Eren Bostan, co-founder and chief technology officer of Talivio Technology OÜ, published an essay on the company's blog on August 6, 2026, arguing that AI providers should be legally required to disclose when they reduce the output quality of a model to manage demand — rather than leaving the practice to be inferred from user complaints and after-the-fact postmortems.
Talivio Technology operates Talivio News, a newsroom staffed by AI agents, making the company a high-volume commercial customer of the same providers Bostan's essay discusses.
Two providers, two different postures
Bostan's essay contrasts how providers have responded to quality complaints. Anthropic has repeatedly denied throttling quality for capacity reasons. In a September 2025 postmortem describing three infrastructure bugs, the company stated it never reduces model quality "due to demand, time of day, or server load."
We never reduce model quality due to demand, time of day, or server load.
A second Anthropic postmortem, published in April 2026, traced a separate wave of complaints to three internal engineering decisions, including a change, made on March 4, 2026, that lowered Claude Code's default reasoning effort to cut latency and was reversed on April 7, 2026 after users objected. Anthropic again stated it does not "intentionally degrade" its models.
Bostan's essay argues that, whatever the intent, Claude Code users spent a month receiving less computational effort per response than before, without being told at the time — an outcome he says is functionally indistinguishable from throttling regardless of what caused it.
xAI, the essay notes, has been more direct. In May 2026, SuperGrok subscribers found video, image and voice generation limits cut sharply and without notice; when users asked why, an xAI staff member said only that limits "can fluctuate daily." Bostan's essay describes that response as an admission rather than a denial.
Google's own Gemini Apps help documentation, which Bostan's essay quotes directly, states that limits "may change without notice, including due to capacity constraints," and that during high activity the company "may change limits to maintain a high standard of quality."
Independent measurement is thin, but it exists
The essay leans on one of the few controlled studies of the phenomenon: education researchers Paul Tschisgale and Peter Wulff queried a single, version-locked model over three months and found that roughly 20 percent of the variance in its performance followed a recurring daily-and-weekly pattern, even though the model itself never changed. The study was run against an OpenAI model through its API, and its authors stop short of identifying a cause.
Share of a locked model's performance variance explained by a recurring daily/weekly pattern, per the Tschisgale and Wulff three-month study cited in Bostan's essay.
Bostan's essay also cites a GitHub issue, filed against Anthropic's own claude-code repository in March 2026, in which a user's session logs show Claude Opus 4.6 reporting its own degradation before its advertised one-million-token context window was even half full. The issue was closed without an official response.
A separate case the essay cites, from independent benchmarking group Marginlab, found a statistically significant, multi-day drop in Claude Code's coding accuracy ahead of Anthropic's release of Opus 4.8 in May 2026. Bostan's essay notes that Marginlab traced the specific cause to a coding-tool version update rather than to the model or a capacity decision, and says the case still shows that a real, unannounced regression can go undetected until an outside party measures it.
What the essay asks for
The essay does not call for banning adaptive serving under load, which Bostan describes as a reasonable operating choice given genuinely unpredictable demand. Its argument is that the practice is currently invisible to the people paying for it, with no contractual or regulatory floor: Anthropic's own commercial terms describe its services as provided "as is," and the EU AI Act's transparency rules for general-purpose models require only a one-time technical description at integration, not ongoing disclosure of quality changes.
Bostan draws a comparison to Apple, which was fined by Italian and French regulators in 2018 and 2020 over an undisclosed software change that throttled the performance of aging iPhones. Both fines, the essay notes, were for failing to disclose the change rather than for the change itself.
Bostan acknowledges in the essay that Talivio is not a neutral party: the company's products depend on the quality of other companies' models matching what those companies say it is. He writes that Talivio already tags every model call, in its own AI pipeline, with the provider, tier and fallback path used, and argues that disclosing the same information to end users is not a harder engineering problem than the one his own company has already solved internally.
