Subscribe to GEN
Login to GEN
GAIN, GEN's Artificial Intelligence Network, is the AI platform we developed in 2020 and have run in production ever since. It is not one large model trained on everything, but a network of smaller models, each fine-tuned on vetted data for a particular kind of task, which divide the work between them, check each other's results and agree on the answer before it is given.
It is not the fastest AI available, and it is not meant to be. It is built to be right, and it runs on our own hardware, so what you send it stays with us.
The mechanics of modern language models are remarkable. What lets most of them down is the training data. A general model learns from the internet at large, where a forum post carries the same weight as a peer-reviewed study. Ask one about the economic prospects of Bradford and you receive a blend of everything written on the subject, most of it by people who have never studied Bradford's economy, rather than the considered conclusion of the research.
The providers then add long lists of rules to stop the worst answers getting through. The result is AI that is often fluent, frequently entertaining, and confidently wrong: fine for drafting, unreliable for anything that matters. GAIN takes a different path. Its models are trained for specific tasks on properly curated data, and none of them answers alone.
Think of it as a panel of specialists rather than a single consultant. Each brings its own expertise, and together they reach conclusions none of them could reach alone.
A complex request is broken down into specialised sub-tasks, each small enough to be answered well.
Each sub-task goes to the model fine-tuned for that kind of work.
Models work in parallel or in sequence, as the task requires, passing results between them.
Results are reviewed by other models at every stage, so an error in one is caught by another.
The models vote, and a result moves on only when enough of them agree that it is correct.
The final output is validated once more before it is delivered.
What began in 2020 as an experiment in orchestrating models has become one of GEN's most valuable assets. GAIN powers some of our most demanding client projects, in telecommunications, finance, healthcare and logistics, and it lets a business of our size deliver work that would otherwise need far larger resources. New specialised models are added as the need arises.
The same approach, local models on our own hardware, runs much of what you see elsewhere on this site: the chatbot that answers from our own content, call and voicemail transcription, AI call routing on our phone systems, and the first analysis of new HelpDesk tickets. Developers can reach our local models through the AI endpoints of our Arkane API.
Current charging rates for the platform. Token rates are normalised to one million tokens, and request rates to 1,000 requests, for easy comparison.
| Metric | Rate | Equivalent |
| Requests | £0.005000 | £5.00 per 1,000 requests |
| Prompt tokens | £0.000004 | £4.00 per 1M tokens |
| Completion tokens | £0.000012 | £12.00 per 1M tokens |
| Thinking tokens | £0.000030 | £30.00 per 1M tokens |
Development, fine-tuning and integration work is charged separately at our published rates.
Long before AI was a marketing term, GEN built its first machine learning network. In 1990 we developed a system that predicted when telephone calls would end, and it did so with remarkable accuracy, at a time when the industry was still sizing its lines with Erlang calculations.
That set the pattern for everything since. Through the growth of the internet, the move to the cloud and the current wave of AI, we have tried new approaches early and in our own way. Not every experiment succeeds, and some cost us more than they returned, but that is how new things get built, and GAIN is the result of that habit. It is why we can say we were working on AI before it was fashionable, and mean it.
No. It is a network of language models, each fine-tuned and validated for a particular kind of task. A request is broken into parts, each part goes to the model best suited to it, and the models check each other's results and agree on them before an answer is returned.
Because it checks its work. Several models cross-validate each stage and must agree before the result moves on. That costs time, and we accept the cost, because a fast answer that is wrong is worth less than a slower one that is right.
On GEN's own hardware, so prompts and documents are not sent to a cloud AI service. Where a customer needs it, the models can instead be deployed on their own premises.
Per request and per token, at the rates in the table on this page, which are normalised to 1,000 requests and one million tokens for easy comparison with other providers.
Yes. Fine-tuning models on a customer's own terminology, documents and workflows is a large part of what our developers do, and it is usually what turns a generic model into one that is useful for your business.
GAIN earns its keep on the work where a wrong answer is expensive. Tell us what you need AI to do, and what it would cost you if it got it wrong, and we will tell you whether GAIN, a fine-tuned model, or no AI at all is the right answer.
For what businesses do with it day to day, see business AI and AI agents.