Four new model names this week. Which ones you can actually open

Four model names went round the AI internet at once this week, and only two of them refer to something that shipped in the last seven days. Google put out Gemini 3.8 Flash on Wednesday. Anthropic put out Claude Fable 5.1 on Tuesday. Meta's Muse Spark and OpenAI's Astra were carried along in the same conversation, one of them a flagship that has been out since spring, the other a model nobody outside OpenAI has ever used.

If you pay for an AI subscription, the question underneath all of it is short. Can I open the thing, and if not, what is stopping it. That deserves a straight answer per name instead of a summary of the press coverage. So this is the state of all four on 4 September 2026: what each one is for, who can reach it, and where it stands in MultiChats, which currently runs 60 models from 18 makers on a single subscription, 21 of them on the free plan.

None of the four is in our picker today. The newest of them shipped two days ago and we expect to carry it. The other three have been outside the picker for longer, and each is outside for its own reason. One of those reasons has been sitting in our own code since August, written down at the time with a date on it.

The one that shipped this week

Google shipped Gemini 3.8 Flash on 2 September. Its API documentation lists the endpoint as gemini-3.8-flash, marks it New Stable, and describes it as "Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows". 9to5Google, covering the rollout the same day, quotes Google calling it "our most intelligent workhorse model" and reports it going out to the Gemini app for Google AI Pro and Ultra subscribers, plus AI Studio and the API. A second variant, Gemini 3.8 Flash Cyber, does not appear on Google's developer model list at all; the announcement says it is "available to trusted defenders through our new Fairwind Program".

The tempo is the part to sit with. Gemini 3.7 Flash launched on 13 August. Three weeks later there is a 3.8. Google is now shipping its workhorse line faster than most people can form an opinion about the previous one, which is a large part of why a catalogue that only ever holds one company's models ages badly.

Gemini 3.8 Flash was not in our picker on 4 September, two days after Google shipped it. The entry has to exist before the claim does. A separate piece covers what it changes for the models already in the Gemini row, and the rest of this article is about the other three.

What a model has to clear before it reaches the picker

A model reaches our picker once two questions have been answered, and both of them are about plumbing more than about the model itself. Reading the announcement and liking the benchmarks is the easy part.

The first is where the request physically lands. Every model in our catalogue reaches you through some company's servers. For 32 of the 60 that company is the maker itself, on a direct API contract. The other 28 route through OpenRouter, which sits in front of many hosts at once and picks one per request. That choice is not left open. Twenty of those entries carry a hard routing restriction that permits DeepInfra, Fireworks, Google Vertex, Amazon Bedrock, Groq or Azure and blocks everything else, including the model creator's own endpoint, any host under PRC jurisdiction, and anonymous GPU networks. A test in our build fails if someone adds a model in that category without the guard, so it cannot be forgotten in a hurry.

The second question is what the host does with your conversation. We do not train on your conversations. Some providers might, depending on which model you pick, and that is worth knowing before you type. Our OpenRouter account carries a policy setting that refuses providers which retain requests for their own purposes, and 26 of the 28 routed entries repeat a refusal at request level as well. The two that do not are the two whose info cards say plainly that requests may be briefly retained. Six models in the catalogue carry a notice like that on the card, and Grok's is the bluntest of them: "xAI may use your conversations to train Grok models."

Those two filters do real work, and sometimes they remove a model we would otherwise want.

Muse Spark, and the exact sentence in our registry

This is the one where we can show our working, because the reason was written down at the time and it has a date on it.

Muse is Meta Superintelligence Labs' line after Llama. Spark is the closed flagship, first shipped in April 2026 as the company's first frontier model without open weights. TechCrunch, writing up the smaller open sibling in August, called that sibling "essentially an open version of Meta's most powerful closed model, Muse Spark". A frontier model from a lab whose other models we already serve belongs in a catalogue like ours.

Here is what our model registry says about it, in a comment written on 11 August and unchanged since:

The Spark flagship itself is NOT here: its only OpenRouter host is Meta's own first-party endpoint, which our account privacy policy rejects outright.

The comment then records what was checked, on 11 August 2026. Every request came back as a 404 reading "no endpoints available matching your guardrail restrictions and data policy", and a request-level data_collection: allow "cannot widen an account-level setting". It ends with a single instruction to our future selves: "Revisit if a vetted third-party host picks Spark up."

Unpack that. The only place we could send a Spark request is Meta's own endpoint. Our account refuses hosts that retain requests for their own purposes, Meta's endpoint does not clear that setting, and so the request never leaves the building. The last clause closes the door properly. A per-request override cannot loosen an account-level rule, which means there is no clever flag we could set to make this go away while pretending nothing had changed.

The proof that a route is what blocks Spark is sitting in the picker already. Muse Glimmer 30B, the open weight 30 billion parameter distill of Spark released on 10 August, is in MultiChats on the free plan. It reads images, it shows its reasoning, and free accounts can push its thinking effort up to medium. Same lab, same model family, one of them in and one of them out, and the difference is entirely a question of which server the request lands on.

So the "not yet" here is precise, and it is the registry's own wording: revisit if a vetted third-party host picks Spark up. If DeepInfra or Fireworks or Vertex lists Spark tomorrow, the objection disappears and the model becomes a normal evaluation. Until then there is no route we are willing to use, and we would rather tell you that than quietly leave a gap in the list.

Fable 5.1 shipped. We do not carry any Fable model

Anthropic released Claude Fable 5.1 on 1 September. Its model page gives the identifier claude-fable-5-1, a one million token context window, 128,000 maximum output tokens, adaptive thinking that is always on, and a knowledge cutoff of June 2026. Alongside it came Claude Mythos 5.1, the same capabilities restricted to participants in Anthropic's Project Glasswing programme.

MultiChats does not carry Claude Fable 5.1. We do not carry Claude Fable 5 either, and that second fact is the more informative one, because Fable 5 was released on 9 June 2026 and has been sitting there ever since. The gap here is older than this week's release. No model in the Fable tier has been through our evaluation, and until one has, we are not going to put a card in the picker with a name on it and hope.

The interesting part is that Anthropic's own guidance points the same way for most people. Its documentation for Fable 5.1 says: "For most workloads, start with Claude Opus 5". It positions Fable for "demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short", which describes a coding and agent-building workload with a test harness wrapped around it. That is a long way from a conversation. We carry Claude Opus 5, along with Sonnet 5, Opus 4.8, Opus 4.7, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. For the work most people bring to a chat app, the vendor's recommended starting point is already in our picker.

If you specifically want Fable today, Anthropic's support page is the place to check the terms: Fable models are on every paid Claude plan and none of the free one, and on Pro and standard Team seats they sit outside the plan's usage limits and run on purchased credits instead. Max plans and premium Team seats include them up to half of the weekly limit. A sibling article covers what was actually confirmed about Fable 5.1 and the Opus 5.1 rumours.

Astra is a different kind of absent

OpenAI named Astra on 1 August, introducing it with ten results in mathematics and theoretical computer science formalised in Lean, so a proof assistant could check them instead of a referee taking the model's word. Six days later, on 7 August, the company said it had paused internal work on aspects of Astra that did not meet newly tightened security guardrails, saying its preliminary evaluations were strong enough that it could not rule out the Critical capability level in cybersecurity.

There is no release date for Astra. No price, no context window, no benchmark table anyone outside OpenAI can run, and nothing from OpenAI stating that Astra reaches ChatGPT or any chat interface at all. OpenAI presented it as work aimed at research problems that stay open for years rather than as a chat product.

That makes "Astra is not in MultiChats" a true sentence and a boring one. Nobody has Astra. There is no API to integrate, no host to vet and no card to add, so the two questions in this article do not apply to it yet. We wrote about the pause separately, because the reason OpenAI gave for it is unusual. A model with no release cannot be missing from anything.

Any piece that hands you an Astra launch date is speculating, whatever tone it is written in. The date is not coming from OpenAI, because OpenAI has not published one.

How each of these three could change

Three "no" answers, three different reasons, and they are genuinely different. Spark is blocked on a route. Fable is waiting on work we have not done. Astra has nothing to block or evaluate.

None of them is permanent, and we are not going to pretend to know which way any of them goes. The registry line about Spark says "revisit", and that revisit is a real trigger: a vetted host listing the model, followed by the same evaluation every other entry gets. Fable is a queue position. Astra is a wait for OpenAI to ship something, if it does.

The alternative would be to say nothing about the models we do not carry and let the list speak for itself. A catalogue of 60 models is a set of decisions, and some of those decisions are refusals. You are better served knowing that a Meta flagship is out because of where the request would land than by finding a Meta-shaped hole in the picker and guessing at the reason.

The list moves constantly. It moved this week: Gemini 3.8 Flash landed on Wednesday, three weeks after 3.7 Flash, and we expect it in the picker shortly. Somewhere in the next few weeks one of the other three names here changes status, and when it does the change will be for a reason we can name. The current plans and what each one includes are on the MultiChats homepage. If you want to put two of them on the same question, switching model part way through a thread keeps everything above it and hands the conversation to the next one.

Last verified: 4 September 2026.