US AI accelerates cost reduction | Rewire Morning News

AI vendors are fighting over every call, while crypto companies recalculate revenue at every layer of infrastructure. Technology can keep expanding, but who bears the cost and who takes the cash flow is becoming a harder dividing line.

1 | Small models cut prices, and the chat box begins to take over the tool entrance

On October 7, Anthropic released Haiku 5.5. For requests of no more than 100,000 tokens, the price per million input and output tokens is $0.10 and $0.50, respectively, 90% lower than Haiku 4.5. The company estimates that, considering changes in task consumption, average operating costs fall by about 75%.

On the same day, OpenAI began rolling out GPT-6 with an interactive interface, first for paid tiers, with free tiers expanding the next day.

The two companies are competing for different entrances. One lowers the cost of summaries, classification, and subtasks so agents can call them more often. The other puts charts, forms, and widgets directly into answers, aiming to let users complete tasks within the conversation.

From this, application vendors face dual pressure. Basic intelligence is cheaper, but lightweight tools such as calculators and information comparison are also more easily absorbed by the chat entrance. Opportunity shifts toward proprietary data and real business processes.

(Source: Anthropic⁠ / OpenAI⁠)

2 | Microsoft and Nvidia pull agents' execution rights back to the computer

On October 7, Microsoft opened preorders for the Surface Laptop Ultra, equipped with Nvidia RTX Spark and up to 128GB of unified memory. Nvidia said the related laptops go on sale on October 16, with compact desktop machines going on sale in November. At the same time, Microsoft announced the general availability of execution containers, allowing agents to run continuously under operating system control.

Hardware and permission management were introduced together, showing that competition in local AI has moved from whether a model can fit to who manages agent execution. Microsoft wants enterprises to continue allocating permissions and observing behavior through Windows, while Nvidia extends CUDA to personal computers. Connected to the price cuts of small cloud models, local devices must prove their value in long-term operation, data retention, and response speed. Inference costs have not disappeared; they have partly shifted from pay-per-use to device procurement, maintenance, and enterprise management expenses.

(Source: Microsoft⁠ / Nvidia⁠)

3 | Abstract shuts down, and brand traffic did not sustain a chain

According to a CoinDesk report on October 7, Pudgy Penguins parent company Igloo will shut down Abstract on December 15 and is asking users to migrate their assets in advance. The company has already invested tens of millions of dollars, is unwilling to keep pulling money from the Pudgy Penguins business to support the network, and has decided not to continue raising funds through a token issuance. The report, citing network disclosures, said cumulative transactions exceeded 325 million, with on-chain corporate revenue surpassing $40 million.

The problem lies in revenue attribution. An app making money does not mean the underlying chain can recover its operating costs. The report, citing DefiLlama data at the time, said chain fees over the past 24 hours were about $3,900, while app revenue was about $39,000. A brand can bring wallets and transactions, but it does not necessarily bring enough chain revenue. After Blast announced its shutdown, Abstract once again reminds consumer-facing projects that owning a chain and owning a sustainable business need to be proven separately.

(Source: CoinDesk)

4 | Stablecoin companies compete for licenses, public chains compete for settlement standards

CoinDesk reported on October 7 that stablecoin payments company Rain applied for a U.S. national trust bank charter, planning to conduct digital asset custody, reserve management, and stablecoin issuance and redemption. The application is still pending approval; the bank will not take deposits, make commercial loans, or provide federal deposit insurance. On October 6, the Solana Foundation released an open-source delivery-versus-payment program that allows assets and payment to settle simultaneously in a single atomic transaction.

The two pieces of news point to different layers of financial infrastructure. Rain hopes to reduce reliance on third-party banks for custody and redemption, while Solana hopes to turn bespoke settlement contracts into reusable standards. In contrast to the shutdown of a consumer chain, institutional business is competing for the role of continuously processing assets. Speed is only one condition; asset ownership, compliance, and privacy still need to be addressed. JPMorgan only provided settlement experience input to Solana, and the announcement explicitly stated it does not constitute an endorsement.

(Source: CoinDesk / Solana Foundation)

5 | Oil market accelerates reserve release, crypto market still constrained by interest rates

On October 7, IEA members supported accelerating the release of oil reserves pledged in March and prioritizing diesel supply as much as possible. The IEA said about 325 million barrels had already been released, with about 100 million barrels of commitments still outstanding. This is an acceleration of old commitments and cannot be treated as new supply.

CoinDesk reported the same day that Bitcoin fell below $84,000, with rising oil prices accompanied by higher U.S. Treasury yields and a stronger dollar. It cited about $547 million in 24-hour crypto liquidations at the time.

Releasing reserves can ease immediate supply, but it cannot directly fix shipping risks, nor can it guarantee that inflation expectations will fall. For crypto assets, progress in institutional settlement and stablecoins may not offset the pressure of a tightening rate environment. From this set of signals, industry adoption and the price of capital are two different lines. On-chain infrastructure continues to expand, while coin holders still have to endure volatility transmitted jointly by energy, the bond market, and leveraged positions.

(Source: IEA / CoinDesk)

Also worth knowing ↓

Anthropic divides cybersecurity capabilities into three tiers of access. The plan updated on October 6 is aimed respectively at defense, authorized red teams, and testing for certain high-risk systems, with different tiers corresponding to different review and control requirements. The commercialization boundary of model capabilities is beginning to be determined by customer identity and usage permissions. (Source: Anthropic)

OpenAI published a batch of mathematical results generated by internal models. The repository released on October 6 includes some Lean formal proofs, reasoning summaries, and compute estimates. Machine-checkable proofs help with verification, but not all published results can be directly equated with findings already confirmed by the mathematics community. (Source: OpenAI)

Atlassian expands its enterprise agent cooperation with OpenAI. An October 6 announcement said the model will be combined with the Teamwork Graph that connects people, projects, and documents, and will enter Rovo. After models become cheaper, enterprise software can still retain value through business context and work entry points. (Source: OpenAI)

ChatGPT's teen protections were questioned by an independent evaluation, and OpenAI rebutted the testing method. Common Sense Media said on October 7 that some parental alerts did not work as promised. OpenAI pointed out that after account linking, several hours are needed for activation, while the evaluator stood by its conclusion. The dispute centers on whether protection features can take effect in time in real-world use. (Source: Common Sense Media / KQED)

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia