The facts: On 9 July, OpenAI opened up its new GPT-5.6 model family to the general public, in three sizes : Sol, built for advanced reasoning and long agentic tasks ($5 / $30 per million input/output tokens) ; Terra, an « everyday » model performing close to GPT-5.5 at half the price ($2.50 / $15) ; and Luna, the fastest and most affordable of the family ($1 / $6). All three share a one-million-token context window and up to 128,000 output tokens. OpenAI also unveiled ChatGPT Work, an agent designed to carry out entire assignments rather than answer one-off questions.

The Codito take: This is not the first time a vendor has segmented its line-up by size : Anthropic does it with Haiku/Sonnet/Opus, Google with Flash/Pro. What changes with GPT-5.6 is the weight it lends to that approach : OpenAI, the most consumer-facing player on the market, now structures its offering explicitly around the principle of « the right intelligence for the right task ». The price gap between Sol and Luna is a factor of 5 : using Sol for a task Luna would have handled perfectly well means paying five times the fair price, every month, on every task concerned. What this segmentation confirms : the market is no longer heading towards one universal model, but towards a portfolio of models that has to be arbitrated, precisely the subject an SME leader now needs to master, just as they choose their suppliers.

What it means for you: If your organisation uses a single AI model for everything, from proofreading an email to analysing a complex contract, you are probably overpaying on 80 % of your use cases, or getting inadequate results on the remaining 20 %. The action to take this week : map out your three or four most frequent AI use cases and ask yourself, for each one, whether it really requires the most expensive model you pay for, or whether a « mid-range » model would do just as well, at a third or a fifth of the price.

Number of the week Factor of 5. That is the price gap between GPT-5.6 Sol ($30 per million output tokens) and GPT-5.6 Luna ($6). On monthly usage of several million tokens, which becomes routine as soon as an agent runs continuously on a business process, that gap can represent several hundred euros a month, for results that are often equivalent on everyday tasks. Calibrating the model properly is no longer a technical detail : it is a budget line in its own right.

Faced with a portfolio of models (Sol/Terra/Luna, Opus/Sonnet/Haiku, or the equivalent from your provider), the temptation is always to take the most powerful one « just to be safe ». That is a miscalculation as common as it is costly. Here is a simple framework for deciding, task by task.

1. The entry-level model for everything repetitive and well defined

Summarising an email, rewriting a text, extracting information from a structured document, sorting incoming requests : these are low-variance tasks, where the expected result is predictable and the tolerable margin of error is wide. This is exactly where the cheapest model belongs (Luna, Haiku, or equivalent). If you use a flagship model for this kind of task, you are paying a premium for capability you never use.

2. The mid-tier model for day-to-day work that actually matters

Personalised sales writing, meeting summaries, a first draft of a quote, reviewing a CV : tasks where a mistake has a cost (time spent proofreading, the image you project to a client) but where the complexity stays manageable. This is the territory of the « mid-range » model (Terra, Sonnet, or equivalent), the best value for money for the bulk of an SME's daily work.

3. The flagship model, reserved for what genuinely warrants it

Complex legal analysis, multi-step reasoning on an ambiguous case, an autonomous agent that has to sustain a long task without supervision : these are the cases where top-end capability genuinely changes the outcome, not just the speed. Reserve the most expensive model (Sol, Opus, or equivalent) for those specific situations, and check regularly whether a mid-tier model now does the job just as well, since capabilities improve every quarter.

« The most expensive model is no guarantee of quality : it is a tool for a specific need, like any other. »

The action to take this week : go back through your AI subscriptions and usage, and ask yourself, use case by use case, which tier is the right one. A quick calibration audit is often enough to find 20 to 30 % in savings with no perceptible loss of quality.

Book a strategy call

Discover Codito Ongoing Support

To close this edition, we are turning the mirror on ourselves.

For a few weeks now, an internal tool has been analysing Codito's CRM every morning and sending Émile and me a Business Brief : a short list of concrete actions to take that day, drawn straight from the real signals in the pipeline. Not a generic to-do list : priorities calculated on precise criteria, a high-fit prospect left without a reply after a won meeting, an active deal stuck at the same stage for several weeks, a follow-up that is dragging on.

What this tool has taught us

• It does not just suggest, it holds us accountable. On Friday 10 July, the daily review had not been completed. The tool sent a second message : « There are still 35 unprocessed items in the CRM and the Executive workspace. Finish your review. » It is not comfortable, and that is exactly why it works.
• The model running this tool is deliberately not the most powerful one. Extracting structured signals from a CRM and prioritising a short list is a repetitive, well-defined task : exactly the first tier of the framework in Part 2. The result does not depend on the power of the model, but on the quality of the CRM data upstream and on well-defined prioritisation rules.
• A CRM nobody opens daily is a dead CRM. Obvious on paper, this is hard to sustain in practice when the diary is overflowing. The tool removes the excuse : the brief lands in the inbox, whether or not we take the initiative.

A point that ties Part 1 and Part 2 together We could have run this brief on the most powerful model on the market. We did not, because the task did not call for it. That is precisely the principle GPT-5.6 has just formalised this week, and one we apply, including to our own internal tools.