Anthropic's share fell from 69% to 40% in two months: why enterprises should not lock themselves into one model

The past 24 hours brought another dense round of AI news: Elon Musk's Colossus 2 has scaled to 550,000 GPUs with plans to reach 1.21m by year-end; Nvidia is reported to be acquiring Hugging Face; and Google DeepMind confirmed Gemini 4 has entered late-stage training.

Amid the noise, one development matters more for your business: the ranking of large models is being reshuffled on a monthly basis.

In two months, the leader's share halved

Model capabilities keep advancing, but the stability of the market landscape is gone.

According to Daily AI Insights (2026-09-26), citing data from Vercel AI Gateway, Anthropic's share fell from 69% to 40% in two months, while OpenAI rose from 10% to 24% and topped token volume; open-weight models absorbed roughly half of Anthropic's lost share. Claude Opus 5.5 captured 10% within two days of launch.

The report's more memorable judgment: shares are highly volatile and driven by new product releases.

In plain terms, today's front-runner may be a chaser three months from now.

What does that mean for enterprise users? If you bind your business logic, prompts, evaluation sets and workflows entirely to one vendor's API, "switching models" becomes a painful reconstruction. If your architecture keeps an adaptation layer, switching models is a matter of changing a few lines of configuration.

The point of model selection is not to pick a winner, but to stay able to switch at any time.

Prices keep falling, so last year's math can be redone

Claude Opus 5.5 was released on September 22. According to the official announcement, the cost of typical token-billed workloads is about 40% lower than the previous generation Opus 5, cache read prices fell further, and output speed improved by about 30%. OpenAI then launched GPT-6 Sol and Luna, with API pricing cut by 50% versus the previous generation, aimed at technical tasks and high-throughput scenarios respectively (Xinzhiyuan, 2026-09-24).

Over a longer horizon, a flagship model was updated every 73 days on average in 2023, a figure that has shrunk to 18 days in 2026; OpenAI and Anthropic together have released 42 major models. Epoch AI's observation is that the cost of achieving the same benchmark score keeps declining.

Chinese players have not stood still either. Alibaba unveiled the Qwen-Audio 3.1 speech model series at its Yunqi Conference, covering speech recognition, speech synthesis, audio understanding and real-time interaction, with some API prices cut further, while also disclosing plans for Qwen 4 (DAMO Developer Matrix AI News Daily, 2026-09-25).

Separately, prices compiled by an aggregation service show GLM-5.2 at roughly RMB 3.2 per million input tokens and RMB 11.2 for output after channel discounts, and Kimi-K3 at about RMB 12 per million input tokens after discount. Note these are channel discount prices, not what you pay calling the official API directly, but the direction is consistent.

For enterprises, what matters is not "how much cheaper it got" but "when this line turns from loss to profit". Tasks such as bulk contract comparison, customer service quality inspection and document review that were rejected last year because of high unit costs may now be viable. We suggest a stocktake: list every scenario shelved because it was too expensive, and recalculate it at the new prices.

The barrier is shifting from "can it be done" to "can it be controlled"

This section, we think, deserves the most attention from operators.

On September 25, New York City Council Speaker Menin unveiled a set of AI regulatory proposals: "bounty hunters" who report dangerous AI could share in fines; AI systems would have to pass independent third-party verification (data quality, bias, privacy) and be equipped with a human-controlled emergency shutdown switch, with fines of USD 25,000 per violation; individuals substantially harmed by AI could sue developers directly. An October 5 hearing is expected to summon the CEOs of OpenAI and Anthropic.

A few days earlier, in an NBC interview broadcast on September 25, Bill Gates warned that AI, if used maliciously, "could be enough to cause an event that kills a billion people", and argued that corporate self-regulation alone is insufficient.

In the same period, OpenAI announced it was pausing training of its most powerful model until additional safety guarantees and alignment improvements are in place (2026-09-27 Financial News Summary). Some analysts call September 2026 a month of concentrated "agent loss of control", with the judgment that the moat for AI agents is not just model capability but controllability, auditability and revocability.

In other words, before putting an agent into a core process, the question is no longer "is it smart" but three things: if it gets something wrong, can you trace which step? Can its conclusions be traced back to the original evidence? Does its data stay inside your intranet?

How to put this into practice

If your business already has a step where "AI first produces a result that can be checked item by item", you can look at our own approach. SHEYU AISITE (AI 独立站) aside, SHEYU AIZHISHEN (舍予AI智审) turns technology project application materials into a checkable evidence table: it reads tens to hundreds of pages in seconds, every extracted value carries a page number and original text, and it runs dual-engine cross-checking; four-value results do not pass by default; data does not leave the intranet. It addresses the third question above — not letting AI make decisions for you, but having it hand over a draft you can review and hold accountable at any time.

One judgment to take away

A share can double and halve within two months, prices can be cut in half within a year, and capabilities turn over every six months. At this pace, enterprises should protect two things: an architecture that is not locked to any single vendor, and a process where every step leaves evidence. The first keeps you from falling behind; the second keeps you from crashing.