Product Update, GPT-6 Astra and Fable 5.1 now available

The September model catalogue is live. Three new models are selectable across NearSync, and every one of them now tells you what it costs in credits before you choose it.

This post explains what is available, how to choose, and what each one draws from your allowance.

Which AI models are new

Gemini 3.8 Flash is now the recommended Google model. A one million token context window, and the lightest model in the catalogue on credits.

Claude Fable 5.1 is a frontier model with a one million token context window.

GPT-6 Astra is also frontier, with the same one million token window. Access is limited by OpenAI rather than by us, so it will not be available to every account.

The catalogue now sorts every model into a tier. Recommended is what we suggest for everyday work. Premium is a step up in reasoning. Frontier is the most capable thing each vendor sells. Previous generations stay selectable and are marked legacy.

How credits work

A credit is roughly one question to the assistant. That is measured rather than estimated, from real usage across more than a thousand requests.

Every seat gets its own allowance each month. Standard is 2,000 credits, Pro is 4,000, and you can top up at any time. Allowances are per seat rather than a shared pot, so one person running a heavy week does not empty the team's allowance.

Recorded meetings do not draw on credits. They are bundled by the hour at the organisation level, and so is email.

What the new models cost in credits

A more capable model does more work for each answer, so it draws more credits. NearSync states that as a multiple of the model we recommend from the same vendor, because a multiple is a number you can act on and a token price is not.

Against their vendor's recommended model, per answer:

Model	Tier	Credits per answer
Gemini 3.8 Flash	recommended	the Google baseline
Claude Sonnet 5	recommended	the Anthropic baseline
GPT-5.6 Terra	recommended	the OpenAI baseline
Claude Opus 5	premium	about 2.5x Sonnet 5
Claude Fable 5.1	frontier	about 5x Sonnet 5
GPT-6 Astra	frontier	about 4.2x GPT-5.6 Terra

In practice, on a Standard seat with 2,000 credits, that is roughly two thousand questions a month on a recommended model, or around four hundred on a frontier one. Same allowance, different answer to how far it goes.

How to choose an AI model

For most of what a business asks software to do, the recommended model is the right answer. Summarising a thread, drafting a reply, pulling a figure out of a document, classifying an email: these are not hard problems for a current model, and a frontier model does not answer them better. It just answers them for five times the credits.

Frontier models earn their place on long reasoning and on code. A complex multi-step analysis across a large document set is where the difference is real and worth paying for.

Why the context window matters

All three new models carry a one million token context window. Several older ones in the catalogue are 128,000 or 200,000.

The context window is how much material a model can hold in view at once. For a short question it is irrelevant. It becomes the deciding factor when the work is large: reading a full contract alongside the thread that negotiated it, or reasoning across a year of invoices rather than a sample.

It is a separate thing from capability, and worth knowing, because a recommended model with a large window often beats a more capable model with a small one on exactly the tasks people assume need the expensive option.

Why we show what each model costs

Every premium and frontier model carries a line stating its cost per answer against the recommended model from the same vendor. Option labels say it too: premium models are marked high usage, frontier models very high usage.

That appears wherever a model is chosen, underneath the active model in the Control Centre, and again where a department's model is set. If you select the most capable option we offer, you are told what it draws before you commit to it, and told again while it is running.

We know this makes our own top-tier options look expensive. A picker that lists eighteen models sorted by how new they are will move more frontier usage than one that writes the multiple beside them. The alternative is somebody meeting that number for the first time when their allowance runs out mid-month.

Your default AI model did not change

Adding three models changed no organisation's default.

The easier version of a catalogue update moves everyone to the newest model, on the reasoning that newer is better and most people never open the setting. It would have improved some answers. It would also have burned through allowances nobody agreed to spend and changed the behaviour of every automation running against the previous model, mid-month, without notice.

Previous generations stay exactly where they are, still selectable, marked legacy so you can see they are a generation behind. Moving is a choice you make.

Why some models are limited access

Tier tells you what a model draws. It does not tell you whether you can have it.

Most models are generally available. Gemini 3.1 Pro is a preview, which means the vendor can change or withdraw it. GPT-6 Astra is limited, which means OpenAI decides who gets access, and a key without that access is refused.

We list limited models rather than hiding them until everyone can use one, and the picker says so plainly.

Where to change your AI model

Model selection lives in Intelligence HQ. Your organisation default sits in the Control Centre, and routing lets you put a specific department on a specific model. Both read the same catalogue, so a model added centrally appears everywhere at once. Changing the model is an admin action.

Routing per department is worth using. Finance and Support ask different things of a model, and there is no reason both should sit on the same one. A department doing document-heavy reasoning can run on a premium model while everything else stays on the recommended one, so the higher rate applies only where it earns it.

If you are not sure which tier your work needs, start on the recommended model. It is the lightest on credits and it handles more than people expect.