APIRank
EN 中文

OpenRouter Jevons Paradox: Why GPT-5.6 Discounts Worked (Aug 2026)

OpenRouter's August 25, 2026 insights blog dropped a Jevons Paradox data set on the GPT-5.6 Terra + Luna discount window (July 27 to August 14): daily Terra token volume rose 5.6x and daily Luna volume jumped 13.8x, while Sol — which stayed at list price in the same window — only saw a 1.11x bump. That three-way comparison is the cleanest natural experiment on discount elasticity the public LLM API market has published. This article walks through the data, what made Luna 13.8x while Terra was 5.6x, who actually stole the 7.8% share Terra/Luna grabbed on OpenRouter, what happened to retention once the window closed, and what the rankings look like two weeks after the promo ended. All numbers are sourced from the OpenRouter insights blog post, the OpenAI pricing page, and OpenRouter's /api/v1/models + /api/frontend/v1/rankings/models live data as of August 29, 2026.

The data: pre / program / post window

OpenRouter defined three windows in their analysis. Pre-period: July 8 to July 26 (19 days). Program: July 27 to August 14 (19 days). Post-period: August 15 to August 20 (6 days, the analysis was published August 25 so the post window was still incomplete). The headline numbers:

  • Terra tokens: 5.6x daily volume in the program window vs pre-period average.
  • Luna tokens: 13.8x daily volume in the program window vs pre-period average.
  • Sol tokens: 1.11x — a "gentle bump" that works as a control group because Sol was un-discounted.
  • Other OpenAI models: tokens "fell slightly" during the window (cannibalization).
  • Outside OpenAI: other models "rose gently" — Anthropic was the only major author whose tokens did not rise.

The Jevons framing matters because the obvious critique of any discount program is that you're just moving spend around — a customer who would have used GPT-5.5 anyway now uses Terra at half price, revenue doesn't grow, profit compresses. The blog's contribution is showing that's not what happened. The token volume grew by 5x-13x while Sol (a model many of those same customers could have used instead) barely moved. The demand was latent; price unlocked it.

Why Luna 13.8x while Terra was only 5.6x

Luna is OpenAI's cheap tier in the GPT-5.6 family. On the official list Luna was $1.00/M input and $6.00/M output. OpenRouter's program discount cut that 50% to $0.50/M and $3.00/M. On July 30 OpenAI also cut its own Luna list by 80% on top, taking the upstream price to $0.20/M and $1.20/M. With both cuts combined, Luna's effective price on OpenRouter from July 30 onward was $0.10/M prompt and $0.60/M completion — a 90% total discount versus the original list.

Terra, in contrast, is OpenAI's mid tier. Original list $2.50/M input and $15.00/M output. OpenRouter's 50% cut took it to $1.25/M and $7.50/M. OpenAI's additional 20% list cut on July 30 took the upstream to $2.00/M and $12.00/M. With both cuts, Terra's effective price from July 30 onward was $1.00/M prompt and $6.00/M completion — a 60% total discount.

The two effective price points explain the asymmetry. Luna was 10x cheaper than its starting list and ~10x cheaper than Terra's final discounted price ($0.10/M vs $1.00/M prompt). The elasticity was non-linear: a 90% cut produced a 13.8x volume bump, a 60% cut produced a 5.6x bump, and a 0% cut produced 1.11x. If you model price elasticity as roughly log-linear, that's consistent with each 1% price cut producing ~0.1-0.15% volume growth — which lines up with what OpenRouter showed in their footnotes.

Who lost the 7.8% share Terra/Luna grabbed

Terra + Luna moved from 0.7% to 7.8% of all OpenRouter tokens — a 7.1 share-point gain. The blog breaks down where that share came from:

  • 5.3 points came from non-OpenAI competitors (other labs' models on OpenRouter).
  • 1.9 points came from cannibalization within the OpenAI family (older GPT-5.x and other OpenAI models).

That's roughly three quarters of the gain came from outside OpenAI. Across the entire OpenAI family, share grew from 7.1% to 12.4% and "occasionally crested over 15%" on specific days during the program. So the competitive story is: OpenRouter's 50% discount pulled demand from Claude, Gemini, DeepSeek, and Qwen customers who were already evaluating a GPT-5.6-tier model, and the price made them switch. The GPT-5.6 promo did not, contrary to a cynical reading, just move the same customers from one OpenAI model to another.

One footnote that didn't make the headline: Anthropic was the only major author whose tokens did not rise during the window. Anthropic customers appear to be sticky on Claude — pricing alone does not move them. If your routing layer has an "auto-fallback to Claude on failure" pattern, this is evidence that Claude customers are a population you do not want to churn by default — they will stay through pricing changes and only leave on capability gaps.

The promotion stacking problem

OpenRouter's blog is understated on one mechanic that matters for buyers: the discounts stacked. OpenRouter cut 50% from July 27. Three days later OpenAI cut Luna 80% and Terra 20% from its own list. The combined effective discount from July 30 onward was 90% for Luna and 60% for Terra. From a buyer's perspective, that was an unexpected second price cut inside an already running promo window. From OpenRouter's perspective, the page wasn't lying — both discounts were real and simultaneous — but the headline "50% off" understated what the channel was actually charging.

Tier Original list (per M) OpenRouter promo (per M) + OpenAI cut Jul 30 (per M) Effective total discount
Luna $1.00 / $6.00 $0.50 / $3.00 (-50%) $0.10 / $0.60 (-90%) 90% off
Terra $2.50 / $15.00 $1.25 / $7.50 (-50%) $1.00 / $6.00 (-60%) 60% off
Sol $5.00 / $30.00 No promo No list cut 0% (control)

All numbers above verified against the OpenAI pricing page (platform.openai.com/docs/pricing) and OpenRouter's /api/v1/models endpoint on 2026-08-29. Cached input and cache-write prices are documented in the OpenAI pricing table but were not part of the Jevons analysis — the headline 5.6x / 13.8x / 1.11x numbers are uncached prompt + completion only.

What happened after the discount ended

The retention data is the most important number for buyers thinking about whether to commit production traffic to a discount-tier model. OpenRouter measured it two ways:

  • By customer count: of 100,000+ customers who used Terra or Luna during the program, ~32% retained some usage in the six days post-program, and ~18% ran at or above their program pace.
  • By token volume: the post-program period saw 1.38x the daily Terra + Luna tokens of the program period on average — meaning the retained accounts are far larger than the median program user.

That's a meaningful distinction. If you only looked at customer count, you'd conclude "two-thirds of trial users churned when the promo ended, this is just a discount-driven spike." That's true on a customer-count basis. But by token volume, the post-promo period actually exceeded the promo period — because the 32% who stayed are heavier users than the median trial user. From a routing-layer perspective: the cheap tier keeps its seat at the table after the promo ends, even if you lose most of the casual trial traffic. The blog flags that the post-period is "6 days so far vs a 19-day program, so the story may change as more data comes in" — so the 1.38x number is preliminary, but the qualitative pattern (small core of heavy retained users + large drop in casual trial traffic) is the typical shape for a successful discount program.

Sol as a control group

The blog explicitly used Sol as a control. Sol was the flagship model in the GPT-5.6 family — same launch date (July 9, 2026), same OpenAI family, same OpenRouter channel — and it was not discounted during the Terra/Luna window. Sol saw a 1.11x token bump. The footnote explains why that bump exists: Sol averaged 79.1B tokens/day during the Terra/Luna program against 71.2B/day pre-period, an effectively flat control group. However, the shaded region in the chart above from Aug 17 onward is Sol's own 50% discount. Sol jumps immediately, reproducing the Terra/Luna pattern.

Translation: as soon as Sol got its own discount on August 17, it jumped in exactly the same pattern. That replicates the Jevons effect within the same family and rules out the alternative explanation that "all OpenAI models were rising in this window because of brand momentum." Brand momentum was not the variable. Price was.

What the rankings look like two weeks after the promo ended

Pulling OpenRouter's /api/frontend/v1/rankings/models?window=week on 2026-08-29 (six days post-program, with Sol now in its own promo), the top six slots by prompt-token volume on the daily rankings are all Chinese open-weight models:

  1. z-ai/glm-5.3-flash-20260826 — 4.43T prompt tokens (free standard).
  2. tencent/hy4-preview-20260827 — 1.89T prompt tokens.
  3. MiniMax-M3 (free) — 1.81T prompt tokens.
  4. MiniMax-M2.7 (free) — 136B prompt tokens.
  5. inclusionai/ling-3.0-flash-fin-20260827 (free) — 127B prompt tokens.
  6. qwen/qwen3.8-flash-20260826 — 47B prompt tokens.

None of the GPT-5.6 family appears in the top 15 of the daily rankings as of 2026-08-29. That's the second-order finding the headline blog post does not make explicit: the Jevons bump was real but ephemeral. By the time the discount window closed and Sol was the only remaining GPT-5.6 family member on promo, the cumulative volume ranking was dominated by models with much lower list prices (or free tiers). The discount unlocked latent demand, but the demand redistributed toward the cheapest available option once the relative-price gap closed.

What this means for your routing layer

Three concrete takeaways for anyone running a multi-model LLM stack:

  1. Price elasticity is large but non-linear. A 90% cut produced 13.8x volume; a 60% cut produced 5.6x; a 0% cut produced 1.11x. The relationship between price drop and volume bump is steep enough that a small change in relative pricing on your routing layer can shift a meaningful fraction of traffic. If your gateway logic treats GPT-5.6 Luna and Qwen3.8 Flash as interchangeable cheap-tier options, the daily rankings suggest you'll see Luna spike when it's under-promo and crater when it's not.
  2. Promo-stack detection matters. OpenRouter's "50% off" understated the effective 90% off Luna that customers actually got once the OpenAI list cut landed on July 30. If your cost-monitoring alert is keyed to the OpenRouter promo page and you don't re-fetch the upstream OpenAI list daily, you'll under-estimate the actual discount for at least a few days after the upstream cut.
  3. The post-promo 1.38x is the real number for capacity planning. Don't plan capacity around the 13.8x in-program peak; plan around the 1.38x post-promo baseline plus a healthy margin. The 13.8x was a transient. The 1.38x represents the customer base that actually chose to stay at the higher price, and they will likely be the long-tail volume for as long as the model remains competitive.

Wiring the failover

The pattern the Jevons data argues for: route your cheap tier through a gateway that can swap the model on price-drop news without code changes. The two cheapest paths for that today are Cloudflare AI Gateway (which exposes OpenRouter as a Unified Billing provider and lets you re-point a route in the dashboard) and Portkey (which lets you set per-route model fallbacks and per-environment pricing tiers). For a workload that was running on GPT-5.6 Luna during the July 27 to August 14 window, the cheapest equivalent post-promo is one of the top-five open-weight models above — and the routing logic that handles the swap is the same logic that will handle the next discount window.

Running this routing layer across multiple providers and watching the price-drop news in real time? FreeModel bundles OpenRouter, direct OpenAI, Anthropic, Gemini, Cloudflare Workers AI, DeepSeek, and Qwen behind one OpenAI-compatible API with daily price-drop alerts and per-route failover — useful when the next discount window opens and you need to flip your cheap-tier route within the hour rather than the next deploy cycle.

How to keep this data fresh

The headline numbers in this article were published by OpenRouter on August 25, 2026. The post-period observation window was August 15 to August 20 (6 days). For an updated view as more post-program data lands, three endpoints worth bookmarking:

Re-fetch the rankings once a week and the models endpoint once a day if you depend on a specific tier being in promo. The OpenRouter blog explicitly flagged the 6-day post-period as preliminary — by the time the August 31 weekly rollup lands, expect a follow-up post with a longer post-program window.

Bottom line

OpenRouter's Jevons data is the cleanest evidence to date that prompt price is a first-order variable in API adoption. The 13.8x / 5.6x / 1.11x trio is a price-elasticity ladder, not a brand story. The retention curve shows that a successful discount program leaves behind a small heavy-user core that exceeds the program's median daily volume — which is the durable revenue. And the post-program rankings show the cheap-tier throne is contested by every open-weight model under $1/M, so the next Jevons window will probably belong to whoever cuts list price fastest, not whoever runs the largest promo.

For API buyers, the action is the same as it was last quarter: keep your cheap tier in a gateway that can re-route within an hour, watch the OpenRouter rankings weekly, and don't lock production traffic to a single discount-tier model without a fallback to an open-weight equivalent that costs less than $1/M prompt.