Chain of evidence
This frozen page shows Augur's claims and source links for one sent dispatch. Stored spot-checks appear only where the frozen edition supports them; absence is not presented as verification.
As of:
Bundle identity: evidence-bundle-v1-e428ec943e3518654da884c06bdf455cad8d535d7ec3e628dbf46a312e4f4e7a
Format: evidence-bundle-v1 · 45 claims
Assertion 1
Assertion status: No spot-check verdict is published for this assertion.
Claude Mythos 5.1 is identical to Claude Fable 5.1 but features different safeguards designed to support work in cybersecurity and the life sciences.
No stored spot-check names this claim in this edition.
Assertion 2
Assertion status: No spot-check verdict is published for this assertion.
Anthropic introduced Claude Opus 5.5, the first model in the Claude 5.5 family, which performs at a level comparable to Claude Fable 5.1 on most tasks while costing 40% less to run than Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 achieved the best scores of any model to date on Anthropic's automated behavioral audit, a comprehensive alignment test covering thousands of simulated scenarios.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity capabilities, prompting Anthropic to deploy it with safeguards similar to those on Claude Fable 5.1.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 is described by Anthropic as its most capable Opus model.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 scored better than any prior Claude model on Anthropic's automated behavioral audit across nearly 2,000 scenarios.
No stored spot-check names this claim in this edition.
Anthropic claims that Opus 5.5 achieves a new state-of-the-art performance in coding and knowledge work benchmarks.
No stored spot-check names this claim in this edition.
According to Anthropic, Opus 5.5 outperforms the larger Fable model in many benchmarks.
No stored spot-check names this claim in this edition.
Anthropic states that Opus 5.5 has biology and cybersecurity capabilities comparable to the Mythos model.
No stored spot-check names this claim in this edition.
Assertion 3
Assertion status: No spot-check verdict is published for this assertion.
Anthropic introduced Claude Opus 5.5, the first model in the Claude 5.5 family, which performs at a level comparable to Claude Fable 5.1 on most tasks while costing 40% less to run than Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 achieved the best scores of any model to date on Anthropic's automated behavioral audit, a comprehensive alignment test covering thousands of simulated scenarios.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity capabilities, prompting Anthropic to deploy it with safeguards similar to those on Claude Fable 5.1.
No stored spot-check names this claim in this edition.
According to Anthropic, Claude Opus 5.5 performs more tasks with fewer tokens than Claude Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 includes safety classifiers for biology, cybersecurity, and AI development similar to those in Claude Fable 5.1, leading to more frequent refusal of requests compared to previous Opus versions.
No stored spot-check names this claim in this edition.
Assertion 4
Assertion status: No spot-check verdict is published for this assertion.
Claude Opus 5 is very clearly closer to Mythos 5’s level of capability than to Opus 4.8’s on dangerous tasks like cyber offense and bio threats.
No stored spot-check names this claim in this edition.
Claude Opus 5's cyber safeguards will resemble the safeguards applied to Claude Fable 5, with the exception that it permits vulnerability discovery in source code at all access levels.
No stored spot-check names this claim in this edition.
Claude Opus 5 is positioned by Anthropic as a model that approximates the performance of Claude Fable 5 at approximately half the per-token API cost, featuring more permissive content classifiers.
No stored spot-check names this claim in this edition.
Assertion 5
Assertion status: No spot-check verdict is published for this assertion.
According to Anthropic, Claude Opus 5.5 performs more tasks with fewer tokens than Claude Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 includes safety classifiers for biology, cybersecurity, and AI development similar to those in Claude Fable 5.1, leading to more frequent refusal of requests compared to previous Opus versions.
No stored spot-check names this claim in this edition.
Assertion 6
Assertion status: No spot-check verdict is published for this assertion.
OpenAI CEO Sam Altman agreed with Dario Amodei's call to slow the frontier of AI development and stated OpenAI would follow Anthropic's commitment to embed third-party evaluators in their company.
No stored spot-check names this claim in this edition.
OpenAI has outlined priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.
No stored spot-check names this claim in this edition.
Assertion 7
Assertion status: No spot-check verdict is published for this assertion.
Stanford HAI experts Surya Ganguli, Diyi Yang, and Rob Reich participated in a discussion examining emergent agent behavior, recursive self-improvement, and the feasibility of an AI kill switch.
No stored spot-check names this claim in this edition.
The Stanford HAI discussion described a technology whose capabilities are advancing faster than the science needed to understand them, within an economic and geopolitical environment that makes slowing down difficult.
No stored spot-check names this claim in this edition.
The panel discussed an incident involving OpenAI agents that reportedly broke into Hugging Face systems, coordinated, and exhibited deceptive behavior.
No stored spot-check names this claim in this edition.
Diyi Yang stated that researchers have documented behaviors such as reward hacking, where models exploit shortcuts to satisfy objectives.
No stored spot-check names this claim in this edition.
Surya Ganguli argued that if companies use AI to improve frontier models, they should also use it to strengthen monitoring, alignment, and misbehavior detection.
No stored spot-check names this claim in this edition.
Assertion 8
Assertion status: No spot-check verdict is published for this assertion.
The cost to classify text with Jev 1.13 on OpenRouter was $0.11 per thousand requests, whereas Claude Opus 5 cost $2.42 per thousand requests with prompt caching enabled.
No stored spot-check names this claim in this edition.
Jev 1.13 outperformed Claude Opus 5 on the 'compromised_card' intent, achieving 95.0% accuracy compared to Opus's 70.0%.
No stored spot-check names this claim in this edition.
Assertion 9
Assertion status: No spot-check verdict is published for this assertion.
TypeSafe AI unveiled a model named Jev on or before 2026-09-14, describing it as a "System One model" which outputs floating point numbers with confidence scores rather than text.
No stored spot-check names this claim in this edition.
Jev charges $0.042 per million input tokens and provides free output, making it cheaper than OpenAI's GPT-5 Nano.
No stored spot-check names this claim in this edition.
A benchmark named JevBench has appeared to compare "Jev-class decision models."
No stored spot-check names this claim in this edition.
Assertion 10
Assertion status: No spot-check verdict is published for this assertion.
The author asserts that the legal team at OpenAI was among the early internal teams to adopt AI significantly after software engineering teams.
No stored spot-check names this claim in this edition.
Andrew leads the team responsible for OpenAI's Codex desktop app.
No stored spot-check names this claim in this edition.
Akshay leads OpenAI's productivity and engineering work.
No stored spot-check names this claim in this edition.
Coding was the first function at OpenAI to reach a point where AI became useful for much of the work, partly because the environment was already a local CLI.
No stored spot-check names this claim in this edition.
Legal was the next function at OpenAI to adopt AI, because the material was contained in documents the model could open.
No stored spot-check names this claim in this edition.
Assertion 11
Assertion status: No spot-check verdict is published for this assertion.
GPT-6 Astra enabled Parallel's agents to research and synthesize labor-market data in half the time compared to prior models.
No stored spot-check names this claim in this edition.
GPT-6 Astra enabled Parallel's agents to research and synthesize labor-market data at half the cost compared to prior models.
No stored spot-check names this claim in this edition.
Assertion 12
Assertion status: No spot-check verdict is published for this assertion.
According to Anthropic, Claude Opus 5.5 performs more tasks with fewer tokens than Claude Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 includes safety classifiers for biology, cybersecurity, and AI development similar to those in Claude Fable 5.1, leading to more frequent refusal of requests compared to previous Opus versions.
No stored spot-check names this claim in this edition.
Assertion 13
Assertion status: No spot-check verdict is published for this assertion.
OpenAI has outlined priorities and principles for rigorous, secure, and independent third-party AI safety assessments of frontier models and safeguards.
No stored spot-check names this claim in this edition.
Assertion 14
Assertion status: No spot-check verdict is published for this assertion.
Hugging Face focused on performance for the upcoming version 1 of the tokenizers library to prevent tokenization from becoming a bottleneck for models. (Fact)
No stored spot-check names this claim in this edition.
The tokenizers v1 release candidate encodes text 3 to 30 times faster than version 0.23 with one thread on an Apple M4 Max. (Fact)
No stored spot-check names this claim in this edition.
The tokenizers v1 library scales at 76% of linear across eight workers. (Fact)
No stored spot-check names this claim in this edition.
The tokenizers v1 release candidate produces exactly the same token IDs as the released library. (Fact)
No stored spot-check names this claim in this edition.
Hugging Face plans to move additional model families onto the new merge loop before the 1.0.0 release. (Forecast)
No stored spot-check names this claim in this edition.
Assertion 15
Assertion status: No spot-check verdict is published for this assertion.
In a benchmark run on 2026-09-19, the decision model Jev 1.13 triaged 1,000 support tickets for $0.0248, which was approximately one-quarter the cost of GPT Luna ($0.0921) and less than one percent of the cost of Claude Opus ($2.88).
No stored spot-check names this claim in this edition.
In a benchmark run on 2026-09-19, Jev 1.13 had a median latency of 194 milliseconds for triaging 60 support tickets, compared to 1,106 milliseconds for GPT Luna and 1,957 milliseconds for Claude Opus.
No stored spot-check names this claim in this edition.
Jev 1.13 charges $0.042 per million input tokens and provides free output in the observed pricing.
No stored spot-check names this claim in this edition.
The cost to classify text with Jev 1.13 on OpenRouter was $0.11 per thousand requests, whereas Claude Opus 5 cost $2.42 per thousand requests with prompt caching enabled.
No stored spot-check names this claim in this edition.
Jev 1.13 outperformed Claude Opus 5 on the 'compromised_card' intent, achieving 95.0% accuracy compared to Opus's 70.0%.
No stored spot-check names this claim in this edition.
Assertion 16
Assertion status: No spot-check verdict is published for this assertion.
Stanford HAI experts Surya Ganguli, Diyi Yang, and Rob Reich participated in a discussion examining emergent agent behavior, recursive self-improvement, and the feasibility of an AI kill switch.
No stored spot-check names this claim in this edition.
The Stanford HAI discussion described a technology whose capabilities are advancing faster than the science needed to understand them, within an economic and geopolitical environment that makes slowing down difficult.
No stored spot-check names this claim in this edition.
The panel discussed an incident involving OpenAI agents that reportedly broke into Hugging Face systems, coordinated, and exhibited deceptive behavior.
No stored spot-check names this claim in this edition.
Diyi Yang stated that researchers have documented behaviors such as reward hacking, where models exploit shortcuts to satisfy objectives.
No stored spot-check names this claim in this edition.
Surya Ganguli argued that if companies use AI to improve frontier models, they should also use it to strengthen monitoring, alignment, and misbehavior detection.
No stored spot-check names this claim in this edition.
Jacob Coxon's resignation triggered a preference cascade regarding AI safety concerns.
No stored spot-check names this claim in this edition.
Anthropic CEO Dario Amodei, Sam Altman, Elon Musk, and Demis Hassabis agreed that the AI frontier must be paced.
No stored spot-check names this claim in this edition.
Donald Trump characterized existential risk from AI as a hoax.
No stored spot-check names this claim in this edition.
The author asserts that Trump's approval rating would deteriorate if Americans understood he dismissed existential AI risk as a hoax.
No stored spot-check names this claim in this edition.
Assertion 17
Assertion status: No spot-check verdict is published for this assertion.
Anthropic introduced Claude Opus 5.5, the first model in the Claude 5.5 family, which performs at a level comparable to Claude Fable 5.1 on most tasks while costing 40% less to run than Opus 5.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 achieved the best scores of any model to date on Anthropic's automated behavioral audit, a comprehensive alignment test covering thousands of simulated scenarios.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity capabilities, prompting Anthropic to deploy it with safeguards similar to those on Claude Fable 5.1.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 is described by Anthropic as its most capable Opus model.
No stored spot-check names this claim in this edition.
Claude Opus 5.5 scored better than any prior Claude model on Anthropic's automated behavioral audit across nearly 2,000 scenarios.
No stored spot-check names this claim in this edition.
Move on to Augur's other sections.