According to BlockBeats, on June 28, Coinbase CEO Brian Armstrong disclosed that the company has reduced its AI spending by nearly half despite the continued exponential growth in token usage. This was achieved by switching the default AI model to open-source models such as GLM 5.2 and Kimi 2.7, and by combining this with measures such as smart model routing and cache optimization.

Brian Armstrong stated that Coinbase no longer relies on limiting usage limits or spending alerts to control costs. Instead, it defaults to using lower-cost open-source models, only invoking more powerful cutting-edge models when necessary. Furthermore, the company further reduces inference costs through automated model routing, improved cache hit rates, and simplified context management.

It's worth noting that this statement comes at a time when Anthropic's CEO has recently expressed concerns about open-source models and model distillation on multiple public occasions. Coinbase's experience shows that open-source models are becoming an important option for balancing cost and efficiency in enterprise AI applications.

According to BlockBeats, on June 28, in response to a WSJ report that China's Zhipu AI has caught up with Anthropic in some cybersecurity benchmark tests, market analyst Serenity stated that rather than attributing this result to US government policies, it is more accurate to believe that Anthropic has not established a sufficiently robust model protection mechanism to prevent model distillation.

Serenity stated that prior to Anthropic's release of Fable, rumors circulated in the industry that Chinese teams were capable of distilling its models. He believes that the key to AI competition lies not only in investing heavily in capital expenditures (CapEx) to build technological moats, but also in protecting core model capabilities and avoiding "handing over" achievements through methods such as numerous API calls.

According to a recent study by CoinGecko, as reported by BlockBeats on June 28, Bitcoin has experienced seven bear markets since 2014. CoinGecko defines a bear market as a period in which Bitcoin's closing price remains below its 200-day simple moving average (SMA) for more than 30 consecutive days.

Data shows that as of June 28, 2026, the current 2025-2026 bear market has lasted 237 days, making it the fourth longest bear market in history. Bitcoin has fallen from its all-time high of approximately $124,800 in January 2025 to approximately $58,115 on June 25, with a maximum drawdown of 53.43%, the smallest drop in any bear market in history.

CoinGecko believes that the relatively mild decline in this bear market may be due to increased institutional participation, the gradual maturation of market infrastructure, and the influence of macroeconomic factors such as interest rate fluctuations and capital flows into artificial intelligence. In contrast, the three bear markets triggered by major industry events in 2014-2015, 2018-2019, and 2022-2023 all saw maximum declines between 76.7% and 83.6%.

CoinGecko points out that historical data shows Bitcoin typically takes 65 to 166 days to climb back above the 200-day moving average after confirming a temporary bottom. If the low on June 25th holds true, it could potentially climb back above this moving average as early as the end of August this year, but a further extension of the recovery period cannot be ruled out.