Watermarks, Routing, Harnesses and Sampling Settings: Which Controls Recent AI Platform Moves Leave With Developers and Which With Providers

Background

Several recent provider moves, reported between August and the end of October 2026, concern settings that sit between a developer’s code and a model: sampling parameters, text watermarks, model selection and the agent harness. This analysis lays them side by side and records, for each, what the cited sources say about who decides. The grouping and the table below are the Herald’s own; each cell is drawn from a cited source.

Sampling parameters. Anthropic’s migration guide for API version 2026-10-21 says a request that sends temperature, top_p or top_k “returns a 400 error, even at the parameter’s default value”. The guide says models that already accept only the defaults, “such as Claude 4.7 and later models”, sample as before, while on models that accept other values, such as Claude Sonnet 4.6, a request that lowered temperature for more consistent output loses that effect. The Herald reported the release when it appeared.

Text watermarks. According to TechCrunch, the EU AI Act’s transparency rules took effect on August 2 and require AI companies to mark AI-generated content in a way other systems can identify. TechCrunch reported in August that Anthropic said all models released after August 2 would automatically carry watermarking technology for text and files, with older models to follow, applied “at the model level”. In its October report, TechCrunch said OpenAI would watermark ChatGPT and Codex text in the EU only, while developers using its API anywhere could turn the feature on for select models, off by default. The Herald’s earlier report on that announcement covers the same ground.

Current situation

Watermarks: one provider fixes it, another offers a switch

The Decoder reported on October 31 that, in an email to API customers, Anthropic said it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, and that newer models such as Claude Opus 5.5 already carry the mark. According to The Decoder, the email says that on Opus 4.6 and Sonnet 4.6 the watermark applies only when temperature, top_p and top_k are left at their default values, and warns: “If you rely on the watermark for your own compliance, keep these settings at their defaults.” The Decoder notes that the email does not mention that caveat for Mythos 5 or Opus 4.7, which it reads as meaning the mark applies to those regardless of sampling settings.

The Decoder also reports that, because the watermark is embedded at the model level, it works on every platform that hosts the models, including Amazon Web Services, Google Cloud and Microsoft Foundry, and that customers cannot turn it off. It describes OpenAI’s system, textGrain, as shipping disabled for API customers, who can switch it on or off.

Reception is mixed in the reporting. The Decoder says the claim that watermarking leaves text quality untouched “is disputed”, citing blogger John Gruber, who argues that choosing words by a watermark key rather than by meaning degrades precision, and the trade outlet Artificial Lawyer, which raises concerns for law firms. TechCrunch reported in October that Anthropic’s worldwide watermarking drew backlash from some Claude users, who argued they had supplied “the instructions, context, decisions” while Claude was only “the tool”.

Model choice: from picker to router

The New Stack reported on October 28 that GitHub debuted Copilot Fusion, which handles model selection and orchestration for coding tasks. GitHub’s instruction, as quoted there, is to “choose Copilot Fusion and let Copilot intelligently route and orchestrate which models it uses for you.” Developers pick an effort level (Standard, High or Max). The same article says enterprise administrators can disable particular models for Fusion without disabling them for the whole team. GitHub’s performance claims, relayed by The New Stack, are self-reported: Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. The New Stack adds that Ramp described moving model selection into its open-source coding agent Open SWE this month, and that Replit made its “intelligent model routing” the default across every account in the summer.

On the other side, The New Stack quotes Amazon CEO Andy Jassy on the company’s Q3 2026 earnings call saying that “most customers are going to use multiple models, and many will use multiple models in the same generative AI application.” The same article cautions that switching is not free: models differ in context limits, tool capabilities and supported features, so developers still need to test their applications, and a workflow built around one vendor’s agent harness adds complication. It quotes Jassy as saying customers want to “swap them back and forth as they leapfrog each other, as different constraints of their applications require them.”

Agent harness: a choice fixed at creation

The New Stack reported that Microsoft opened its Copilot harness to outside developers in public preview through Foundry Agent Service. Under the heading “Runtime limits and lock-in”, it says the harness “must be selected when creating a prompt agent, and that choice can’t be changed later, even in a new version”, that Microsoft controls the agent’s CPU, memory and idle time, and that bringing your own models “isn’t supported with the managed Copilot harness”. Microsoft’s own post says that “switching the harness after creation isn’t supported yet” and that developers choose “a supported Foundry model”. It offers hosted agents as the alternative for teams that want to bring their own framework, agent loop, middleware, dependencies and runtime behavior while Foundry manages hosting. The post, read on November 1, 2026, lists billing for the harness from December 1, 2026, with harness runtime usage free before that date.

The settings side by side

Setting	Who decides, per the cited source	Source
Sampling parameters on API version 2026-10-21	Anthropic removes them; a request sending them gets a 400 error	Anthropic migration guide
Anthropic watermark	Set at model level; customers cannot turn it off; on two older models it depends on default sampling settings	The Decoder
OpenAI watermark	API customers switch it on or off; EU default for ChatGPT and Codex	TechCrunch, The Decoder
Model selection in Copilot Fusion	Fusion routes; developer sets effort level; administrators can exclude models	The New Stack (relaying GitHub)
Agent harness in Foundry preview	Chosen at creation; not changeable after; own models not supported	The New Stack, Microsoft

Implications

What the sources show is a split, not a single direction. Where regulation or a managed service is involved, the reporting places the decision with the provider: Anthropic’s watermark cannot be disabled by customers, the Microsoft harness cannot be swapped after creation, and Copilot Fusion makes the per-request model choice. Where the sources describe developer choice, it appears as an opt-in or an escape route: OpenAI’s API toggle, Fusion’s effort level and administrator exclusions, Foundry’s hosted agents, and Bedrock’s single-API access to many models, which The New Stack describes as matching Jassy’s argument.

Three consequences follow from what is reported, and each rests on a cited statement rather than on a forecast. First, a team relying on a provider’s watermark for its own compliance is told by that provider, in the email The Decoder describes, to keep three sampling settings at their defaults on two models. Second, The New Stack’s account shows that “multi-model” claims carry limits at the harness layer: Microsoft’s preview allows a choice among supported Foundry models but not a harness change or a model of one’s own. Third, for routers, the only published performance figures in the cited reporting are GitHub’s own, and the reporting does not say whether developers can see which model handled a given request.

Several questions remain open in the sources. The Decoder’s report and Anthropic’s migration guide each concern sampling parameters, but neither cited source says how the 2026-10-21 removal of those parameters interacts with the watermark condition on Opus 4.6 and Sonnet 4.6. TechCrunch’s reporting does not say whether regulators have assessed either approach against the EU rules. And Microsoft’s “yet” is the only signal in the cited material about whether harness switching will be supported later; the post gives no date.