This observation covers the period from 8:11 am on 1 September to 8:11 am on 2 September 2026 in Australia/Sydney time. Three developments that were formally released, formally classified or opened for actual use are included.
Claude Fable 5.1 becomes generally available while Mythos 5.1 remains restricted
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. They use the same underlying model but apply different safeguards. Fable 5.1 is available to all customers through the Claude API and on AWS, Google Cloud and Microsoft Azure, with the API model ID claude-fable-5-1. Mythos 5.1 is limited to vetted cybersecurity and life-sciences participants and is currently available only to selected US organisations.
Both versions provide a one-million-token context window and a maximum output of 128,000 tokens, with adaptive thinking always enabled. The new models do not support forced tool use: setting tool_choice to any or to a named tool returns an error. Earlier Claude models also cannot read thinking blocks created by Fable 5.1. Cache reads now cost US$0.25 per million tokens, while standard input and output prices remain US$10 and US$50 per million tokens. Anthropic reports improvements over the previous model on Terminal-Bench, OSWorld and research tasks, and estimates that the cache-price change reduces the cost of typical workloads by about 25%. These capability scores and cost ratios are primarily developer evaluations and have not received a complete public independent reproduction. The Verge independently confirmed the release, availability boundaries and pricing changes.
OpenAI gives an unreleased model its first Critical cyber classification
On 1 September, OpenAI formally determined that the model under development known as Astra meets the Critical cybersecurity capability threshold in its Preparedness Framework. The threshold covers a model that, with suitable tools and access, can find unknown vulnerabilities and develop ways to exploit them across multiple protected systems without step-by-step human guidance, or execute a novel end-to-end attack strategy.
OpenAI reports that Astra scored 100% on ExploitBench and achieved a higher arbitrary-code-execution rate than GPT-5.6 Sol, with fewer output tokens, on an internal evaluation containing 20 recently disclosed high-severity V8 vulnerabilities. Two zero-day vulnerabilities found during testing are being disclosed to maintainers. These are developer-reported results. The internal dataset, vulnerability details and full system card are not yet public, and there is no independent reproduction. Astra has not been released. OpenAI says only that limited testing access will begin soon, that its strongest cyber capabilities will remain restricted, and that refusal training, misuse protections and monitoring able to stop potentially unauthorised activity have been added. Reuters confirmed that this is the first OpenAI model to trigger safeguards at this level and that no specific release date has been provided.
Waymo begins accepting public riders in three cities
On 1 September, Waymo began gradually welcoming its first public riders to fully autonomous services in Denver, San Diego and Tampa, bringing the number of cities in which it provides fully autonomous trips to 14. Compared with earlier road testing, employee rides or invitation-only trials, the new status allows members of the public to register through the Waymo app and progressively receive access.
The rollout remains phased rather than immediately available to every user in the three cities. Waymo did not disclose initial fleet sizes, dates for unrestricted access or operating results from this launch. Reuters independently confirmed the public-rider status. It also reported that Zoox's new activity in Houston and San Diego begins with manually driven mapping and testing, so it is not an operational driverless commercial service in those cities.
Sources
Anthropic, Claude Fable 5.1 and Mythos 5.1, 1 September 2026
https://www.anthropic.com/claude-fable-and-mythos-5-1
The Verge, Anthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work, 1 September 2026
https://www.theverge.com/ai-artificial-intelligence/987830/anthropic-claude-fable-mythos-5-1
OpenAI, Path to Astra: critical capabilities and frontier safeguards, 1 September 2026
https://openai.com/index/path-to-astra/
Reuters, OpenAI says upcoming model is so capable it requires stronger guardrails, 1 September 2026
https://www.reuters.com/business/openai-says-upcoming-model-is-so-capable-it-requires-stronger-guardrails-2026-09-01/
Waymo, Welcoming our first riders in Denver, San Diego, and Tampa, 1 September 2026
https://waymo.com/blog/2026/09/ride-in-denver-san-diego-tampa
Reuters, Amazon's Zoox, Alphabet's Waymo expand robotaxi services to more US cities, 1 September 2026
https://www.reuters.com/business/amazons-zoox-alphabets-waymo-expand-robotaxi-services-more-us-cities-2026-09-01/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.