DeepSeek Revises API Pricing: Off-Peak Rates to Apply All Weekend
要闻
DeepSeek announced last night that starting at midnight on August 23 (today), its API will no longer distinguish between peak and off-peak hours on weekends, with off-peak rates applying throughout the day. The existing time-based peak and off-peak pricing will remain in effect on weekdays.
DeepSeek announced that it will adjust its peak and off-peak pricing rules starting at 00:00 Beijing time on August 23, 2026 (today): On weekends (Saturday and Sunday), there will no longer be any distinction between peak and off-peak hours, and all API usage will be charged at off-peak rates throughout the day; on weekdays (Monday through Friday), the existing time-based peak and off-peak pricing will remain in effect. Charges incurred before the new rules take effect will still be settled under the previous pricing rules. According to the company, the change is intended to give users greater flexibility in scheduling operations on weekends without worrying about peak-hour costs, while balancing compute loads across the network. Users who continue using the service after the adjustment will be deemed to have accepted the new pricing model; those who do not agree may stop using the service and request a refund.
Firecrawl has launched Firecrawl Developer Index, a search index designed specifically for coding Agents. The company says it covers more than 70 million primary sources and delivers the highest recall among comparable coding indexes.
Firecrawl has officially launched the Firecrawl Developer Index, a search index built specifically for coding Agents. The index provides access to more than 70 million primary sources across repositories, documentation, and issues. Its content includes issues, merged PRs, and README files from popular public repositories and documentation sites, enabling Agents to independently answer questions about code behavior, API contracts, error messages, known bugs, and more. According to official data, it achieved 63% recall@10 across 1179 real-world developer queries covering repositories, documentation, and PRs, which Firecrawl says is approximately 10% higher than the next-best external provider.
Tibo Says Cache Hit Rates Have Dropped for Some Codex Users as Team Investigates
开发生态
Codex lead Tibo posted an update on unusual Codex usage, noting that cache hit rates for some users have fallen below their previously stable levels this week, which may explain the faster usage consumption. The team is investigating by reviewing logs and monitoring dashboards and plans to provide another update later.
Tibo, head of Codex, posted an update on Codex rate limits, stating that some users had lower cache hit rates this week than the stable levels seen in previous weeks. He said this could explain why those users were consuming their usage allowances slightly faster, as consistently hitting the cache is an important part of efficient usage. The team is investigating and will provide an update. Tibo also said that he is currently reviewing logs and monitoring dashboards with the team. Multiple users replied that their allowances were being consumed unusually quickly, while others disputed that cache hit rates were the cause, with some reporting that their allowances continued to decline even when they were not using the service. Other users said they had encountered no issues, and the cause has yet to be confirmed.
uv Creator Reworks Codex CLI Startup, Making It About 25 Times Faster
开发生态
Charlie Marsh, the creator of uv who joined OpenAI, said he reworked the startup process in the latest version of Codex CLI, making it about 25 times faster and enabling near-instant responsiveness.
Charlie Marsh, the creator of the Python tool uv who has joined OpenAI, said that he reworked the startup lifecycle in the latest version of Codex CLI, making codex launch instantaneously—approximately 25 times faster than before—and respond immediately. The optimization has been released with the new version and is available to users upon upgrading.
Cursor Updates Official Documentation: SuperGrok Heavy No Longer Includes Cursor Ultra
开发生态
Cursor's official documentation now shows that linking SuperGrok Heavy only provides Grok Bot usage, with references to complimentary Cursor Ultra access removed. A staff member said users who have already paid and successfully activated the free Ultra benefit will not be affected.
Cursor’s official documentation now explicitly states that linking a personal SuperGrok Plus or SuperGrok Heavy account grants a Grok Bot usage allowance rather than a Cursor plan, and SuperGrok Heavy no longer includes Cursor Ultra. Regarding existing accounts that have already claimed the benefit, staff said that users who had paid and successfully activated the complimentary Ultra plan would not be affected.
Anthropic Responds to Performance Degradation Concerns, Says Claude Code Configuration Test Had No Impact
开发生态
Responding to concerns that Claude Code had been degraded after its model displayed “10” at the “high” reasoning-effort setting, Anthropic member Thariq said the discrepancy was caused by different numerical mappings during an API service configuration test. He said users' selected effort levels remained unchanged and that internal evaluations found no performance regression.
Users previously found that, starting with Claude Code 2.1.237, the value read by the model at the “high” reasoning-effort setting was 10 out of 100—the same value as the “low” setting in earlier versions. Anthropic team member Thariq responded that the team sometimes tests API service configurations in Claude Code before rolling them out more broadly, and that a test currently underway maps numerical effort values differently, causing some users to see “10” at the high setting. He said the scale does not run from 0 to 100 and that the number itself has no independent meaning; the effort level users select is the effort level they actually receive, and in-depth internal evaluations have confirmed that the change does not affect model performance. He added that the model itself has not changed at all, and that the team is rolling out more convenient mechanisms for submitting feedback and reporting regressions. If users notice a clear regression, they can submit it via /feedback and send him the ID to receive credits. He also acknowledged that Opus 5 is a “spiky” model, and said making it more consistent and more “Claude-like” is a high priority for the team.
Note: This content was created with AI assistance and may contain hallucinations and errors.
