Sonnet 5.5 Is 30% Cheaper: Unlimited Agents Erase Savings

Anthropic has released Claude Sonnet 5.5.

According to the company, the model runs more than 30% faster and costs up to 30% less for most tasks, while the per-token price remains the same as Sonnet 5. That same day Simon Willison showed how a single setting wipes out that saving and multiplies the bill by 22.

Cheap intelligence pays off for a company that controls its agents' budgets, access, and auditing.

Without this control, a company risks paying more for the new model than it paid for the old one.

In Willison's observations, Sonnet 5.5 outperforms Sonnet 5 on every benchmark at the same price. Anthropic promises savings of up to 30% on a typical task.

The token price has not changed, so the savings are likely due to the model needing fewer tokens and less time to achieve the same result.

This is a hypothesis: Anthropic has not disclosed the exact mechanism.

These 30% are noticeable in high-volume tasks: generating product card descriptions, tagging a catalog by taxonomy, creating search facets, and processing incoming requests.

At KT.Team, we handle these tasks in PIM systems (Akeneo, Pimcore) across flows involving tens of thousands of SKUs, where every percentage point of model-call cost is visible in the budget.

On a one-off analytical task, nobody will notice a 30% difference.

Willison ran his usual test: he asked the model to draw an SVG of a pelican on a bicycle. Claude has a thinking effort parameter, or reasoning level.

It determines how many tokens the model can spend reasoning before answering.

At the max level, Sonnet 5.5 spent 128,000 tokens and $1.28 on reasoning, hit the limit, and never produced the SVG.

Willison writes that the same error occurs with Opus 5.5.

At the xhigh level, the task cost 5.74 cents, took 41 seconds, and produced the pelican.

The more expensive option cost 22 times more and produced no result.

Anthropic sets the per-token price, while your engineer sets the reasoning level, token ceiling, and failure behavior.

If an agent sets itself to max in a loop and retries after failure, a 30% saving can become overspending within the first week. The rule I would enforce is this: assign the reasoning level to the task class, set a hard token ceiling for every call, and trigger a predefined fallback instead of retrying when it is reached.

That same day, Willison quoted an author using the handle @joedaroo, apparently an employee of one of the AI labs based on the text.

According to him, the leap in model capabilities for cybersecurity and swarm agent work surprised the team itself. He believes companies need to build security into their processes, while people adapt more slowly than labs release new models. In the summer

Willison examined a story described as the first "escaped" agent: an OpenAI agent allegedly hacked Hugging Face systems during benchmark tests. He himself

Willison doubts whether it was an incident or a marketing stunt, and I consider the story unsubstantiated.

The mechanism is real: an agent with an unlimited token budget and network access gives an attacker a broad attack surface. Willison himself provided the counterexample. Together with Alex

With Garcia, he audited Datasette using three models: Claude Fable 5.1, GPT-5.6, and GPT-6 Astra.

The models found vulnerabilities, after which two people spent nearly a week coordinating and checking the fixes.

The result was releases 1.0a39 and 0.65.4.

Models find vulnerabilities in hours, while people need days to fix them, and that week of manual review cannot be removed from the plan.

### One gateway for all model calls Every request to Claude, GPT, GigaChat, or YandexGPT goes through a single point. Our LLM & Security Gateway sets token budgets per task and team, permitted reasoning levels, model routing, call logging, and personal-data masking.

With this setup, moving the product-description flow from Sonnet 5 to Sonnet 5.5 requires only a configuration change, not rewritten integrations.

If the new model performs worse on your data, rolling back will take just as long. ###

Agent tools with least privilege

The agent accesses systems through MCP servers, and each server exposes only the operations required for the task. An agent that writes product descriptions needs permission to read attributes from PIM and write to the description field.

There is no reason to give it permission to change prices in 1C or make external network requests.

If it fails, such an agent is limited by its permissions: at worst, it can damage a product description that is easy to roll back. ###

Models as auditors, people as judges In AI-native development, we use several models for cross-reviewing code and finding vulnerabilities, as Willison did with Datasette.

The engineer decides whether to merge, and we schedule time for the review in advance.

Sonnet 5.5 is cost-effective by the numbers, so switching to it is worthwhile.

The savings will benefit companies with a gateway enforcing limits, agents with restricted permissions, and engineers reviewing the models' findings.

Without it, the monthly bill will show calls that nobody capped and an agent that nobody stopped.

The model itself is inexpensive, but the surrounding controls require engineering work, which must be done before the agent gets access to production.