DeepSeek revises API pricing, applying off-peak rates all weekend
要闻
DeepSeek will revise its API peak and off-peak pricing from 00:00 Beijing time on August 23, 2026. All Saturday and Sunday usage will be charged at off-peak rates, while the existing time-based pricing will remain in place from Monday through Friday.
DeepSeek announced that its revised API peak and off-peak pricing rules will take effect at 00:00 Beijing time on August 23, 2026. Peak and off-peak periods will no longer be distinguished on weekends, with all usage charged at off-peak rates, while weekdays will retain the existing time-based pricing structure. Charges incurred before the effective time will continue to be settled under the previous rules. According to the company, the change is intended to let users schedule weekend operations without incurring peak-period prices and to balance compute load across the network. Continued use of the service after the adjustment will be treated as acceptance of the new pricing method; users who do not accept it may stop using the service and request a refund.
Firecrawl launched Firecrawl Developer Index, combining 70M+ docs, READMEs, issues, and merged PRs in one index. The company says it has the highest recall among comparable coding indexes.
Firecrawl has released Firecrawl Developer Index for developer search, with 70M+ docs, READMEs, issues, and merged pull requests. The index supports finding a repo by capability without knowing its name, connecting a bug report with the PR that fixed it, and locating a page that answers a how-to question. According to Firecrawl, most sources are refreshed.
The page demonstrates the results with the CLI query `firecrawl developer "why is my retry backoff not firing on 429" --limit 10`: the client treats 429 as a terminal status, so backoff does not run; the corresponding fix adds 429 to the throttling error list and applies exponential backoff before each retry. Firecrawl says the index has the highest recall among coding-specific indexes. The page uses native web search as the control and says the same model scores 0.01 without a search tool.
Tibo says cache hit rates have declined for some Codex users, with the team investigating
开发生态
Codex lead Tibo said some users’ cache hit rates this week fell below the stable levels seen over the previous several weeks, possibly causing usage allowances to be consumed slightly faster. The team is reviewing logs and monitoring dashboards, and the cause has not been confirmed.
Codex lead Tibo published an update on Codex rate limits, saying that some users’ cache hit rates this week were below the stable levels seen over the previous several weeks. The team is investigating and will provide further updates. According to Tibo, consistently hitting the cache is an important part of using Codex efficiently, and lower hit rates could explain why affected users are consuming their usage allowances slightly faster. He is currently reviewing logs and monitoring dashboards with the team. In the replies, multiple users reported unusually rapid allowance consumption. Others argued that the issue was unrelated to cache hit rates, including one report that allowances continued to decline without any usage, while another user said they had not encountered the problem. The cause has not yet been confirmed.
uv creator revamps Codex CLI startup, making it about 25 times faster
开发生态
Charlie Marsh, the uv creator who has joined OpenAI, says the new Codex CLI includes his rebuilt startup mechanism, making startup about 25 times faster than before and enabling immediate responses after launch.
Charlie Marsh, creator of the Python tool uv and now at OpenAI, reworked the startup lifecycle in the latest Codex CLI release, with the change taking effect in that version. According to Marsh, the rebuilt codex completes startup instantly, starts about 25 times faster than before, and can respond immediately after launching. The optimization is included in the latest version, and users can access the revised startup flow by upgrading Codex CLI.
Cursor updates official documentation: SuperGrok Heavy no longer includes Cursor Ultra
开发生态
Cursor says personal SuperGrok Heavy no longer includes Cursor Ultra. Only that plan and personal SuperGrok Plus can grant Grok Bot usage to a linked account, and the link cannot be removed or transferred once created.
Cursor’s official documentation defines a SuperGrok link as a Grok Bot usage entitlement rather than a Cursor plan. Linking personal SuperGrok Heavy does not grant or switch the account to Ultra, and it does not cancel or replace an existing Pro, Pro+, Ultra, or Teams subscription. Eligibility is limited to personal SuperGrok Plus and personal SuperGrok Heavy; SuperGrok Team and SuperGrok Enterprise are not supported. While the subscription remains active, the allowance applies to the paired Cursor account, and Cursor rechecks the subscription status at every renewal. A link pairs one Grok account with one Cursor account and remains permanent after creation, with no option to unlink or transfer it. An existing paid Cursor plan that includes Grok Bot continues to be billed and renewed separately, and the link does not add a second usage meter. The “Link Grok account” entry appears only on the pre-login “Quickstart” paywall. Users with account or usage issues must contact support and provide the email addresses for both accounts and a screenshot of the plan page.
Anthropic responds to performance degradation concerns, saying its test Claude Code configuration does not affect performance
开发生态
Anthropic addressed concerns that Claude Code 2.1.237 showed 10 at the “high” setting, saying the number came from an API configuration mapping test, was not based on a 0–100 scale, and had no performance impact according to internal evaluations.
Anthropic team member Thariq said that, beginning with Claude Code 2.1.237, some users saw 10 under the “high” reasoning-effort setting because a current API service configuration test used a different mapping for numerical effort values; the model itself had not changed, and internal in-depth evaluations confirmed no impact on performance. Users had previously noted that the value matched the old “low” setting. According to Thariq, the team sometimes tests API service configurations in Claude Code before rolling them out; the scale does not run from 0 to 100, the number has no standalone meaning, and the effort level selected by a user remains the level actually provided. The team is rolling out a simpler mechanism for feedback and performance-regression reports. Users who encounter a clear regression can submit it through /feedback and send the ID to Thariq to receive credits. He also said Opus 5 has a spiky performance profile and that improving consistency and reinforcing its Claude-like style are high priorities for the team.
