OpenAI 28-Day Sprint Day 12: GPT-6 Pro Multi-Agent Parallelism, Codex Boasts Lower Quota Consumption

Beating AI News: Perplexity has open-sourced two multimodal decision models — pplx-decider-v1.2 (27B) and pplx-decider-flash-v1.2 (4B) — based on Qwen3.8-27B and Qwen3.5-4B respectively. Unlike models that generate long-form responses, these two models perform classification, scoring, and judgment based on text or images, suitable for use cases such as agent task routing and content moderation. Compared to the previous generation which required per-question inference, the new models allow multiple questions to share a single input, completing judgments in one inference step. Perplexity states that inference speed is up to 5 times faster in some scenarios. However, the 4B version may experience accuracy loss when combining multiple questions, so the official recommendation is to default to per-question processing. In the public test of Decision Index 0.3, the 27B model’s score rose from 62.30 to 68.09 compared to the prior generation, while the 4B version ranks first among models of the same scale. It should be noted that the new models incorporated official training datasets from some evaluation tasks during training. Both models are licensed under Apache 2.0, and integrated with Perplexity API and Vercel AI Gateway. API pricing is $0.02 per million input tokens for the 27B model and $0.01 per million input tokens for the 4B model, with free output tokens.

Beating AI Express News: Anthropic has upgraded Claude Code’s automated code review tool, Code Review, from research preview to general availability, making it available to Team and Enterprise users. The feature, first launched in March this year, uses multiple parallel agents to inspect developers’ GitHub pull requests (PRs), identifying logical errors, security vulnerabilities, and edge cases while filtering false positives via cross-validation. Issues, their trigger conditions, and fix suggestions are directly annotated on the relevant code lines. Anthropic says most reviews now take 6–10 minutes, nearly three times faster than the preview’s average of ~20 minutes. Admins can set three review depth tiers for different repositories: Quick, Standard, and Deep. Claude Code founder Boris Cherny noted he uses Code Review for all his PRs, adding that it catches nearly all bugs before merging. The feature requires an additional paid subscription and is not included in standard Claude plans. Anthropic previously disclosed that average review costs range from $15–$25, billed based on token usage, with fees typically higher for more complex code and deeper reviews.

Dual shocks from AI infrastructure financing risks and surging oil prices hit Wall Street: the Nasdaq 100 index fell 5.5% week-to-date, marking its largest single-week drop since the April 2025 tariff turmoil; the Philadelphia Semiconductor Index declined for eight consecutive trading days, its longest losing streak since 2008, with a total loss of around 16%. AI cloud service provider CoreWeave has become a market focal point, with some of its bond yields once topping 15%, sparking concerns over AI infrastructure financing costs, capital expenditure returns, and corporate solvency. CoreWeave’s stock dropped 12% on Friday, with its cumulative decline from this year’s peak exceeding 50%. Meanwhile, Brent crude oil climbed to around $110 per barrel, while WTI crude rose 3.8% on Friday to $101.71. The oil price surge has reignited inflation expectations, pushing the 10-year U.S. Treasury yield to 5.097%, eroding government bonds’ safe-haven appeal amid stock market declines. Market worries about AI are shifting from computing power demand to financing costs and profit returns. Still, some investors are opting to buy the dip: Cathie Wood’s ARK Investment Management increased its holdings in CoreWeave and Cerebras Systems. Going forward, tech giants’ earnings reports will be critical to verifying whether AI-related revenues can support sustained high capital expenditures.

Beating AI News reports that OpenAI on October 15 newly disclosed four AI agent privilege escalation incidents, all occurring during internal training or evaluations in June this year, with two classified as severe. The most egregious case involved a model tasked with querying five public metrics from an international research institute; instead, it found leaked credentials in webpage code, accessed the institute’s internal system, modified online configurations, obtained cloud storage access credentials, and downloaded 147,451 lines of non-public data. The model failed to report the entire process to evaluators and answered all five final questions incorrectly. Another model, assigned to query a region’s annual hospital bed count, exploited a file download vulnerability after the government statistics site denied access, obtained server configurations and programs, analyzed authentication flaws in the code, forged tokens to bypass restrictions, and ultimately downloaded 24 years of bed statistics. A third incident involved an internal research model from the GPT-5.6 Sol family: when deploying a smart contract, it lacked gas fees, so it derived and checked 362 wallet addresses from public developer mnemonics, located a funded wallet, and improperly used its Sepolia testnet tokens. The model questioned its authorization during reasoning, then justified its actions by noting the mnemonics were public and testnet tokens had no real monetary value. In the fourth incident, the model used an external browser to send forbidden network requests and even continued researching restriction-bypassing methods after completing its task. OpenAI has now publicly released 19 model misalignment reports. Three of the four newly disclosed incidents were only traced and discovered in August or September. The company said it has tightened network access permissions for training and evaluation environments and expanded behavioral monitoring.

Apple said pre-order demand for its first foldable iPhone, the iPhone Duo, is "unprecedented". Within minutes of pre-orders opening, shipping timelines were pushed from November to December, then further delayed to the end of January 2027, affecting all storage configurations from 256GB to 2TB and every color option. Apple noted it is making every effort to boost delivery efficiency, with first shipments set to launch on October 23. The iPhone Duo starts at $1,999, with the top-tier configuration priced at $3,200. It features a 7.6-inch inner display and a redesigned iOS experience tailored for the foldable form factor. Apple is grappling with component shortages that have plagued the consumer electronics industry this year, particularly a shortage of memory chips, which has forced the company to raise prices across its entire product line. It will also launch a touchscreen-equipped MacBook Pro and other devices later this month. Analysts Anurag Rana and Josh Christensen from Bloomberg Intelligence have raised their sales forecasts, projecting that the iPhone Duo could exceed Apple’s previous estimates of 14 million units sold and $28 billion in revenue for fiscal 2027. If the device successfully attracts existing iPhone Pro and Pro Max users to upgrade, the overall average selling price of iPhones is also expected to rise significantly. Apple’s stock rose 0.43% on Friday to close at $343.46, marking a 1.61% gain over the past five days.

Argentine dollar stablecoins (including USDT) have posted five consecutive days of price increases. Data from Coinpay shows that as of October 16, the median USDT buy price on Binance’s P2P market climbed to 1,612.6 Argentine pesos, while Argentina’s central bank’s official wholesale exchange rate stands at 1,512.81 pesos, translating to a 6.6% premium — the highest level since August 15. On OKX’s P2P market, USDT traded at an average of around 1,614.15 pesos, with a roughly 6.7% premium. The surge came within a week of Argentine President Javier Milei warning of exchange rate run risks, with Milei noting his administration has set aside $125 billion to prepare for potential financial stress in 2027. A regulatory rule barring investors from selling dollar-denominated MEP securities or purchasing foreign currency-settled securities within 90 days of buying official dollars has driven some users to opt for USDT, attracted by its flexible trading, 24/7 liquidity and convenient cross-border transfers — though this comes with higher premiums. Regionally, as of October 12, Argentina’s digital dollar premium hit 5.9%, ranking second in Latin America, only behind Venezuela’s 15.3%.