Codex lead Tibo shares tutorial for enabling 1M context in Codex, sparking pricing controversy
要闻
Codex lead Tibo shared configuration instructions for enabling a 1-million-token context window for GPT-5.6 Sol. Users need to modify the config.toml file and restart the application. Tibo noted that the default settings have been optimized for performance and cost. However, another OpenAI employee said tokens beyond the default window count at twice the rate toward usage limits—a claim Tibo had previously denied.
Codex lead Tibo shared how to enable a 1 million-token context window for GPT-5.6 Sol in Codex: add the following configuration at the top level of ~/.codex/config.toml:
Save the file, restart the Codex client, and start a new session for the change to take effect. Tibo said Codex’s default context limit has been optimized for performance and cost, and that he documented this option in response to frequent requests. However, another OpenAI member, Vaibhav (VB) Srivastav, warned that tokens beyond the default window count at 2 times the rate toward usage limits and that the setting should be disabled when not needed, though Tibo had previously denied this claim.
Codex lead Tibo confirms Astra integration is coming to Codex
要闻
OpenAI Codex lead Tibo confirmed in a post on X that the next-generation Astra model will be integrated into Codex in the future, but did not disclose a launch date or further details.
OpenAI Codex lead Tibo posted a list of Codex features: nearly 100% reliable, occasional resets, and open source. He added “will have Astra” in parentheses at the end, indicating that Codex will integrate Astra in the future. OpenAI has not officially announced a launch date, and the original post contained no timing information.
DeepSeek API officially raises prices and introduces peak and off-peak pricing
开发生态
DeepSeek's previously announced new API pricing took effect at midnight today. The new structure introduces peak and off-peak pricing, with substantial price increases across the board.
The new API pricing structure previously announced by DeepSeek officially took effect at midnight today, and the prices in its official API documentation have been updated. Pricing for deepseek-v4-flash and deepseek-v4-pro will shift from a flat rate to peak and off-peak pricing, with an overall increase. Peak hours are 9:00—12:00 and 14:00—18:00 each day, with all other times designated as off-peak. Off-peak prices are half the peak-hour rates.
OpenCode adjusts Go quotas following DeepSeek's official price increase
开发生态
OpenCode announced that it has updated OpenCode Go quotas following DeepSeek's official price increase. OpenCode founder dax said the team is testing multiple deployment options this week in an effort to continue hosting DeepSeek at prices close to previous levels. Community users observed that as DeepSeek model prices rose, the v4 flash model allowance fell from $60 to $15, while estimated monthly requests dropped from 158,150 to 18,900.
OpenCode team member dax said DeepSeek has officially raised its prices, and OpenCode Go allowances have been updated accordingly. The team is testing multiple deployment options this week in an effort to continue hosting DeepSeek at prices close to the previous levels. dax noted that dozens of providers claiming to match the old pricing have in fact failed to do so, and that no provider had ever truly matched DeepSeek’s pricing before. Community users observed that DeepSeek models on OpenCode Go had become more expensive while allowances had also shrunk: the allowance for the flash model was cut from USD 60 to USD 15, peak-period surcharges were introduced, and the page’s estimated monthly request count fell from 158150 to 18900.
SuperGrok Heavy subscribers can use Cursor Ultra for free while their subscription remains active
开发生态
SuperGrok Heavy subscribers can use Cursor Ultra for free by linking their Grok account, with access continuing as long as their subscription remains active. Cursor clarified on its community forum that the “one month” period previously listed on its help page was outdated and that access now lasts for the duration of the active subscription.
Cursor is offering Cursor Ultra free of charge to SuperGrok Heavy subscribers. Under the official terms, the subscription can continue to be used for free as long as the SuperGrok Heavy plan remains active at renewal. The help page previously stated that the offer included 1 free month, but a staff member on Cursor’s official forum said this was part of the original launch terms and is now outdated. To claim the offer, users must link a Grok account with an active SuperGrok Heavy subscription through Grok Bot. Each Grok account can be linked to only one Cursor account, and the offer does not apply to existing Ultra subscribers, team accounts, or subscribers to other Grok plans.
Dario Amodei responds to criticism over AI risk communication and regulation
技术与洞察
Anthropic CEO Dario Amodei published a response to criticism over AI risk communication and regulation, rejecting claims that his remarks have been overly negative. While acknowledging negative public sentiment toward AI, he argued that it stems from a long-standing crisis of trust. He also announced that Anthropic is accelerating its biology and medical research and expects to publish early results in the coming months as part of its commitment to improving the world.
Anthropic CEO Dario Amodei responded on X to recent criticism of him, denying that his personal communications about AI have been “disproportionately negative” and saying his remarks have been broadly balanced between risks and benefits. He acknowledged that the public holds negative views of AI, but argued that these stem from a long-standing crisis of trust rather than warnings from him or other AI leaders. He said the “most accurate criticism” of AI companies is that they have yet to deliver on their major promises to improve the world. To address this, he said Anthropic is accelerating its investment in biology and medicine, with early results expected in the coming months. On regulation, he rejected the binary choice between “centralized regulation or broad distribution,” arguing that well-designed rules can both constrain the power of frontier AI companies and leave room for open-weight models. He also voiced support for the U.S. government’s reported approach of pre-deployment testing.
Stripe reportedly finalizes deal to acquire OpenRouter for over $7 billion
前瞻与传闻
Payment company Stripe has reportedly finalized an agreement to acquire OpenRouter for more than $7 billion. The deal has not yet been officially confirmed.
Bloomberg, citing people familiar with the matter, reported that payment processor Stripe has finalized an agreement to acquire startup OpenRouter for more than $7 billion. The report has not yet been officially confirmed. Just a few months ago, OpenRouter completed a funding round at a reported valuation of $1.3 billion.
OpenAI has reportedly disbanded the Preparedness team responsible for assessing catastrophic risks from its models, splitting its work among existing teams. Co-founder Greg Brockman said safety work has been integrated more closely into model development.
According to the Financial Times, citing OpenAI insiders, OpenAI disbanded its Preparedness team at the end of July. The team was responsible for assessing whether the company’s own AI models could pose severe or catastrophic risks, and its work on biological and cyber risks has been split up and absorbed into existing teams. Several safety staff members have recently left the company. OpenAI co-founder Greg Brockman said the company has integrated safety work more closely into model development. Dylan Scandinaro, the former head of the team, is now focused on safety risks arising from “recursive self-improvement” AI—systems capable of optimizing themselves and training other models.
Note: This content was created with AI assistance and may contain hallucinations and errors.
