Frontier AI
Calls to the deepseek-flash model now run DeepSeek-V4.1-Flash-1017, which DeepSeek says finishes tasks faster with shorter output, at unchanged pricing.
Why it matters. Existing integrations get the new model with no code change, so rerun your evals, and DeepSeek recommends Harness 0.2.1-rc.1 or later.
Source: DeepSeek API change log
GPT-6 Sol, Luna and Astra on Amazon Bedrock now take global inference profile calls from Singapore, Jakarta, Malaysia, Thailand and Taipei.
Why it matters. Billing, quotas, CloudTrail and CloudWatch stay in your source Region, while Bedrock picks where each request is processed.
Source: AWS Machine Learning blog
The generally available APIs return ranked web results or a page as Markdown, at $0.75 per 1,000 successful requests and 5 requests per second.
Why it matters. Agent builders get another web search option, and calls through Claude Code, Codex or DeepSeek Harness stay free until November 1, 2026.
Source: DeepSeek Search and Fetch guide
The low, medium and high presets now use Fast Search for web_search, cutting its cost from $2.50 to $1.00 per 1,000 invocations.
Why it matters. Agents on those presets pay less and get results about 800 ms faster, while the xhigh preset keeps standard search.
Source: Perplexity API changelog
Large requests to api.anthropic.com now use gzip compression even when telemetry is disabled, and /advisor arrives on Claude Platform on AWS.
Why it matters. Teams whose proxies inspect request bodies can set CLAUDE_CODE_GZIP_REQUEST_BODIES=0 to turn compression off.
Source: Claude Code releases