← Back to archive中文
NeuronX AI Daily

OpenAI launches GPT-6.2 Astra and Sol with million-token context windows at two price pointsModel competition focuses on token efficiency; Nvidia doubles quarterly revenue and excludes data center compute revenue from China from its next-quarter guidance.

November 18, 2026 Wednesday Sources · X · Blog · TechCrunch
About this issue: This brief is automatically compiled, grouped and rewritten from public sources (X / podcasts / blogs and newsletters). Every item links to its original source — please defer to the original; AI rewriting may contain errors, and corrections against the source are welcome.

In one paragraph

OpenAI has launched GPT-6.2 Astra and Sol, both with 1-million-token context windows. Astra costs $8 per million input tokens and $40 per million output tokens; Sol costs $2 and $10, respectively. Nvidia reported quarterly revenue of $115.8 billion, up 103% year over year, and announced it would stop selling H200 chips in China. Its $130 billion revenue guidance for next quarter excludes data center compute revenue from China. Google is making Gemini Tasks free for users in six countries, connecting email, calendars and payment services. Sharing payment or private account information requires user permission.

🚀Model efficiency moves to task-level accounting

Token prices, task success rates and actual output volume are becoming cost metrics that need to be assessed together.

XOpenAI launches two GPT-6.2 modelsLaunch

OpenAI has released GPT-6.2 Astra and Sol, emphasizing token efficiency. According to Vals, both support 1-million-token context windows. Astra costs $8 per million input tokens and $40 per million output tokens, while Sol costs $2 and $10, respectively. Sam Altman said the new models are highly capable but acknowledged that the collaboration experience still needs improvement, making it a priority for future releases.

Read original →
BlogMicrosoft releases MAI Code 2 coding modelsLaunch

Microsoft has released MAI Code 2 and a Flash variant, both now integrated into GitHub Copilot. Microsoft says Code 2 averages 98% of GPT-6.1 Sol's performance across four enterprise coding benchmarks at roughly 40% of the price. Neither model has yet been tuned for basic scientific research or very long-horizon agent tasks. Flash is available only in a limited-access preview on Foundry.

Read original →
BlogQwen adapter improves coding task success ratesResearch

A lightweight adapter for Qwen3.8-27B has been published on Hugging Face, using reinforcement learning to adapt the model to the pi coding agent. Reported results show that its success rate on the 51 hardest SWE-bench Pro tasks rose from 39.2% to 64.7%. Without enabling thinking, it matches the base model's score at its highest thinking setting while generating roughly a quarter as many tokens. These results apply only to the pi environment and cannot be compared directly with full leaderboard results obtained using other frameworks.

Read original →

🔍 Analysis: MAI Code 2's 98% performance figure is measured against GPT-6.1 Sol, so it cannot establish how the model compares with the newly released GPT-6.2. The Qwen adapter's token savings are also specific to the pi environment. That means GPT-6.2's two price points cannot simply be combined with those two results to produce a value-for-money ranking. Developers need to measure both success rates and actual token consumption on the same tasks in the same runtime environment.

💰The market boundaries of Nvidia's growth

Data centers account for more than 90% of revenue, but data center compute revenue from China is excluded from next-quarter guidance.

BlogNvidia's quarterly revenue doubles to $115.8 billionEarnings

Nvidia reported revenue of $115.8 billion for the third quarter of fiscal 2027, up 103% year over year. Data center revenue reached $107.6 billion, up 110%, accounting for more than 90% of total revenue. The company said Vera Rubin NVL72 has begun shipping and is running at cloud providers including Microsoft and Google. It forecasts next-quarter revenue of $130 billion, plus or minus 3%, excluding data center compute revenue from China.

Read original →
TechCrunchNvidia to stop selling H200 chips in ChinaPolicy

According to TechCrunch, Nvidia's acting CEO Colette Kress said the company would stop selling H200 chips in China. The previous sales arrangement, brokered by Jensen Huang and Trump, required the chips to be routed through the United States so the U.S. government could collect a 25% tariff. The move comes as the United States introduces new rules banning sales of advanced U.S. chips to overseas data centers that remotely serve Chinese customers.

Read original →

🔍 Analysis: Nvidia's $130 billion revenue guidance for next quarter does not depend on data center compute revenue from China, so its growth outlook should not be read as a bet on continued H200 sales to China. The halt in H200 sales, combined with new rules restricting overseas data centers that remotely serve Chinese customers, also means Chinese customers' access to advanced computing cannot be assessed solely by whether chips enter the country directly.

🛠️AI takes on hands-on software engineering

From retaining project context to fixing tests and finding vulnerabilities, engineering outputs require different forms of validation.

XDevin Bots gain separate memories and persistent identitiesLaunch

Cognition has introduced Devin Bots, with one bot leading a team of agents on engineering tasks ranging from writing firmware to merging code. Swyx explained that teams can create dedicated bots for different projects, each with its own memory, goals and persistent identity, and able to communicate with one another when needed. Unlike ad hoc groups of agents, these bots bring cross-session context by default, retaining project background over time.

Read original →
BlogNvidia releases Nemotron tools for scientific agentsLaunch

Nvidia has launched Nemotron for Science, combining open models, chemistry datasets and post-training recipes to help researchers build specialized scientific agents. Its chemistry dataset contains nearly 50,000 questions for training and evaluating models' chemical reasoning. Nvidia says an agent developed by KTH Royal Institute of Technology in Sweden using Nemotron fixed 81% of test cases in the molecular dynamics software GROMACS during a single overnight run, with all fixes independently verified by the software.

Read original →
BlogCalif uses AI to discover Java4Shell vulnerabilitySecurity

Calif says it used AI to discover Java4Shell, a memory corruption vulnerability in OpenJDK's Base64 decoder that can lead to sensitive information disclosure or remote code execution. Its researcher found the vulnerability in one hour and built the first working exploit in another two hours, at a token cost of less than $200. The flaw is in low-level acceleration code, but there is no universal exploit: the specific attack steps depend on how an application uses Base64.

Read original →

🔍 Analysis: Devin Bots retain project context in cross-session memory, addressing continuity between engineering tasks. The 81% repair rate in the Nemotron example was independently verified by GROMACS, addressing how results are validated. Calif discovered a vulnerability and built the first exploit for less than $200 in token costs, but exploitation depends on how an application uses Base64, so that cost cannot be extrapolated directly to arbitrary applications. For engineering teams, a persistent identity does not guarantee reliable delivery: completion criteria still depend on test validation and application-specific conditions.

🤖Gemini Tasks opens up errands and payments

BlogGemini Tasks becomes free for users in six countriesLaunch

Google announced that Tasks, powered by Gemini 4 Hydrogen, is now free for users in the United States, Japan, South Korea, Canada, Australia and India. It connects email, calendars and payment services, and can make phone calls to book appointments, analyze bills and complete purchases through Google Pay. Advanced features are launching first in the United States. It identifies itself as AI when a call connects, businesses can opt out, and sharing payment or private account information requires user permission.

Read original →

🔑Key terms this issue

KEYWORD 01
Token efficiency
Whether a model saves money depends not just on its price per million tokens, but also on how many tokens it actually generates to complete a task and whether the result meets acceptance criteria.
KEYWORD 02
Persistent identity
Devin Bots retain separate memories and goals for projects, allowing bots to carry project context across sessions.
KEYWORD 03
Compute export boundaries
Restrictions on advanced chips for China cover not only direct sales but also supplies to overseas data centers that remotely serve Chinese customers.
Worth watching (reference points, not predictions or advice)
📺 Channel updates today · 5
NeuronX AI Daily
AI moves fast; you don't have to chase all of it. We read the primary sources and keep the few things that matter.
📮 Subscribe freeRSS