Fable 5.1 is Here, OpenAI's New Update is Still Pending

Anthropic releases Fable 5.1 and promises more efficiency per Euro

Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, two models with identical weights that differ only in the level of their safety guardrails. Fable 5.1 is generally available and delivered via AWS, Google Cloud, and Azure. Mythos 5.1 is only available through a vetted access program, open to cyber defenders and life scientists at US organizations. For the biology capabilities, Anthropic has set up an access program together with the US government; registration for researchers is expected to open soon.

The base prices remain unchanged: $10 per million input tokens, $50 per million output tokens. The price for Cache Reads has been lowered from $1 to $0.25 per million tokens. According to the provider, this makes typical workloads around 25 percent cheaper than with Fable 5, and up to about 45 percent for highly agentic tasks. An independent preliminary evaluation presents a different picture per task: $3.76 per task in the Intelligence Index, 20 percent more than Fable 5, because the model uses about 1.7 times as many output tokens; the cache reduction saves around $1.40 in this case.

In terms of metrics, the new science benchmark stands out: On Terminal-Bench-Science 0.1, Fable 5.1 achieves 52.6 percent compared to 24.7 percent for Fable 5, according to Anthropic. On Terminal-Bench 4.0, Fable 5.1 scores 55.8 percent and Mythos 5.1 scores 60.9 percent, quantifying the capability gap between the two safety levels. Humanity’s Last Exam is at 60.9 percent without tools, GDPval-AA v2 at 1,853 Elo. In external measurements, the model achieves 66 points in the Intelligence Index, the highest score ever recorded there, ahead of Claude Opus 5 with 63 and GPT-5.6 Sol with 61.

According to Anthropic, it is also addressing customer criticism regarding data retention and overly strict safety mechanisms. The announced Enterprise Frontier Safeguards are intended to store data on customers' cloud servers instead of Anthropic’s and will be rolled out over the course of the fall. Fable 5.1 is designed to block basic biology questions less frequently than its predecessor, while Mythos 5.1 retains the same restrictions. For cybersecurity, Anthropic now allows Fable 5.1 to identify software vulnerabilities, but continues to refer tasks like penetration tests, exploit generation, and binary-based vulnerability scanning to the Opus models. Among the early-access customers, Every CEO Dan Shipper and Box CEO Aaron Levie, among others, commented positively on the model’s speed and accuracy.

Specific real-world cases are being reported: At the investment firm Millennium, the model found the cause of a rare crash that the firm’s own engineers had failed to solve for years. Cognition is shifting its Opus-5 traffic in Devin to Fable 5.1 at launch, and Jane Street reports better readability over long task chains. In editorial tests, the model completed agent tasks in about 60 percent of the time of Opus 5, but delivered 43 quotes in response to a request for eight to twelve, including quotes missing from the source, and temporarily ignored stop instructions at the highest effort setting. The release comes at a time when an IPO from Anthropic is expected and OpenAI’s Astra model is still pending. → anthropic, artificialanalysis, axios, AI Weekly, Every

Synthszr Take: A model that raises the science benchmark from 24.7 to 52.6 percent is shifting the hierarchy among institutions faster than any appointments committee. The real turning point lies in the access model: The biology capabilities of Mythos 5.1 are distributed through a program run by Anthropic in conjunction with the US government, and it is open to vetted organizations in the US. This means a provider and a government jointly decide which labs get to work with the best available co-researcher, while peer review presupposes that the reviewers have the same tools at their disposal. The Millennium case shows the direction: A cause that engineers failed to find for years is discovered in a single run, and such results first emerge where a token budget is available, not in a university department with a third-party funding application. The next reputational advantage in research will be established at the negotiating table for access rights, and universities have rarely had a seat there so far.

OpenAI classifies Astra as a critical cyber risk and restricts access

OpenAI has reportedly classified Astra as the first model at the “Critical” level for cybersecurity capabilities, the highest category in its in-house Preparedness Framework. According to OpenAI, this threshold is considered reached when a model can develop functional Zero-Day-Exploits in hardened real-world systems without human guidance or build a complete attack chain from a vaguely formulated objective. On the public ExploitBench, Astra achieved a score of 100 percent, according to the company. Due to potential contamination of the training data, OpenAI also built an internal test set of 20 recently disclosed V8 vulnerabilities. During this evaluation, the model reportedly found two previously unknown vulnerabilities and incorporated them into an exploit chain; disclosure to the maintainers is still in progress. In expert-led tests, Astra is also said to have broken out of a hardened browser sandbox and engineered a privilege escalation to root in an operating system. → OpenAI

Synthszr Take: The figures presented by OpenAI are from the Daybreak-Blue configuration and not from the production version that will be delivered later. This means that nothing about this 100 percent score on ExploitBench is externally verifiable. For security researchers outside the company, this classification effectively means a waiting period, as the advanced capabilities are initially available only to a select group of testers, and the admission criteria are not published anywhere. Additionally, OpenAI has trained the model to more reliably reject malicious cyber requests, and these guardrails first affect the people who are legitimately searching for vulnerabilities in their daily work (a pentest prompt, after all, reads like an attack)..

World Labs introduces Atlas, a world model that builds 3D scenes from a few photos

World Labs, the startup founded by computer vision researcher Fei-Fei Li in February 2024, has unveiled its world model, Atlas. According to siliconangle.com, Atlas generates up to one minute of 1440p video from a single 2D image, which can be viewed from any angle while remaining geometrically consistent. Technically, the company describes the architecture as a multimodal autoregressive diffusion transformer, that processes camera paths and geometry as native inputs, rather than being controlled by text prompts. In addition to video frames, the model also outputs 3D assets such as point clouds and Gaussian Splats, according to the provider. World Labs names robotics as its target field: developers are meant to capture a real space with a standard smartphone camera and reconstruct it as a simulation in which a robot can navigate, receiving RGB images and depth sensor data. In a blind evaluation conducted by the company itself on camera path adherence, Atlas was reportedly preferred over Gemini Omni Flash and FLUX. → siliconangle.com

Synthszr Take: One minute of 1440p video from a single smartphone photo, plus depth data and point clouds: that’s a lot of useful material for a robot’s perception training. Navigation and object recognition under varying lighting conditions can be simulated more cheaply this way than with a hall full of physical setups. But it stops at grasping, because the friction and compliance of an object are not contained in any reconstructed geometry, and this is precisely where grippers fail in real-world operation.

Meta’s Muse Voice Transcribe undercuts Google Cloud’s price by 80 percent

Meta Superintelligence Labs yesterday released Muse Voice Transcribe, the lab’s first real-time model that combines streaming transcription, speaker separation, and end-of-speech detection in a single model. Via the Meta Model API, the service costs $3 per 1,000 audio minutes, or $0.18 per hour; Google Cloud Speech-to-Text charges $0.96 per hour in its standard tier. On the independent AA-WER streaming benchmark from Artificial Analysis, Muse achieved a word error rate of 3.1 percent, according to Implicator.ai, ahead of Cartesia Ink-2 with 3.4 percent, ElevenLabs Scribe v2 Realtime with 3.6 percent, and Google’s Gemini 3.5 Transcribe Live with 4 percent. The test is based on about eight hours of English audio and says nothing about the more than 70 training languages, of which Meta recommends 25 as verified at launch. For diarization, i.e., assigning speech segments to individual speakers, Meta itself reports an error rate of 17.5 percent. → Implicator.ai

Synthszr Take: $0.18 versus $0.96 per hour, that’s a factor of five, and next to that, a 0.3 percentage point lead in word error rate looks like window dressing. Meta isn’t just undercutting Google: Cartesia is at $4 per 1,000 minutes, ElevenLabs and Deepgram at $6.50, and they all play in the same error class between 3 and 4 percent. If you calculate one million transcription hours a year, that’s a $780,000 difference compared to Google, and for that, a buyer will accept a 17.5 percent error rate in speaker diarization without a second thought.

According to WSJ, Google is about to launch a coding model called 'Skimaki'

According to a report by the Wall Street Journal, Google DeepMind is launching a model focused on programming in the coming days: Gemini 3.8 Flash, internally called 'Skimaki', could be released as early as Wednesday. The newspaper relies on anonymous insiders who attribute significantly improved coding capabilities to the model. It was reportedly tested with an internal tool called 'Jetski'; the developers involved are said to prefer it over Anthropic’s Claude Opus. Unlike OpenAI’s as-yet-unreleased Astra, Skimaki does not belong to the trillion-parameter range but to the Flash series, which is designed for speed and low cost. → Gizmodo

Synthszr Take: In March, OpenAI pivoted to coding, now Google is following suit, and the pattern has become so reliable that you can read one lab’s roadmap from its neighbor’s (we described this copycat carousel back in early May, it’s just spinning faster). More interesting than the imitation is the internal trigger: Hassabis, who was still hitting the brakes last year, has not been DeepMind’s CEO for a month, Brin is pushing the pace, and four weeks later, a coding model is on the doorstep. The fact that Google is choosing a Flash model for this and not the trillion-parameter beast shows where the math works out: agents burn through tokens every second, so fast and cheap beats the biggest model on the shelf.

OpenAI counters Apple’s lawsuit: Apple failed to properly protect its own secrets

In the ongoing dispute over trade secrets, OpenAI is deflecting responsibility, stating that Apple has only itself to blame for any potential leak. Reuters reports that OpenAI is advocating this position in a legal filing in the case where Apple alleges the theft of confidential information. According to OpenAI, the core of its defense is that Apple did not adequately protect the information in question; this has not been legally established. Legally, this targets a central point of U.S. trade secret protection: a plaintiff must prove that they took reasonable measures to maintain secrecy, otherwise the information is not legally considered a protectable secret. → Reuters

Synthszr Take: OpenAI is pulling the most inconvenient lever that trade secret law provides: a secret only exists legally if you can prove you treated it like one. For decades, Apple has lived on its reputation as the most secretive company in the industry, and now that very reputation is becoming a burden of proof. In practice, secrets are compromised by access rights that were never revoked upon departure and by shared repositories that entire departments can view.

Physical Superintelligence launches with a $58 million seed for virtual physicists

On September 1, 2026, Matt Pines, Alex Klokus, and Alexander Wissner-Gross brought the company Physical Superintelligence (PSI) out of stealth and announced a $58 million seed round led by Breakthrough Energy Ventures. According to the announcement, participants include Dragon Global, SV Angel, Susa, as well as individual investors from the circles of OpenAI, Nvidia, Oracle, and Hugging Face. The core of the offering is the platform Emmy, named after the mathematician Amalie Emmy Noether: According to the provider, it breaks down research questions into trees of verifiable hypotheses and tests them in parallel in a curated set of simulations. PSI names the Multiphysik design of AI data centers, both terrestrial and orbital, across power, cooling, network, and compute as its first commercial application. In addition, PSI is a technical founding partner of the Fermi Explorer Mission, a non-profit project aiming to send a probe with a one-kilogram payload to Alpha Centauri before the end of 2029, for a total cost of less than $15 million. → Unite.AI

Synthszr Take: The feasibility study notes that every quantitative result was generated by the platform itself and double-checked by its own verification agents, without full external peer review. Parallelized hypothesis trees accelerate what physicists already delegate to machines; physics is still decided by experimental setups, and those aren’t found in any set of simulations. The first revenue comes from power, cooling, and network design for data centers—in other words, from very well-paid engineering optimization being sold here under the headline of new laws of nature.

ChatGPT Ads reach a $1 billion run rate, Europe gets self-serve access

OpenAI puts the value of its advertising business in ChatGPT at an annualized Run Rate of one billion dollars, less than six months after its launch. This figure is derived by extrapolating the current monthly revenue times twelve, thus corresponding to about $83 million per month and not the year-to-date revenue. In parallel, since August 31, the provider has been opening up its Ads Manager as a beta in self-serve access for approved advertisers in 31 European markets, who previously could only book through selected agency partners. This allows small businesses and startups to set up campaigns without an agency, provided their category is approved and the ads pass internal policy reviews. Dave Dugan, VP of Global Ad Solutions at OpenAI, speaks of a 'new chapter for advertising', according to the company, and sees the billion in under 200 days as proof of the size of the opportunity. → Techpresso

Synthszr Take: $83 million a month is money someone is paying to appear in an answer that the user reads as advice, not as an ad. On Google, the result is next to an ad block; in the chat, the recommendation is in the same flow of conversation as its justification. As soon as brands can book their own ads via the self-serve manager in 31 markets, every product mention will be read with a silent question: Is it there because it’s good, or because it was paid for?

John Ternus takes over Apple: Cook’s legacy now hinges on the AI strategy

John Ternus officially took over as CEO of Apple on September 1, succeeding Tim Cook after fifteen years at the helm. Ternus is 51, has been with the company since 2001, and made a name for himself as the head of the hardware division. Cook is not leaving Apple: he will remain on board as Executive Chairman, responsible for, among other things, liaising with political decision-makers worldwide, according to the company. Ternus’s first public appearance as CEO will be at the annual product event on September 9, where reports suggest the first foldable iPhone will be unveiled. Cook will be in the audience but will not take the stage.

Cook’s financial record is undisputed. The stock rose by around 2,000 percent during his tenure, the market capitalization briefly exceeded five trillion dollars in July, and from 72 million iPhones sold in 2011, it will be 255 million this year, according to estimates by Counterpoint Research. In addition, there is the services business with Apple TV+, Apple News, and iCloud, which has converted hundreds of millions of device owners into subscribers. On the downside are the canceled car project and the Vision Pro, which survives as an expensive device for VR enthusiasts without having found a mass market.

The open question is the AI strategy. The vision of Apple Intelligence presented at WWDC 2024 was designed to be privacy-centric, with personal data on the device or in a protected cloud environment. It was followed by delays, and the centerpiece, the revised Siri, is only now arriving with iOS 27, with Apple relying on models from Google for it. Forgoing its own frontier model saves Apple the sums that Google, Meta, Microsoft, OpenAI, and Anthropic are investing in model development and infrastructure, but it makes the company dependent on Google for central AI functions.

Ternus also inherits a personnel problem and a cost issue. Foundation Models chief Ruoming Pang, Engineering chief Frank Chu, and Design VP Alan Dye have moved to Meta, and hundreds of former Apple employees to OpenAI. Because AI companies are buying up chips on a large scale, memory is becoming scarcer and more expensive, and Apple is passing these costs on to customers. In the pipeline, according to reports, are an AI home hub, AirPods with a camera, smart glasses, and a robot assistant. → Morning Brew, Axios AI+, Business Insider

Synthszr Take: Steve Ballmer tripled Microsoft’s revenue and is still remembered as the man who missed the boat on the smartphone. Cook has the more impressive numbers, 72 million iPhones in 2011 versus 255 million this year, and carries the same risk: a single platform question retroactively decides fifteen years of operational excellence. Nokia had the better factories in 2007, Blackberry the more popular keyboard, and both lost on a software bet they considered secondary. Apple is now borrowing Google’s models for Siri and has simultaneously lost its head of Foundation Models to Meta, which is an uncomfortable combination for the coming years. The verdict on Cook will be delivered on the day that OpenAI or Anthropic bring their own hardware into series a production; the foldable iPhone from September 9th counts for little in this.