Fireworks says it found impacts to account and billing information, plus secrets, datasets and incomplete model weights for a limited set of accounts. Its October 9 bulletin had identified exposed evaluation-job secrets.
The company says a third-party cloud provider was compromised on October 23. On October 24, an intruder reached Fireworks development machines and ran a workload that caused additional customer-data impact.
Billing information includes addresses and invoice history. Fireworks says it does not store payment card numbers and those numbers were not affected.
Self-service dataset uploads and downloads for batch inference and dataset-based training are disabled during the investigation. The company says it is notifying affected customers with account-specific actions.
Fireworks reports no detected impact to its default zero-data-retention inference, Training SDK or Nexus services. It says it has not detected further successful unauthorized access to its customer platform since October 25.
Caveat: The investigation is ongoing, so absence of detected impact is not proof that every service or customer is unaffected.
(Techmeme · collected Forbes report · Fireworks incident update)
Both use the same underlying model. Fable has additional cybersecurity and biology safeguards; direct Mythos access requires verification, although Mythos also powers Claude Security for Enterprise customers.
Standard pricing is $8 per million input tokens and $40 per million output tokens, units of information the model processes and generates. Optional fast mode doubles both rates and Anthropic claims up to 2.5 times faster output.
Anthropic reports 49.1% for Fable 5.2 at extra-high effort on FrontierCode 2.0, a fixed set of repository tasks, versus 32.4% for Fable 5.1 and 45.5% for Opus 5.5 in its comparison.
Artificial Analysis reports a 65 on its composite test at maximum effort, alongside an average $26.54 and 45 minutes per test task. These are test-specific costs and durations, not a normal request quote.
Anthropic’s system card says Fable’s fallback to an older model increased attack success in prompt-injection tests, where malicious text tries to redirect an assistant.
Caveat: The coding comparison is vendor-reported, and the independent composite score does not establish performance or cost on your own tasks.
(HN 696 points · 378 comments · Techmeme · announcement, system card and independent test)
Meta lists muse-ember-1.0 on its Model API and Muse Code at $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs $0.004 per million tokens.
The advertised context window holds up to one million tokens, with up to 512,000 output tokens. Meta says the model accepts text, images, audio and video.
On Artificial Analysis’s 132-task FrontierCode 2.0 test, Meta reports Ember at 27.2% and Haiku 5.5 at 34.9%. Its methods report distinguishes independent results from Meta’s own tests.
For Terminal-Bench 4.0, Meta reports 40.9% for Ember and 39.2% for Haiku. The published comparison uses each model’s best reported surrounding tool software, so the setups differ.
Caveat: These are selected test results and hosted prices; they do not guarantee savings or quality on a particular workload.
(HN 167 points · 70 comments · Techmeme · release and evaluation methods)
The November 8 version 3.13.0 adds transcription, speech generation and real-time audio. wllama runs llama.cpp through browser WebAssembly, a way to run compiled programs on the web.
The browser needs two isolation headers for multithreading. The project says model files larger than 2 GB need splitting; smaller chunks can reduce memory pressure.
Caveat: The documented continuously streaming transcription mode needs model support that LFM3-Omni does not have; its committed audio buffers can still be transcribed.
(X @ggerganov · collected announcement · release and version-pinned documentation)
A Gemini connected-app privacy claim → STALE: Google’s documented change is dated October 16; the newly surfaced account does not establish a fresh policy change. Its autonomous retrieval claim was not independently reproduced here.
The Opus degradation discussion → HYPE: the discussion does not supply a controlled comparison checked in this run that separates model changes from settings and workload.