Anthropic Fable

Evidence (17)

• 7The New Stack·10/6/2026security
  Anthropic Expands Cyber Verification Program

  Anthropic is launching an expanded Cyber Verification Program with three access tiers for qualified security professionals. The program gives vetted teams access to stronger model capabilities and reduced cyber safeguards for defensive work, red teaming, and specialized safety-system testing.

• 5Hacker News·9/23/2026research
  GPT-6 Astra tops a car-driving benchmark

  A Hacker News post shows GPT-6 Astra achieving the best result in a car-driving benchmark, completing the course on its second attempt. The results table compares it with other models, while GPT-6 Astra reaches 100% progress and finishes in 5:22.

• 7The New Stack·9/21/2026model_release
  Grok 4.7 improves long-horizon coding, but still lags top rivals

  SpaceXAI released Grok 4.7 with training focused on long-running, multi-hour tasks, emphasizing self-verification and long-context management. The model improved on endurance benchmarks such as Terminal-Bench 4.0 and CursorBench 4.0, but still trails Anthropic’s Claude Fable 5.1 on Terminal-Bench and remains unreliable in many long coding workflows.

• 5Hacker News·9/21/2026research
  Anthropic Fable 5 shows a drop in thinking tokens in August

  A Hacker News post reports that Fable 5, after being made permanently available in Anthropic subscription plans, appeared to perform worse than before. The author says measurements across five methods showed August produced far fewer thinking tokens than July, suggesting a decline in the model's reasoning output.

• 7Ars Technica AI·9/15/2026research
  Open AI models nearly match frontier models at much lower cost

  A Mozilla report says the performance gap between frontier AI models from US companies and top open-weights models from Chinese companies has narrowed to 4.4 months. It also finds that open models are often the better default for routine work, while closed frontier models mainly justify their premium for expert work, long-context tasks, and heavy retrieval.

• 7The New Stack·9/14/2026research
  Real-SWE benchmark shows top coding agents fail most of the time

  The Real-SWE benchmark tested coding agents on private, real-world company codebases and found that Claude Fable 5.1 led with only 38.8% success, still failing more than 60% of attempts. Results varied sharply by model and task, and no agent was consistently reliable across unfamiliar production code.

• 6Hacker News·9/13/2026research
  Claude Fable 5.1 deciphers a 370-year-old cipher

  Claude Fable 5.1 reportedly solved Sir Thomas Urquhart’s Cyphral Distich by using the book’s own structure as the key, rather than an external cipher alphabet. The same approach was then applied to Urquhart’s larger Cyphral Octastich, yielding a mostly complete plaintext and suggesting a historical text-based encoding method.

• 7Simon Willison·9/11/2026security
  Datasette 1.0a39 and 0.65.4 security fixes released

  Datasette has released two security patch versions, 1.0a39 for the alpha line and 0.65.4 for the stable 0.65.x series. The fixes are recommended for public-facing instances, especially deployments that mix public and private tables, after an audit uncovered subtle security issues.

• 5The New Stack·9/10/2026research
  Fable 5.1 and Fable 5 tested on real benchmark tasks

  The article compares Anthropic’s Fable 5.1 and Fable 5 on a subset of Terminal-Bench-Science tasks using a user-like budget rather than the benchmark’s published conditions. In these runs, Fable 5.1 performed better on some tasks, but both models still failed several tests and the gap was much smaller than the headline benchmark score suggests.

• 7Hacker News·9/6/2026research
  GPT-6 Astra Outperforms Fable on Robot Arm Task

  A comparative benchmark on YAM robot arms found OpenAI's GPT-6 Astra completed a block-to-bowl manipulation task far more often than Claude Fable 5 and 5.1, while also using less time and lower estimated cost per run. On a more difficult puzzle-piece insertion task, Astra matched Fable 5.1's low success rate and stalled at the same final stage, suggesting stronger performance on one manipulation task but not a broader breakthrough in dexterous control.

• 7Hacker News·9/5/2026framework_update
  Artificial Analysis updates Intelligence Index v4.2

  Artificial Analysis has released Intelligence Index v4.2, adding more complex benchmark tasks, private held-out test sets, and grading improvements to reduce gaming. The update introduces AA-Briefcase and GDP.pdf, while also reweighting the index and highlighting new leaderboard results led by Anthropic and OpenAI.

• 2Simon Willison·9/4/2026product_launch
  Simon Willison releases August sponsor newsletter

  Simon Willison says the August edition of his sponsor-only monthly newsletter is now available to sponsors. The issue covers recent AI and model updates, including OpenAI cyberattack details, Claude auto mode, ChatGPT Work, and model releases, along with assorted notes from his projects and current setup.

• 8OpenAI Blog·9/3/2026model_release
  OpenAI rolls out GPT-6 Astra

  OpenAI is rolling out GPT-6 Astra to a limited set of organizations, with broader availability coming soon to ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS. The model is priced at $10 per million input tokens and $50 per million output tokens, and is reported to score strongly on several benchmarks, especially long-context and security tasks.

• 5The New Stack·9/3/2026research
  PhiloLabs built a virtual Union Square with AI agents for $33

  PhiloLabs used Claude Fable 5.1 agents to build a browser-based 3D reconstruction of San Francisco’s Union Square from geographic data and reference images. The experiment paired Playwright screenshot comparisons with agent review loops to catch visual issues that conventional tests would miss, and the full run reportedly cost about $33 in API usage.

• 8The New Stack·9/1/2026model_release
  Anthropic launches Fable 5.1 with lower cache costs and looser safeguards

  Anthropic released Fable 5.1 and Mythos 5.1, claiming better performance at lower cost mainly through a 75% cut in cache read pricing. Fable 5.1 is generally available for API and premium plan users, while Mythos 5.1 stays in trusted access with safeguards aimed at cybersecurity and life sciences. Anthropic also says Fable 5.1 is less over-cautious, with cyber and biology safeguards intervening less often and new benchmark gains in research-related tasks.

• 2Hacker News·9/1/2026research
  GPU World Story Contest

  Hacker News is promoting a story contest that asks writers to imagine a future where frontier AI is widely available because compute becomes abundant by 2040. The premise assumes AI progress stops improving dramatically after 1 September 2026, but cheaper and faster GPUs continue scaling, enabling universal access to frontier LLMs. The contest includes cash prizes, a judging panel, and a deadline of October 31, 2026.

• 6Simon Willison·8/23/2026research
  Anthropic’s top model faces weak adoption despite revenue growth

  A Financial Times report says Anthropic’s annualized revenue rose sharply to about $65 billion in July, with the company expecting Q3 profitability and 6,000 customers spending at least $100,000 a year. The same report suggests its most expensive models are seeing limited adoption, while cheaper alternatives are gaining traction based on Ramp billing data.