Quantitative Analysis: Fable 5 Banned for 18 Days — Who Ate Anthropic's Pie?

On June 9, 2026, Anthropic released Fable 5, a Mythos-class coding model the community called a “beast.” Three days later, the Department of Commerce imposed an export control order, forcing Fable 5 offline worldwide. On July 1, the ban was fully lifted, and access was restored on July 3. Anthropic experienced a dramatic reversal over those 18 days: first releasing the world’s strongest coding model, then having its own government pull the plug.

Is this a net positive or negative for Anthropic? Intuition offers two opposing takes. One holds that the ban window gave Fable 5 global exposure — the government effectively stamped it as “truly frontier” — and demand will rebound once the ban lifts. The other argues that the ban window let competitors rush into the vacuum, and developers who migrate won’t come back. We approached this from two angles using semi-quantitative methods: the objective angle pulls daily token traffic data for all models from the OpenRouter API, and the subjective angle tracks community sentiment evolution on HN and Reddit. Both signals point to the same conclusion: the pie is growing, but Anthropic’s share is shrinking — eaten by GLM.

The Pie Is Growing, But Anthropic’s Share Is Shrinking

There are 446 trackable models on OpenRouter. We grouped them by vendor and computed each vendor’s weekly prompt token volume and share using daily API data.

OpenRouter’s total traffic grew steadily through June. Week 1 (6/2–6/8) saw ~24T total, Week 2 (6/9–6/15, Fable 5’s release week) jumped to 29T, Week 3 (6/16–6/22, GLM-5.2’s open-source week) hit 32T, and Week 4 (6/23–6/29) held steady at 32T. The pie is indeed growing, but that doesn’t mean all vendors are benefiting.

Anthropic’s share is declining continuously. Week 1 stood at 19.1%, Week 2 rose to 20.7% (buoyed by the Fable 5 release), Week 3 dropped to 17.7%, and Week 4 fell further to 17.6%. From peak to trough, Anthropic lost 3.1 percentage points in two weeks.

GLM is the biggest beneficiary. GLM’s share on OpenRouter was zero in the first two weeks of June. In Week 3 (GLM-5.2’s release week), it appeared at 4.7%, and Week 4 rose to 6.7%. GLM went from zero to 6.7% in just two weeks.

Other Chinese models are growing too. Xiaomi rose from 12.3% to 15.1%, DeepSeek went from 24.5%, first rising then falling to 22.1%. OpenAI saw modest growth on OpenRouter, from 3.4% to 4.5%. Google declined from 10.5% to 7.3%.

The key takeaway: the 3.1 percentage points Anthropic lost were primarily eaten by GLM. But GLM’s growth (6.7 pp) exceeds Anthropic’s loss (3.1 pp), meaning GLM took share from other vendors as well, especially DeepSeek and Google. The pie is growing, but Anthropic isn’t participating in that growth — its absolute traffic fell from 5.95T in Week 2 to 5.63T in Week 4, shrinking in a growing market.

Daily Data Exposes a Puzzling Intermediate State

Weekly data tells us “who ate whom,” but switching to daily granularity reveals a phenomenon that at first glance seems hard to explain — and that reveals how the migration happened.

GLM-5.1 is GLM-5.2’s predecessor. In the first half of June, its daily average traffic was just 63B — a nearly forgotten old model. After the ban, it fell normally to 43B over the weekend. But on June 16 (Tuesday), it suddenly spiked to 231B, and on June 17 to 345B — 4 to 5 times the baseline. On those two days, GLM-5.2 had just been released and its traffic was still tiny. GLM-5.1’s traffic was actually higher than GLM-5.2’s. Then on June 19, GLM-5.1 collapsed to 67B, subsequently stabilizing at the 30B level — half the baseline. Meanwhile, GLM-5.2 took over 250–390B of daily traffic.

This traffic shape suggests that GLM-5.1 first absorbed a wave of traffic on the day GLM-5.2 launched, then was replaced by GLM-5.2 within 48 hours. Whether this was automatic downgrade routing in agent toolchains, or developers manually trying and switching, we currently lack logs or developer self-reports to distinguish between the two possibilities. What can be confirmed: the substitution occurred within 48 hours, and GLM-5.1’s spike and collapse precisely coincide with GLM-5.2’s release timing.

HN and Reddit Sentiment Followed Two Opposite Arcs

Behind the gateway traffic shock lies a shift in the developer community’s trust landscape. On HN and Reddit, sentiment during the ban window traced two arcs in opposite directions.

The first arc was anger at the government, peaking on June 12, the day the ban order was issued. On HN, Anthropic’s ban announcement post drew more attention than the launch post from three days earlier, and Anthropic’s official statement on X received 13.3 million views. At this point, Anthropic was the victim, and community sympathy for it reached its highest point of the entire event. But this arc began declining from then on; anger at the government gradually diminished over two weeks.

The second arc — the collapse of trust in Anthropic itself — began warming up on June 14 and surpassed the first by July 2. The turning point was a highly upvoted analogy on HN: imagine a company called the “Doomsday Device Company,” constantly promoting its doomsday device as the best and most powerful, while simultaneously arguing that doomsday devices are too dangerous and should be regulated. Meta Chief AI Scientist Yann LeCun publicly mocked: “The fearmongering about AI has finally been punished, hoisted by its own petard.” (Business Insider KOL roundup) The community began using “Feeble” to mock an expectedly neutered Fable upon return.

The real inflection point came on July 1, the day the ban was lifted. The returning Fable 5 came with three conditions: automatic fallback to Opus 4.8 for coding and debugging tasks, a weekly usage cap at 50%, and a switch to credits-based billing after July 7. The reaction on HN was: bait and switch. The redeployment post on HN that day got only 60 points and 12 comments — a brutal contrast to the massive attention from two weeks earlier. By July 2, negative sentiment toward Anthropic’s business strategy had surpassed anger at the government.

Competitive Positioning and Summary

GLM-5.2’s role in this substitution wave is reflected not only in OpenRouter’s traffic data. Within 16 days, it escalated from HN technical threads to the mainstream agenda of Reuters, NYT, CNBC, and ABC News — and this narrative escalation is itself a signal.

GLM-5.2 was released on June 16, exactly one day after Fable 5 was banned. MIT license. Priced at one-tenth of Fable 5. In independent security testing by Semgrep, GLM-5.2’s vulnerability detection F1 score surpassed Claude Code, and the test report reached the HN front page with 1,099 points. Polylabs positioned it as: “If you take Fable out of the equation, GLM-5.2 is currently the best coding model.” VentureBeat reported that it beat GPT-5.5 on multiple long-horizon coding benchmarks at one-sixth the cost. Cline and Kilo Code integrated it on release day.

Over those 16 days, mainstream media formed a remarkably consistent narrative framework: the U.S. tied its own hands, China rushed into the gap, and open-source models are closing in on closed-source. NYT DealBook ran it in its capital-markets decision-maker column rather than the tech section; Reuters, after interviewing Z.ai at their Beijing headquarters, wrote GLM-5.2 and the Fable ban in the same sentence; David Sacks on the All-In podcast called it “just a tick below Opus 4.8”; Marc Andreessen said “first Chinese AI model to match and often beat the American big lab.” Reuters also cited RAND data: within two months of DeepSeek R1’s release, Chinese LLMs’ global market share jumped from 3% to 13%. (Reuters 6/25, Reuters 7/2)

But narrative escalation to the mainstream agenda doesn’t mean ecosystem lock-in has already happened. Discussions on HN are more honest than mainstream media. Some developers said they spent a weekend and $20 with GLM-5.2 to complete a Matrix bot and a Rust agent, and “nothing felt off.” But others said GLM-5.2’s reasoning traces read like “the internal monologue of an anxiety disorder patient” — going in circles, doubting itself, hallucinating side quests. One technical detail ignored by mainstream media: GLM-5.2’s per-token price is low, but its token consumption is about 42k, comparable to Opus 4.8’s 41k, while GPT-5.5 takes only 16k. The task-level cost advantage isn’t as large as the headline price difference suggests. Poe Zhao’s assessment in the second Reuters piece is closer to reality: “partial routing, not overnight replacement” — developers will use GLM-5.2 as one routing option, not as a wholesale replacement.

Western frontier models have fallen into a self-limiting dilemma. GPT-5.6 was likewise asked by the government to restrict its release (CNBC). Sam Altman publicly criticized: “I don’t like the government picking winners and losers for access to AI.” But OpenAI complied. AI benchmarker Chris poured cold water on it: “GPT-5.6 will be an incremental/solid improvement over GPT-5.5, not a Fable killer.” (ExplainX) OpenAI technical staffer Ryan Brewer’s take best captures the developer consensus: “If the U.S. government continues down this path, you’ll end up with frontier intelligence only accessible from a handful of buildings in the Bay Area.” The community doesn’t view GPT-5.6 as a Fable replacement, but rather as yet another American frontier model that’s equally gated.

Other competitors have limited substitution capacity. Sakana AI launched the Fugu model, which got 247 points on HN, but it’s a 7B-parameter coordinator LLM — its technical scale is not in the same league as Fable/Mythos. China’s 360 launched Tulongfeng, which founder Zhou Hongyi called “China’s version of Mythos” (Reuters), but discussion in the Western developer community focused on the geopolitical dimension, and it didn’t enter the coding agent toolchain.

This ban window episode leaves the global industry with three layers of lessons.

First, multi-model hot-standby has gone from a cost-saving measure to a survival imperative. TrueFoundry put it most precisely: “When you depend on a single model from a single provider, you accept several failure modes you can’t control.” OpenRouter launched its Fusion API, offering an orchestration path to “Fable-level performance at half the price.”

Second, policy risk has formally surpassed technical benchmarks as the primary variable in architecture design. The Cloud Security Alliance characterized the event as a “regulatory kill-switch moving from theory to operational fact,” recommending that enterprise procurement processes incorporate export control compliance reviews.

Third, Forrester judged that, since Fable 5 survived only 72 hours and enterprise integration hadn’t yet deepened, the direct stock loss is limited. But the OpenRouter data points to a different kind of loss: the ban window inserted GLM-5.2 into a position it wasn’t supposed to occupy, the very day after the ban, and within two weeks it went from zero to 6.7% share, while Anthropic dropped from 20.7% to 17.6%. The pie is growing, but Anthropic is the only major player not joining the feast. The 18-day ban window itself is over, but the way the pie is sliced has already changed.

Appendix

Weekly Vendor Shares

GLM-5.1 and GLM-5.2 Daily Prompt Tokens (B)