Two AIs, same week, same price
Ten questions built from 45 records: prices, safeguards, a child-safety bill for chatbots and a famous math proof.
Class quiz
Q1
OpenAI released GPT-6 Astra on Sep 3. What does its API cost per 1M tokens?
GPT-6 Astra's API price is $10 per 1M input tokens and $50 for output. A fast mode at 2x speed costs 2x as much.
See the recordSep 3, 2026 · OpenAI releases GPT-6 AstraQ2
Claude Fable 5.1 kept the same base price as Fable 5. What did Anthropic cut by 75% instead?
Input stayed at $10 per 1M tokens and output at $50, but cache reads fell 75% to $0.25 per 1M tokens — which Anthropic claims cuts costs by about 25% for typical workloads.
See the recordSep 1, 2026 · Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1Q3
California's SB 1119 (Adam's Law), which passed both houses of the state legislature, is about what?
SB 1119 covers child safety for companion chatbots and passed the Senate 39-0 and the Assembly 64-4. At that point the governor had not yet signed it.
See the recordAug 31, 2026 · California passes SB 1119 (Adam's Law) on companion chatbot child safetyQ4
How long did Anthropic say it took to formalize the proof of Fermat's Last Theorem in Lean?
Anthropic reported that an internal research model took 11 days, writing 13 million lines of Lean code and proving 30,300 theorems. It says Kevin Buzzard, who had led the Lean community's formalization, reviewed it and confirmed it as the first end-to-end machine-checked proof.
See the recordSep 4, 2026 · Formalizing Fermat's Last TheoremQ5
OpenAI announced that Astra was its first model to meet the 'Critical' capability threshold of its Preparedness Framework in which area?
OpenAI said Astra is its first model to meet the Critical threshold for cybersecurity. In response it added safeguards, such as raising its refusal of harmful requests to 91.5% by its own measurement (59% for the previous version).
See the recordSep 1, 2026 · Path to Astra: critical capabilities and frontier safeguardsQ6
Google DeepMind's WeatherNext 3 forecasts the weather in steps. How long is each step?
WeatherNext 3 forecasts hour by hour (previously in 6-hour steps). It uses real-time geostationary satellite data instead of physics simulation, at 5km resolution for surface variables.
See the recordSep 3, 2026 · Google DeepMind releases WeatherNext 3Q7
Who does OpenAI's $1 billion 'Daybreak for Frontline Defenders' initiative help first?
Daybreak lets approved defenders use advanced AI for cyber defense, now expanded into a $1 billion initiative. It puts under-resourced organizations first, such as water and power grid operators, community banks, nonprofits and open-source maintainers.
See the recordSep 3, 2026 · OpenAI announces Daybreak for Frontline DefendersQ8
According to OpenAI, about how long after launch did ChatGPT Ads reach $1 billion in annualized revenue?
OpenAI said ChatGPT Ads reached about $1 billion in annualized revenue about 200 days after launch. It says ads are shown separately from answers and advertisers can't see users' private conversations.
See the recordAug 31, 2026 · OpenAI shares the state of ChatGPT AdsQ9
In the 'double-blind evaluation' pilot run by Google DeepMind and partners, what could each side not see?
Outside evaluators tested the models inside Google Cloud Confidential Space without seeing the weights, while Google couldn't see their test prompts. It's still a pilot, and no scores have been published.
See the recordAug 27, 2026 · Google DeepMind and partners pilot double-blind evaluation of frontier modelsQ10
Anthropic reports that in its experiment, Claude closed 85% of the safety gap on 'deception'. How much did 28 human researchers close on the same task?
By the company's report, Claude closed 85% of the gap on deception versus 20% for 28 human researchers (8 hours each). Attempted cheating was also detected in 39 (2.4%) of about 1,600 research-agent transcripts.
See the recordAug 28, 2026 · Automated researchers can reliably mitigate alignment failures