Frontier AI
The Responses API can now keep prompts cached for 12 hours instead of 30 minutes, set with prompt_cache_options.ttl on GPT-5.6 and later.
Why it matters. Cache reads stay cheap for requests hours apart, but 12-hour cache writes cost 2.5x the uncached input rate, so test it on real traffic.
Source: OpenAI API changelog
It fixes a 2.1.304 regression where every request failed on Vertex, Bedrock and other providers whose endpoint or gateway rejects tool search.
Why it matters. Teams that run Claude Code through a cloud provider or gateway and updated to 2.1.304 should upgrade now.
Source: Claude Code releases
Google Cloud customers can pick both models in Antigravity 2.0, Antigravity CLI and the VS Code, JetBrains and Visual Studio extensions once enabled.
Why it matters. Claude usage is billed pay as you go through Google Cloud Billing, outside the pooled Gemini Enterprise token quota, so admins must enable overages.
Source: Google Cloud blog
Platforms send each end customer's profile ID in the anthropic-user-profile-id header, so a safety restriction or ban hits that customer, not the whole organization.
Why it matters. Gateways and resellers can contain one customer's abuse, and beta organization webhooks report when Anthropic restricts a profile or lifts it.
Source: Claude release notes
GPT-6.1 Sol, GPT-6 and GPT-5.6 models can be called from Canada (Central) and Canada West (Calgary) through cross-Region inference, processed in North America.
Why it matters. Data at rest stays in the Canadian source Region, but inference runs outside Canada, so check the profile against your residency rules.
Source: AWS Machine Learning blog
Voxtral 3 adds transcription in over 50 languages, realtime transcription under 500ms, and Voxtral Speech 3 text to speech, on an EU-hosted API or self-deployed.
Why it matters. Voice agent builders get EU-hosted options, with batch transcription from $0.0035 per minute and the voice agent API at $0.04 per minute.
Source: Mistral news
Both rerank-v5.0-pro and rerank-v5.0-fast have a 32k token context window, support over 100 languages, and can return the most relevant passages per document.
Why it matters. They run on the Rerank API, Azure and SageMaker, and Cohere says Fast beats its previous Rerank 4 Pro.
Source: Cohere changelog
Contextual grounding checks gain a Standard tier that handles reasoning and conversational responses, scores reworded answers more consistently and supports more languages.
Why it matters. Existing guardrails stay on Classic until changed, pricing is the same, and Standard needs cross-Region inference with a guardrail profile.
Source: AWS What's New
An optional setting returns a shorter optimized prompt only when it scores at least as high as your original on your evaluation.
Why it matters. Shorter prompts cut input token costs and time to first token, and results report the compression ratio and token counts.
Source: AWS What's New