Weekly Recap · Week 36/2026

AI Recap: GPT-6 Astra, Claude Fable 5.1 and the week four labs shipped at once

A packed week: Anthropic ships Claude Fable 5.1, OpenAI follows with GPT-6 Astra, and in between Google and Meta push cheap workhorses. Buyers are talking about "model fatigue" for the first time. What matters for small and mid-sized businesses.

Jan Malte SanderFounder · BitsAndBucks GmbH · September 9, 2026 · 7 min read

The week in one sentence

Four frontier models in four days: Claude Fable 5.1 gets cheaper and stronger, GPT-6 Astra arrives with its cyber capabilities deliberately throttled, Gemini 3.8 Flash and Muse Spark 1.3 push prices down. For SMBs that means the model becomes a per-task choice, not a matter of faith.

September 1 · Models

Claude Fable 5.1 and Mythos 5.1: more power at high effort, a quarter cheaper

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Both are the same model with different levels of safeguards: Fable 5.1 is generally available, Mythos 5.1 only through verification programs for cybersecurity and life sciences. At low and medium effort Fable 5.1 delivers similar or better results than Fable 5, at high effort much more: 55.8% instead of 42.0% on Terminal-Bench 4.0, 41.7% instead of 36.1% on OSWorld 2.0. More interesting for the budget: prices stay at $10 / $50 per million tokens, but cache reads now cost $0.25, about 75% less. According to Anthropic, that cuts costs by around 25% for typical work and up to 45% for highly agentic tasks. In Claude Code the cyber safeguards trigger about 60% fewer false positives. What all of this means day to day is in our GPT-6 Astra vs. Claude Fable 5.1 comparison.

Source: Anthropic Source: 9to5Mac

September 3 → 4 · OpenAI

GPT-6 Astra is here: "generational leap", cyber capabilities behind a locked door

OpenAI released GPT-6 Astra to selected organizations on September 3 and to all paid plans on September 4 (Plus, Pro, Business, Enterprise plus the API and AWS). In ChatGPT it shows up as GPT-6 Astra Pro; there is no "ChatGPT 6". OpenAI calls it a generational leap in computer use, software engineering, science and professional work: computer use runs nearly twice as fast according to OpenAI, and FrontierMath Tier 4 and ARC-AGI-3 are considered saturated. Technically new is a recurrent-depth architecture ("looped transformers") that computes more efficiently but partly obscures the chain of thought. The context window is around 1.05 million tokens, the knowledge cutoff late April 2026. The cyber capabilities are the most discussed and most restricted feature: after the Hugging Face incident in July, in which test models broke out of their isolation, OpenAI delayed the launch and releases security-relevant functions only through a trusted-access program. Via the API, early price lists put Astra at $10 / $50 per million tokens, double that above 272,000 input tokens.

Source: 9to5Mac Source: Wikipedia Source: OpenAI

September 2 · Prices

Gemini 3.8 Flash and Muse Spark 1.3: Google and Meta fight over the workhorses

Between the two flagships, Google and Meta made it about price. Gemini 3.8 Flash still costs $0.75 / $3.75 per million tokens, beats its predecessor on every benchmark Google published and reaches 90.8% on Terminal-Bench 2.1, with a one-million-token context and a locked-down cyber variant. Google did announce that both prices double on January 1, 2027. Meta followed with Muse Spark 1.3: $1.25 / $4.25 on the standard endpoint, and a "contributor" endpoint at roughly $0.10 / $0.20 where Meta may train on your traffic. For SMBs this is the real news of the week: the mid-range is getting better faster than the top is getting more expensive.

Source: The Register Source: 9to5Google Source: Artificial Analysis

August 31 → September 6 · Market

"Model fatigue": buyers can't keep up, Anthropic secures $35 billion of compute

CNBC gave the week its name: model fatigue. Runpod CEO Zhen Lu says the pace is disorienting IT buyers and that "you have to make noise" to stand out. In parallel, more than 1,100 lab employees petitioned Washington to help pace frontier development. Infrastructure keeps growing regardless: according to Bloomberg, Anthropic signed a $35 billion cloud deal with Lambda for a Hut 8 data center in Nueces County, Texas, with Nvidia holding the lease. And OpenAI reports that ChatGPT advertising reached a $1 billion annualized run rate in under 200 days, with advertisers in more than 40 countries.

4 in 4 frontier models in four days: Fable 5.1 (Sep 1), Gemini 3.8 Flash and Muse Spark 1.3 (Sep 2), GPT-6 Astra (Sep 3/4). CNBC calls the result "model fatigue".CNBC, September 6, 2026

Source: CNBC Source: Bloomberg

September 1 · DACH

Scalable Capital opens portfolios to AI assistants, Nvidia reportedly buying Hugging Face

Closer to everyday life: German broker Scalable Capital lets customers connect their portfolio to ChatGPT, Claude or Grok. The assistants analyze portfolios, pull news and market data, manage watchlists, set price alerts and create orders or savings plans; every transaction needs explicit confirmation. The broker calls itself the first bank in Europe to open its platform to the major assistants. To us this is a pattern that will show up in every industry soon: the assistant operates the tool, the human approves. The Information also reports that Nvidia wants to acquire Hugging Face for $12.9 billion. Neither side has confirmed it; for the distribution of open models it would be a turning point.

Source: Digitale Profis

Sources & further reading

As of September 9, 2026. Details on models, prices, benchmarks and contracts as stated by the companies and the cited media, without guarantee. GPT-6 Astra API prices come from the first published price lists and may change at general API availability.

FAQ on this AI week

What is GPT-6 Astra and who gets it?

GPT-6 Astra is OpenAI's new flagship model, available to selected organizations since September 3, 2026 and to ChatGPT Plus, Pro, Business and Enterprise users as well as via the API and AWS since September 4. In ChatGPT it appears on paid plans as "GPT-6 Astra Pro". Its cyber capabilities are restricted through a trusted-access program.

What is new in Claude Fable 5.1?

Claude Fable 5.1 was released on September 1, 2026. It delivers similar or better results than Fable 5 at low and medium effort and much more at high effort, for example 55.8% instead of 42.0% on Terminal-Bench 4.0. Thanks to cache reads that are 75% cheaper, it costs around 25% less for typical work and up to 45% less for highly agentic tasks. Mythos 5.1 is the same model with fewer safeguards for verified security and life-science teams.

Is there a "ChatGPT 6"?

No. OpenAI has not released a product called ChatGPT 6. The version is called GPT-6 Astra, and ChatGPT remains the app in which the model is selected on paid plans as GPT-6 Astra Pro.

Should SMBs switch models now?

Only if a concrete process benefits. For standard tasks such as email triage, quote drafts or document recognition, cheap models like Claude Sonnet 5 or Gemini 3.8 Flash are enough. GPT-6 Astra and Fable 5.1 pay off for long, multi-step agent tasks and coding. The sensible approach is a multi-provider strategy that picks the model per task.

Which model for which process?

In the free 60-minute workshop we put your process on the board and tell you honestly whether it needs a frontier model or a cheap one will do.

Request a workshop

Jan Malte Sander

Founder of BitsAndBucks GmbH. Builds automation for companies that have no time for it, and follows AI development daily because it ends up in real client projects. LinkedIn