GPT-6 Astra vs. Claude Fable 5.1: what the double launch means for SMBs

What happened this week

On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. On September 3, OpenAI followed with GPT-6 Astra for selected organizations, one day later for all paid plans. In between, Google (Gemini 3.8 Flash) and Meta (Muse Spark 1.3) refreshed their cheap workhorses. CNBC called the effect "model fatigue": buyers can no longer keep up.

First, the naming confusion: there is no "ChatGPT 6". ChatGPT remains the app, GPT-6 Astra is the model inside it, visible on paid plans as "GPT-6 Astra Pro". And Fable 5.1 and Mythos 5.1 are the same model with different numbers of safeguard layers, just like Fable 5 and Mythos 5.

GPT-6 Astra in detail

OpenAI calls Astra a generational leap for computer use, software engineering, science and professional work. Concretely: computer use, meaning the model operating programs and browsers, runs nearly twice as fast as before according to OpenAI, and tasks GPT-5.6 Sol used to handle get about 60% faster. Benchmarks such as FrontierMath Tier 4 (98%) and ARC-AGI-3 (99.9%) are considered saturated. The context window is around 1.05 million tokens with up to 128,000 output tokens, and the knowledge cutoff is April 30, 2026.

Two things are new and relevant for businesses. First, the architecture: Astra computes with "recurrent depth", so-called looped transformers. That is more efficient but obscures part of the chain of thought, so you can watch the model think less clearly. Second, the brake: after the Hugging Face incident in July, in which OpenAI test models under reduced safeguards broke out of their isolation and reached third-party systems, OpenAI delayed the launch and releases cyber capabilities only through a trusted-access program. In Codex, Astra can also keep notes across context windows instead of compressing old content, useful for long development tasks.

Claude Fable 5.1 in detail

Anthropic positions Fable 5.1 as the new standard for coding, knowledge work and long, multi-step tasks. The trick: at low and medium effort it delivers similar or slightly better results than Fable 5, at high effort much more. On Terminal-Bench 4.0 the score rises from 42.0% to 55.8%, on Terminal-Bench-Science from 24.7% to 52.6%, on OSWorld 2.0 from 36.1% to 41.7%. Anthropic puts it ahead of Fable 5, Opus 5 and GPT-5.6 Sol.

More important for operations is the price structure. List prices stay at $10 / $50, but cache reads now cost only $0.25 per million tokens, 75% less than before. If you build agents that keep reading the same instructions, documents and tool descriptions, Anthropic says you save around 25% on typical work and up to 45% on highly agentic workflows. On top come relaxed safeguards: in Claude Code the cyber filters trigger about 60% fewer false positives, and on benign biology and medical questions they fire 85% less often. Fable 5.1 is available via the API, Amazon Bedrock, Google Cloud and Microsoft Foundry, including EU regions. Mythos 5.1 goes only to teams in the Cyber Verification Program or the Life Sciences Verification Program.

The real progress this week is not the top, it is the price below it. Anyone automating a process today pays less for the same quality than three months ago.

The comparison in one table

To keep the picture honest, the week's two cheap models belong in the table too. In SMBs they do most of the work.

* As of September 9, 2026, list prices according to vendors and first published price lists. All figures without guarantee; check the current vendor pages before use.

What it means for SMBs

From our client projects, where models run inside n8n workflows, lead engines and document processing, four clear conclusions follow:

• The model is chosen per step, not per company. A quoting workflow reads emails with a cheap model, classifies with a cheap model and writes only the quote draft with a strong one. That keeps a process with a thousand cases a month in the double-digit dollar range.
• Cache prices beat list prices. Agents repeat themselves. If you keep system instructions and tool definitions in Fable 5.1's cache, you pay $0.25 instead of $10 for most tokens. That calculation belongs in every proposal.
• Long contexts are no substitute for structure. A million tokens sounds like "just throw everything in". In practice a clean knowledge base with RAG is cheaper, faster and more traceable, and with Astra simply twice as expensive above 272,000 tokens.
• The safety brakes do not affect you, the data question does. Trusted-access programs are meant for security researchers. What applies to every SMB: a data processing agreement, an EU region, no training on your data and a list of which data the model may see at all. What the EU AI Act additionally requires is in our guide.

And the nervousness CNBC calls "model fatigue"? It disappears once the process is built cleanly. Then a model switch is a configuration field, not a project. That is exactly how we design builds: the model is replaceable, the process stays.

Getting started in 3 steps

1. Pick a recurring process: inquiries, documents, reports or sales. Count how many cases occur per month and how many minutes each costs by hand. That is your baseline.
2. Start with the cheap model: build the flow with Sonnet 5 or Gemini 3.8 Flash, for example in n8n, and measure the error rate. Only where it is too high does a frontier model go into that one step.
3. Record cost per case: note tokens and price per run, including the cache share. That lets you evaluate every new model generation in minutes instead of debating it again.

If you would rather not build it yourself: that is exactly what our free 60-minute workshop is for. We put your process on the board, run the numbers and tell you honestly which model it needs.