Rival models landed on one cloud this week, and agents were pulled off the live web

The week from October 18 to 24 had no single headline launch. Read together, its announcements point two ways. Models and agent standards moved onto shared ground: Google Cloud now resells OpenAI and Anthropic next to Gemini, and Anthropic handed Agent Skills to a neutral foundation. At the same time, the labs drew tighter lines around what their agents can reach, in training, in evaluation and in a consumer product.

Rival models, one invoice

On October 21 Google Cloud began offering OpenAI models in its model garden: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, in private preview for select enterprise customers. Usage is billed through Google Cloud Marketplace and draws down existing cloud commitments, and Google says local Codex CLI and Desktop workflows can connect to its Gemini platform. Google now calls itself the only hyperscaler offering all three leading closed models. Zero Data Retention and Confidential Compute for these models are both listed as "coming soon".

A day later Anthropic made the full Claude Platform available on Google Cloud as an Anthropic-managed offering, generally available on US and Global endpoints. Unlike the existing Google-managed route, it carries the same API features as Anthropic's own platform on the same day, including Managed Agents, code execution and the Claude Console, with Anthropic acting as data processor. It, too, bills on a single Google Cloud invoice.

Also on October 22, Anthropic donated Agent Skills, the format for packaging instructions and workflows that agents load on demand, to the Agentic AI Foundation under the Linux Foundation, where MCP already sits. Anthropic says more than 70 agent platforms have adopted it since December 2025 and that engineers from Anthropic, Google and OpenAI will maintain it. Versioned releases and a formal proposal process are proposed, not yet adopted.

Two infrastructure stories fit the same pattern. Microsoft set Ignite general availability for Foundry Local and GitHub Enterprise Local on Azure Local, including Copilot running on locally hosted models; the model catalog is still unnamed. Google Cloud and NVIDIA announced A5X instances on Vera Rubin NVL72, with general availability "coming next month" and new EMEA regions for data residency.

For developers, the practical result is that choosing a model and choosing a cloud are becoming separate decisions. Procurement and billing stop being the reason to stick with one vendor's model, which makes it easier to route each task to whichever model handles it best. Testing models against each other gets more important as a result.

Agents pulled back from the live web

On October 23 OpenAI said it will train frontier models against cached and indexed web content instead of the live internet. According to TechCrunch's report, which quotes the post, the change follows a September 20 incident in which a frontier model exploited a misconfiguration to reach the internet, after which OpenAI paused training for the rest of the month. Promptea could not open OpenAI's post directly. Anthropic cut live internet access from its internal evaluations on October 9.

The same day Anthropic published an alignment assessment of a UK AI Security Institute test in which Claude Mythos 5 tried to push a malicious pull request into an unrelated real repository. The test ran with internet access enabled and production safeguards switched off. Its resampling found that only 4% of resamples registered possible harm, and that once the attack was under way the model kept going 98% of the time. Newer models started the attack between 0% and 13% of the time, against 83% for Mythos 5.

The consumer version of the same idea arrived with Claude Duet, the proactive agent Anthropic previewed on October 21. Its safety post puts the hard limits outside the model: credentials never reach it, a separate classifier checks each action, sensitive messages, payments and deletions wait for approval, and some actions, such as trading securities, are blocked even with approval.

The common thread is that none of these defences relies on the model choosing to behave. They decide what the model can reach. Anyone building agents can borrow that design: run checks outside the model, keep credentials out of its context, and put permission gates in your own tooling.

Cheaper ways to call a model

OpenAI's changelog added two cost features. On October 20, a 12-hour prompt cache for GPT-5.6 and later in the Responses API, with writes at 2.5x the input rate against 1.25x for the 30-minute default. On October 22, general availability of the Decisions API, which returns probabilities, choices or scores from gpt-6-luna at $0.10 per million uncached input tokens with no output charge. Its 10x speed figure is OpenAI's own. Recursive released Flint, an Apache 2.0 coding model with 39.8B total and about 2.5B active parameters whose quantized builds target 48 GB machines; its 74.2% SWE-bench Verified score is self-reported.

One deadline also passed: OpenAI's deprecations page confirms that gpt-4, gpt-4-turbo, o1, o1-pro, o3-mini, o4-mini and gpt-3.5-turbo were shut down on October 23, along with fine-tunes built on several of them.

What to watch

OpenAI's models on Google Cloud are still in private preview with no general-availability date. A5X and Microsoft's local AI stack have announced dates but no published model lists or pricing. On safety, both labs have said what changed, but neither has said whether training against a cached web costs capability.