reported as a contest of peaks: who built the smartest model. Sonnet 5.5 is a reminder that for most businesses, the peak is irrelevant. What matters is the middle of the range, and what it costs to get a normal Tuesday’s work done.

For Europe’s small firms — margin-conscious, compliance-heavy, short on developers — the story isn’t that AI got smarter. It’s that the capable, affordable option just got close enough to the flagship that the flagship stopped being the default.

The sticker price didn’t move. Check the bill.

Sources

Leave a Reply

Discover more from SIGNALDIGITAL.COM

Subscribe now to keep reading and get access to the full archive.

Continue reading