The facts: On 30 June, Anthropic released Claude Sonnet 5, billed as the « most agentic » model in its mid-range line-up to date : it can plan, use tools (browser, terminal, API) and carry out tasks autonomously at a level that, only a few months ago, required larger and more expensive models. Its performance approaches Opus 4.8 on reasoning, tool use, coding and complex work tasks — at a significantly lower price : $2 per million input tokens, $10 per million output tokens until 31 August (then $3 / $15), with up to 90 % savings through context caching and 50 % with batch processing. Anthropic also reports a lower rate of undesirable behaviours than the previous generation, a key point for autonomous use in production.

The Codito analysis: For two years, « autonomous AI agent » was a term reserved for impressive demos and comfortable budgets : the model most capable of executing tasks autonomously was invariably the most expensive one. What Sonnet 5 changes is that trade-off : state-of-the-art agentic capability is now available at mid-range pricing, not flagship pricing. In practice, an agent that six months ago would have required Opus (and its cost) to run a multi-step task without derailing can now run on Sonnet — at a third of the price, or less with caching. This shift in the price/capability frontier is exactly the kind of move that has historically opened a market to a new category of buyers : the SMEs that were watching AI agents with interest, but whose business case did not yet add up.

What it means for you: If you assessed an AI agent project (automated prospecting, quote processing, document monitoring, customer support) in the spring and concluded that « the ROI isn't there yet », redo the maths this week. Two things have changed : the cost per task has mechanically dropped, and the reliability of autonomous execution has improved — which reduces how often a human has to step in, and therefore the real cost. Don't reopen a buried project out of nostalgia : rerun the numbers with the new pricing and the new reliability, and see whether the break-even point has moved.

Number of the week 90 %. That's the potential saving on the running costs of a Claude Sonnet 5 agent through context caching, for tasks that repeat the same base context (a specification document, a knowledge base, a customer history). Combined with an entry price already below that of Opus, some agents that cost several hundred euros a month to run now come in under €50. Agentic AI is no longer a cost centre — for many use cases, it has become a negligible budget line.

A price cut changes nothing in itself. What matters is what it makes possible that wasn't before. Here are three concrete shifts that cheaper agentic AI opens up for an SME — beyond the simple « it costs less ».

1. Low-volume tasks become profitable again

An agent handling 200 quotes a month struggled to justify its running costs at the prices of six months ago — the maths favoured high volumes or large accounts. With the cost per task divided by three to ten depending on the case (caching included), moderate-volume processes — typically those of a 20 to 100-person SME — finally clear the profitability threshold. That is precisely the segment the price cut unlocks : neither the giant already automating at scale, nor the process so rare it is never worth automating — the middle case, the most common one in SMEs.

2. Iteration becomes affordable, not just deployment

An expensive agent gets tested once, validated, then frozen — every iteration carries a cost that discourages experimentation. A cheap agent can be reworked, corrected and improved continuously without every attempt weighing on the budget. That changes the deployment method : you can now ship an imperfect prototype quickly and refine it week after week with feedback from the field, rather than trying to nail everything down before the first launch. It is a genuine shift in mindset for SMEs, which are often reluctant to invest in a tool that will only be "good" after several cycles.

3. The barrier is no longer the model's price — it's the quality of your data

When execution cost stops being the limiting factor, the limiting factor that remains is the material you feed the agent : your documented processes, your structured data, your open stack. A Sonnet 5 agent costing €50 a month will produce nothing worthwhile if it works on data scattered across three disconnected systems. Cheaper pricing does not remove the groundwork — on the contrary, it makes it the only remaining obstacle. Which is not bad news : it is a problem you control, unlike a model price set by a third party.

« When the model's price stops being the obstacle, what's left is your own preparation. »

Your action for this week : pull up the list of automation projects you set aside for lack of a clear return. Recalculate with the new pricing. And for those that become viable again, ask yourself the real remaining question : is your data ready to feed the agent ?

Book a strategy call

Discover Codito Ongoing Support

To close this edition, the next chapter of a story we started three weeks ago in these pages.

On 16 June, we told you about the first Ongoing Support session at Gustave Rénovation : Amaury and Joseph, a specification document already prepared, a stack (Kalitix, Excel, SharePoint) to work with rather than around. This week, the next step : on 1^st July, I delivered the first version of the prototype — named Gustave Chantier 360 — and on 3 July, we demoed it with Amaury.

What the demo confirmed

• The Kalitix API connector works. It was the technical prerequisite identified in the very first session — without it, everything else stayed theoretical. It is now operational, and that is what makes the prototype viable going forward.
• The dashboard and quote generation are validated. Amaury was able to follow the process end to end and confirm it was fit for purpose — the first concrete proof that the material gathered in M1 translates into a usable tool.
• Model flexibility is confirmed. The prototype can switch between Anthropic and Mistral models — exactly the sovereignty requirement Amaury set out in the first session, and which was then only a promise.

What still needs refining

Three pieces of feedback shape what comes next : being able to edit a quote directly in the tool (or via Kalitix export/import), linking site reports to the project's persistent data, and making the schedule more flexible — moving a work package or a date without rebuilding everything. Nothing alarming : these are exactly the adjustments you expect at this stage of a prototype. The next session is set for 15 July.

The thread running through this week What Gustave Chantier 360 demonstrates in practice is exactly what we described in Part 2 : once the price of agentic AI comes down, what determines the quality of the result is the preparation upstream — the specification document, the mapping of the stack, sovereignty settled from the outset. Three weeks after scoping, the prototype stands up because that preparation was already there.