On 30 June 2026 Anthropic released Claude Sonnet 5. The headlines were about price - 2 dollars per million input tokens, 2 dollars 50 cents for output, as an introductory rate, through the end of August. About the context window - 2.5 million tokens, one of the largest on the market. About the fact that Anthropic is positioning the new model as a cheaper alternative to Opus, GPT-5.5 and Gemini Pro, rather than as another contender for the top of the benchmarks.
That was enough to generate 622 points and 329 comments on Hacker News. For us, at SNOK, what is more interesting than the launch itself is what it actually changes in the design of agentic automation architecture at a client - specifically in the UiPath environment, where we build most such deployments.
The barrier was not technical. It was budgetary
Agentic automation architects know the moment: a project makes sense technically, but at a certain point you have to decide how many reasoning steps the architecture can afford before the cost of tokens is no longer justifiable to the client. Every additional agent in the pipeline - for example an agent that only checks whether the output of the previous step complies with company policy before it reaches a human - is a cost the architect has to explain, not just design.
This is not an invented constraint. It is a real mechanism that slowed the move of agentic automation from pilot to production faster than any model limitation. You can have the best governance architecture in the world and still not deploy it in full, because the third verification step raises operating cost by an amount that is hard to justify in a business case.
Claude Sonnet 5 shifts that boundary. It does not eliminate the cost - it changes its scale enough that the decision “to add or not to add another verification step” stops being a budgetary decision and returns to what it should be: an architectural one.
Where it fits into UiPath architecture
The AI Trust Layer in UiPath does not force a single language model provider. The model is selected via a connector according to where it physically runs - for Anthropic models the realistic path is the Amazon Bedrock connector, where Claude is available as one of many supported models. This separation has a practical consequence: changing the model in a UiPath architecture is a change of connector configuration, not a rewrite of the agent’s logic from scratch.
In practice, Claude Sonnet 5 can enter three different layers of the UiPath stack, each for a different reason:
Agent Builder (low-code) is the default, production-recommended way to build an agent - a prompt plus a set of ready-made tools, no Python, conversational. The lower reasoning cost of Sonnet 5 lowers the barrier to entry for this type of agent in scenarios where the number of model calls used to be the factor limiting pilot scale.
Coded Agent (Python) is the layer where Sonnet 5’s 2.5 million tokens of context make the biggest difference. Custom RAG, complex LangGraph states, MCP integrations, multi-step human-in-the-loop with resumption - these are scenarios where, until now, you had to split documentation, code and decision history into fragments to fit them into the context window. With 2.5 million tokens the complete process documentation, decision history and application code can enter a single prompt without losing continuity between steps.
Maestro, the orchestrator of multiple agents, bots and people in one long-running, monitored process, is where the change in economics has the greatest strategic impact. The lower reasoning cost per step changes the calculus of how many verification gates the architect can place in a process without breaking the budget - and it is the verification gates, not the execution automation itself, that distinguish governed automation from a plain script with a language model inside.
One technical caveat, so this does not sound broader than it is: we are talking here about Claude Sonnet 5 as a cloud option available through Bedrock. That is a different path from self-hosted deployments - for example models run locally via NVIDIA NIM on the client’s own infrastructure, where data residency requirements decide, not the price of tokens. Both paths have their place, depending on whether the client has a regulatory reason (NIS2, DORA, a regulated sector) to keep data from leaving their infrastructure.
What it means in practice - a worked example
So that this change does not remain a mere declaration, it is worth calculating it on a simple example. A process with three reasoning steps per event - classification, policy-compliance validation, generating a report for a human - with an input document on the order of 5 thousand tokens and a response of similar size, at the prices in force before the introduction of Sonnet 5 for models in this class cost several times more for the same number of events than at the introductory rate of $2/$2.5 per million tokens. At a volume on the order of a few thousand events per month - the typical scale of a mid-sized company’s process, not a corporation’s - the difference in the monthly cost of the AI layer runs into thousands of zlotys, not pennies.
This matters in two places within the offer. First, in the ROI calculation of an agentic automation project - a lower operating cost of the AI layer shortens the payback period, which is an argument easy to verify by the client’s CFO, not just the CIO. Second, in the decision the architect makes at the design stage: whether to add a third verification step or limit it to two. With a lower cost, that decision stops being a compromise between governance and budget.
What it means for governance and compliance
Regulated environments - the financial sector under DORA, critical infrastructure under NIS2, projects requiring AI Act compliance - have an additional reason to treat lower reasoning cost as an opportunity, not just a saving. A cheaper model means that a compliance review or an audit of decisions made by an agent - a step that is itself a cost added to the process, not a business value in its own right - becomes easier to justify in the budget. The 2.5 million token context window also allows a complete policy or regulation to be brought into a single review, instead of splitting it into fragments and risking a loss of context between them - which has practical significance in AI Act audits, where the completeness of the review is part of the evidence of compliance.
Why we do not recommend a single model to clients
This leads to a position SNOK has held for a long time and which Claude Sonnet 5 illustrates well: we do not sell the deployment of a specific language model. We sell an architecture in which the choice of model is a configuration parameter, not the foundation on which the entire solution stands.
The consequence of this approach is concrete and measurable. When - as now - a model appears on the market that is cheaper or better at a given task, a client with an architecture based on the AI Trust Layer pays to test and switch the connector. They do not pay to rewrite the agent, rebuild the integration or migrate data. That is the difference between a deployment that has to be redesigned with every model release cycle - and Anthropic, OpenAI and Google release new versions on a cadence of months, not years - and an architecture that absorbs that cycle without an additional project.
The choice of language model should be a governance decision: cost, latency, regulatory compliance, data residency. It should not be a marketing decision or loyalty to a single provider. When a client asks us whether they should “deploy Claude” or “deploy GPT”, the answer begins with a question about the architecture in which that model will run - not with the name of the model.
What to do about it now, and what not to do
The introductory pricing of Claude Sonnet 5 is time-limited - it applies through the end of August 2026, after which it returns to a higher level. It is a demonstration window, not a permanent price, and it should be treated as such in a project’s budget planning. GPT-5.5 and Gemini Pro remain strong competitors in benchmarks for specific tasks - Sonnet 5 wins primarily on cost and context size, not automatically on every quality criterion.
SNOK’s recommendation is therefore simple and the same as the one we apply internally: benchmark on your own, specific use cases before a production migration, do not migrate on the basis of a press headline. For a client with an existing AI Trust Layer architecture, that benchmark is a week of work, not a separate project - and this is the right moment to do it, before the introductory pricing window closes.
For clients designing an agentic architecture from scratch today, Claude Sonnet 5 is a good argument to build from the outset on a model-abstraction layer, rather than on integration with a single provider - so that the next model release, a few months from now, is an opportunity to test, not another migration project.
This material was prepared on the basis of official Anthropic announcements and an analysis of the AI Trust Layer architecture in UiPath. Recommendations regarding a specific client architecture require separate technical verification - we invite you to a conversation about agentic automation architecture.
See also: UiPath opens the platform to AI agents: Claude Code and Codex are already in and UiPath is no longer RPA. It has become the Switzerland of AI agents.