Early September was busier than usual. Between 1 and 3 September, Anthropic, OpenAI and Google each released a new model. The density of major releases over two days was the highest of the year. Below are the facts that matter for Dutch professionals and developers.
Claude Fable 5.1: same price, 75% cheaper cache reads
Anthropic published Claude Fable 5.1 on 1 September with unchanged list prices ($10/1M input, $50/1M output), but with a significant change to the cache rate: cache reads fell by 75%, from $1.00 to $0.25 per million tokens.
In practice, this makes typical workloads 25% cheaper; for agentic applications that repeatedly load the same context, the saving can rise to 45%. Fable 5.1 has a context window of 1 million tokens and a maximum output of 128K tokens. On Terminal-Bench 4.0, the model scores 55.8% against 42.0% for its predecessor Fable 5.
The model is available via the Claude API, AWS Bedrock, Google Cloud Vertex AI and Microsoft Foundry. At the same time, Anthropic launched Claude Mythos 5.1, likewise without a price change, but with the same API changes as Fable 5.1.
See how Claude compares with other models on our comparison page.
GPT-6 Astra: OpenAI's first "AGI candidate" in the API
On 3 September, OpenAI released GPT-6 Astra, the first model the company explicitly regards as a possible AGI. The API price is $10 per 1M input and $50 per 1M output (equal to Fable 5.1), with cached input at $1 per million tokens.
GPT-6 Astra is the first model OpenAI classifies as a "critical cybersecurity risk" under its own Preparedness Framework: it can find and exploit zero-day vulnerabilities in hardened systems without step-by-step guidance. OpenAI applies activation classifiers and reinforced isolation measures. Rollout is staged via the Daybreak Access programme for API developers and has since been available to Plus, Pro, Business and Enterprise subscribers.
GDPR/AI Act point: organisations deploying GPT-6 Astra for processing personal data must explicitly address the model's higher autonomy and offensive cyber potential in their DPIA. The Dutch government has published a beta DPIA assistant that helps you record this in a structured way. See also our safety page for the GDPR profile per model.
Gemini 3.8 Flash: better benchmarks, same introductory price
Google launched Gemini 3.8 Flash on 2 September at the same rate as its predecessor: $0.75 per 1M input and $3.75 per 1M output. That introductory price is, however, temporary, from 1 January 2027, both rates double to $1.50 and $7.50. If you want the benefits of the new model without the price increase, now is the time to set up longer-term contracts and caching strategies.
Alongside 3.8 Flash, Google simultaneously launched a restricted security variant: Gemini 3.8 Flash Cyber, available only to approved partners.
Aithos research: no major model meets the EU compliance standard
Earlier, in June 2026, the Amsterdam AI institute Aithos published the LARA study: more than 3,000 tests on 12 frontier models, including systems from Google, OpenAI and Anthropic. The best model achieved 54% compliance; the worst only 7%.
The tests assessed six provisions of the EU AI Act and four GDPR principles (transparency, lawful processing, data minimisation, purpose limitation). Problems: models generated outputs that facilitated prohibited practices, such as emotion recognition in the workplace or manipulation of vulnerable users. Aithos points out that legal liability lies with the organisation that places the AI system on the market or puts it into service, not with the model provider. Fines can reach €35 million or 7% of global turnover.
For Dutch organisations using AI in HR, customer service or decision-making, independent compliance testing is therefore no longer a nice-to-have. See which sectors are most affected by the high-risk AI requirements of the AI Act.
Sources: VentureBeat (Fable 5.1) · Axios (GPT-6 Astra) · CellCog (Gemini 3.8 Flash) · Digital Watch Observatory (Aithos LARA)