GPT-6 Astra and Fable 5.1: What Changes for Agents

Bottom Line

Two frontier models in 48 hours, both pointed at the same job: agents that do long, multi-step work on a computer.

Neither runs on your hardware. What they change is the price and quality of the calls your local stack routes outward — and for anyone running a hybrid setup, that is the number that matters.

GPT-6 Astra

Released September 3, 2026. API model ID gpt-6-astra.

• 1 million token context window
• Built explicitly for agentic and computer-use tasks: software engineering, cybersecurity, science, professional work
• Manages multi-step workflows across a computer and browser, holds its original instructions across them, and OpenAI cites strong visual judgment
• $10 per million input tokens, $50 per million output

Availability is staged. It opened to a limited set of organisations under OpenAI’s Daybreak Access program, with ChatGPT Plus, Pro, Business and Enterprise plus API developers following. Cyber-sensitive capabilities sit behind a separate trusted-access program.

OpenAI’s framing has been unusually strong — Greg Brockman has described it as the start of AGI, and coverage has repeated the claim. Treat that as positioning until independent benchmarks land. What is checkable today is the context window, the computer-use focus, and the price.

Fable 5.1

Released September 1, 2026, two days earlier.

Anthropic’s stated improvements are coding and science, and cost. The specific claim is that it handles lengthy, complicated programming work better — multi-file software projects, code reviews — while being more economical than the model it replaces.

Claude Mythos 5.1 is the same underlying model with full cyber and biology capabilities retained, under limited access.

A correction worth making, since it circulates: there is no Fable 5.5. The current release is 5.1. Fable 5 arrived in June 2026 alongside Mythos 5.

What this changes for a local stack

Nothing about the case for local weights, and quite a lot about the routing.

The cost shape is the whole argument. A long agent run on a frontier model is metered per token: a run that thinks for an hour costs proportionally more than one that thinks for a minute. The same run on local weights costs electricity, and costs the same whether it takes a minute or a day.

At $10 and $50 per million tokens, an agent that reads a large codebase, reasons across it, and writes back is not a cheap call. A 1M-token context is a capability and a cost multiplier in the same sentence.

So the sensible pattern is the one it already was:

• Local models for bulk and repetition — classification, extraction, formatting, first-pass drafts, anything that runs constantly
• Frontier models for the hard call — the architecture decision, the ambiguous review, the thing where being wrong is expensive

What changed this week is that the hard call got better and, on Anthropic’s side, cheaper. That makes the split more attractive, not less.

The computer-use part is the real shift

Worth separating from the benchmark noise.

Astra is built to operate a computer and a browser across multi-step workflows while holding its instructions. That is precisely what a self-hosted agent stack does today with a local model and a lot of scaffolding.

If a hosted model does that reliably, some of the scaffolding stops earning its place. Not the local inference — the glue around it.

This is the thing to watch over the next month, and it is not answerable yet. Nobody outside the early-access programs has run it long enough to say whether it holds instructions over hours the way the announcement implies.

What we have not tested

Neither model, on our hardware or anyone else’s. Everything above comes from the vendors’ own releases and launch coverage, both under a week old.

Specifically unverified: whether Astra’s instruction-holding survives a genuinely long run, whether the 1M context is usable at the top end or degrades like most long contexts do, and whether Fable 5.1’s cost improvement holds on agent-shaped workloads rather than benchmarks.

We will update this once there is independent testing worth citing. In the meantime, the local half of the stack is unchanged, and the 64GB guide remains the practical starting point for the bulk work these models are too expensive to do.