Google's decision to restrict free Gemini users to the 3.5 Flash-Lite model from October 9, and to limit Plus subscribers to 3.5 Flash-Lite and 3.6 Flash1, adds another vector to a pattern hardening across the industry: AI companies are changing models, tiers, and limits under paying customers, and practitioners are treating those changes as a reliability threat on par with downtime. Separately, Anthropic has taken the unusual step of ending enterprise customers' discounts, which typically reach roughly 15%, once they hit usage limits, forcing renegotiations2. An independent monitoring effort has already caught one provider changing behavior without announcement5.
Why it matters
Production AI systems depend on behavioral consistency from the models they call. When a provider deprecates a pinned model version, quietly introduces a usage cap, or swaps a flagship model for a smaller one, the downstream effect is functionally equivalent to a service outage, except no status page fires. These stories each describe a different mechanism of provider drift, but the operational consequence is the same: teams that built on a stated contract discover the contract has moved.
The big picture
Google will soon introduce limits on which Gemini models free users and AI Plus subscribers can access. Free users will be restricted to 3.5 Flash-Lite from October 9, while Plus subscribers will be limited to 3.5 Flash-Lite and 3.6 Flash. That is a tier contraction: users who previously had access to more capable models will lose it on a fixed date.
Anthropic sells its Claude Max 20x subscription with the promise of 20 times more usage than its $20-a-month Pro plan3. Developers paying $200 a month can still hit a separate weekly ceiling. According to the complaint in an expanded class-action lawsuit filed Tuesday, Anthropic introduced weekly limits in late July 2025, months after Max launched in April. The plaintiffs allege those constraints were not adequately disclosed during subscription. Anthropic's documentation states that its Max 5x and 20x multipliers apply per session, with usage limits resetting every five hours, but the five-hour window is not the only limit: Anthropic also imposes a weekly usage cap across all models and says it may add other restrictions to manage capacity. Claude Code users' interactive coding sessions count toward the same plan limits.
On the enterprise side, Anthropic has been ending customers' discounts once they hit usage limits, forcing renegotiations. Consumer and enterprise customers are now experiencing the same dynamic from different directions: consumer tiers face post-launch constraint addition, while enterprise contracts face mid-term price resets.
At OpenAI, the Go plan lost access to the GPT-Live-1 model and switched to the mini version7. One subscriber noted that the difference between Free and Go keeps getting smaller: in August, Luna chats became unlimited on both Free and Go; in September, three scheduled tasks were added to Free, which had previously been a Go feature. Go costs $8 in the US plus ads. OpenAI's Pro $200 tier is returning with what was described as "half the dollar in API spend"6.
Meanwhile, a Towards Data Science analysis argued that the recurring cost of production AI is not inference but "re-qualification: the eval reruns, prompt retuning, and regression testing you owe every time a model changes under you"4. And an independent developer has been running the same private set of probes every day since August 20 against 34 models from 15 labs. On September 10, DeepSeek's reasoner started using roughly 10 to 12 times the thinking tokens it had used for the previous three weeks on the same probes, with no announcement found. Every reading is hashed into a public transparency log.
ANALYSIS Each strand describes a different vector of drift. Anthropic's consumer case is about post-launch constraint addition: a pricing promise made in April was narrowed by a weekly cap introduced months later, while the 5x and 20x marketing language persisted. Anthropic's enterprise moves represent mid-contract repricing. Google's Gemini change is scheduled tier contraction. OpenAI's Go tier illustrates capability subtraction: a model swap from GPT-Live-1 to its mini variant removed the feature a subscriber identified as the primary reason to pay. The DeepSeek finding represents silent behavioral change: no capability was removed, but cost and latency shifted without disclosure.
Anthropic argued in its motion to dismiss an earlier version of the case that customers could access information about the limits through hyperlinks during purchase. Anthropic compared those disclosures to information on a product label that customers can find by turning over the package before deciding whether to buy it. That defense frames limit disclosures as a pre-purchase transparency question, but the complaint's core issue is that the weekly cap was added after launch, meaning no amount of label-reading at purchase time could have surfaced a restriction that did not yet exist.
Anthropic continued to market the plans with its 5x and 20x usage claims after introducing weekly limits. If the plaintiffs succeed, the case could set a precedent for how clearly AI providers have to explain those restrictions before developers sign up.
The five strands together point to provider drift not as a single failure mode but as a category: it includes constraint addition (new caps), capability subtraction (model swaps), version deprecation (pinned models retired), silent behavioral change (token-count shifts), and mid-contract repricing (discount clawbacks). Each requires a different defensive posture from production teams.
What's next
Google's October 9 date for restricting free Gemini users to 3.5 Flash-Lite is the next scheduled tier contraction. The Anthropic class-action will test whether AI subscription marketing must disclose not just current limits but the provider's right to impose future ones. The independent seismograph project monitoring 34 models from 15 labs continues to publish hashed readings, offering the closest thing the field has to a third-party drift-detection service. For engineering teams, the Towards Data Science framing points toward a practical budget line item: re-qualification cost per provider change.