The Power Switcher Report

Section 1

Who is the Power Switcher?

The default mental model of an OpenRouter user is someone who picks a single LLM and stays there — pin Claude, pin GPT, pin Gemini, never look back. That is not what the data shows.

Across N_users users and token_volume, e.g. "12 trillion" tokens routed in OpenRouter's date_range data, a clear cohort emerges: users who actively switch models — across providers, across sessions, sometimes within a single conversation. We call them Power Switchers.

Power Switchers are a small fraction of the total user base — roughly cohort_pct — but they generate token_share_pct of OpenRouter's tokens and revenue_share_pct of its revenue. They use a median of median_models_pw_per_week distinct models per week, against median_models_typical for the typical user. They are the cohort whose behavior most closely tracks where the AI market is going — because they are the ones doing the picking.

Section 2

When they switch — the triggers

Power Switchers do not switch at random. They switch when something about the work changes — a longer context arrives, a task pivots from chat to code, a model returns a structurally broken JSON object, latency spikes on a paid endpoint, or a coding agent fails its third retry.

We classified N_session_transitions mid-session model transitions across the cohort and grouped them into five trigger families:

1. 01 Task pivot The user's prompt changes domain (writing → code, short-form → long-form, ideation → execution). pct of transitions
2. 02 Length pressure Context length crosses the comfortable window of the current model. pct of transitions
3. 03 Quality failure The previous response was structurally wrong (JSON schema violation, tool-call malformation, hallucinated reference, refusal). pct of transitions
4. 04 Cost optimization The user falls back to a cheaper model after the expensive one delivered. pct of transitions
5. 05 Latency / availability The chosen provider is slow or down. pct of transitions

Section 3

How they route — by task

This is the section dev-Twitter will screenshot. We classified the top tasks Power Switchers route on, then computed the model with the largest revealed-preference share for each — the model the market actually picks, not the model that wins the benchmark.

The clearest patterns:

• finding_coding_python — e.g. "On Python data work, Power Switchers pick {model} {X}% of the time, despite {other model} leading {benchmark}."
• finding_long_context — e.g. "Above 100k tokens, {model} captures {Y}% of routes — even when {other model} costs less per million tokens."
• finding_agentic — e.g. "On agentic coding loops inside {harness}, {model} wins {Z}% — the market has already picked, while public benchmarks still show parity."

Section 4

How they route — by cost

Power Switchers are not uniformly cost-sensitive. We see two clear sub-cohorts inside the cohort: frontier-only switchers (who route exclusively across the top of the price curve) and cost-aware switchers (who fall back to OSS or smaller models on tasks where quality holds).

The cost-aware sub-cohort tells the OSS-share story. When they fall back, they fall back to:

• top OSS escalation target — e.g. "Llama 3.x: N% of OSS fallbacks"
• 2nd OSS escalation target
• 3rd OSS escalation target

Provider economics cross-cut (the Anthropic 12%/46% finding, applied to this cohort): Power Switchers route provider_X_token_share of their tokens to provider_X but pay provider_X_dollar_share of their dollars to that provider. The token-to-dollar gap by provider:

Section 5 · The new conceptual frame

Power Switchers as leading indicator

This is the section that justifies the report's existence. Power Switchers do not just consume the market — they predict it.

For every model launched in time_window, we measured how long it took for Power Switchers to reach a stable share of routes vs how long it took the broader user base. The gap is consistent: Power Switchers reach steady-state adoption a median of lead_days days before the rest of the user base.

This is the framing that makes the cohort interesting rather than just valuable. Watching what the Power Switchers route to this week is watching what the rest of the market will route to next quarter.

Section 6 · The bait

Three predictions for Q3–Q4 2026

Three falsifiable, dated, named-model claims — engineered to be argued with on Twitter and re-checked in the Q4 follow-up report. Each prediction includes (a) the trigger condition, (b) the metric, (c) the threshold, (d) the date the report will check it.

We will publish the check in the Q4 Power Switcher Report. Public predictions, public scores.

Section 7

Methodology

Power Switcher cohort definition: a user is classified as a Power Switcher in week W if they routed to ≥ N distinct models in W and ≥ M distinct models in the preceding 4-week rolling window. The threshold was chosen rationale — e.g. "to capture the top decile of model diversity while excluding bot/eval traffic".

Privacy-preserving aggregation: all reported figures are computed on user-anonymized session data, with all prompts and responses excluded from the analysis pipeline. Only metadata (model, token count, timestamp, task classification) flows to the analysis surface. The classification pipeline itself is open-sourced alongside this report.

Limitations. Self-hosted models routed through OpenRouter are included; users routing entirely outside OpenRouter are not visible. Task classification is a learned classifier with accuracy as noted above on a hand-labeled held-out set. Bot and CI / eval traffic was filtered using the description above; residual contamination is bounded.

Section 8

Acknowledgements + advisory bench

Co-author: Stanford HAI / MIT IDE co-author.

Advisory reviewers — these reviewers provided pre-publication feedback. They did not author the report and the conclusions are OpenRouter's:

• Tamay BesirogluEpoch AI
• Anton KorinekBrookings · Anthropic Economic Advisory Council
• Ethan MollickWharton · Anthropic Economic Advisory Council
• routing literature reviewer — e.g. Tatsu Hashimoto (Stanford)Routing / evaluation
• optional 5th revieweraffiliation

Data team: Megamind · Justin Summerville · data team attribution.