Skip to content
Sign in
Rung 13 Inference & Token-Pricing EconomicsSwitch rung

Inference & Token-Pricing Economics — Timeline

Dates are publication dates of the underlying source unless marked [Rung] (something done to this rung) or [Observed] (a dated price observation). Every cost or price figure carries its as-of date because these numbers go stale fast. Items marked sibling rung are registered elsewhere and appear here as context only.

2026

September

  • Sep 24 — [Rung] Three sources ingested on approval, the first since Aug 6: OpenAI's GPT-5.6 price cut (raw), Anthropic's Fable 5.1 cache-read cut (raw) and Intercom's Fin outcome pricing (raw). Price-rung persistence and the OpenAI pricing page corrected: OpenAI does cut its rungs.
  • Sep 24 (earlier) — [Rung] Full recompile from the fourteen sources already held — no new source. Links to raw files and concept pages that had moved to serving, capital-structure and models were re-pointed; claims those sources carried were moved to "see also" context. The Modular handbook (ingested Aug 6) was compiled for the first time. Several claims corrected (see frontier.md). lastIngested left unchanged on purpose.
  • Sep 23 (about) — [Observed] Intercom's help article on Fin pricing shows its last update: $0.99 per outcome, one charge per conversation however many actions Fin takes; a sales qualification is $9.99. Read Sep 24. (raw)
  • Sep 22 — [Observed] OpenAI releases GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50 per million tokens; its pricing page lists GPT-6 Astra at $10/$50. A new generation steps below the old frontier price, beside a new top tier above it. (OpenAI changelog, read Sep 24.) (raw)
  • Sep 11 — [Rung] Tiered Super-Moore's Law re-read in full (arXiv HTML, 23 pages). The abstract-only reading was upgraded, not a new source: weak fits (economy R² 0.192), price-band tiers, a 62-milestone trend, and a reasoning premium that swings 0.51–83.3× by quarter. (raw)
  • Sep 9 — [Report, sibling rung energy] EIA forecasts record US electricity generation in 2026 and 2027, with data-centre build the biggest regional driver. (raw)
  • Sep 1 — [Disclosure] Anthropic releases Claude Fable 5.1: input/output unchanged at $10/$50, cache reads cut 75%, $1.00 → $0.25; Anthropic estimates ~25% off typical workloads, up to ~45% off highly agentic ones. (raw)

August

  • Aug 21 — [Observed] OpenAI cuts GPT-5.6 Sol from $5/$30 to $4/$20 — the frontier rung moves. Labelled promotional, "available at least through November 21, 2026". (OpenAI changelog.) (raw)
  • Aug 19 — [Rung] The Token Price Index is retired and becomes a closed archive. The rung's price-rung observations stop being a live instrument. (price-rung persistence)
  • Aug 13 — [Observed] Anthropic's new generation shows the same shape as OpenAI's: frontier rung held (Opus 5 = Opus 4.8 at $5/$25), strong rung cut (Sonnet 5 at $2/$10 vs Sonnet 4.6 at $3/$15), with a scheduled return to $3/$15 cancelled on Aug 10. Own-dataset and vendor-page reading; no raw close-read registered on this rung. (price-rung persistence)
  • Aug 10 — [Filing, sibling rung capital-structure] CoreWeave closes a $2.6bn GPU-backed loan at SOFR+550, up from SOFR+450 in May. (raw)
  • Aug 6 — [Rung] The Modular LLM Inference Handbook is ingested on approval — the rung's last ingest to date. The same day's KB cull removes the Patel podcast from the register. (raw)
  • Aug 3 — [Observed] The Token Price Index records OpenAI's GPT-5.6 Terra and Luna cuts (see Jul 30). Own dataset, primary against OpenAI's pricing page; no raw close-read registered on this rung.

July

  • Jul 30 — [Disclosure] OpenAI cuts GPT-5.6 Terra 20% ($2.50/$15 → $2.00/$12) and Luna 80% ($1/$6 → $0.20/$1.20); Sol holds at $5/$30 for now. Stated cause: GPT-5.6 optimised its own serving — 20% lower serving cost from GPU kernel work, 15%+ better token generation from speculative decoding. Fast mode replaces Priority Processing (up to 2.5× faster at 2× the price). (raw)
  • Jul 29 — [Filing, sibling rung hardware] SK hynix reports a 76% operating margin for Q2 2026 — memory scarcity as a margin. (raw)
  • Jul 16 — [Observed] GPT-5.6 recorded into the Token Price Index: Sol and Terra land on the prior generation's exact prices ($5/$30, $2.50/$15); Luna is a new mid point at $1/$6. (TPI)
  • Jul 9 — [Disclosure] OpenAI releases the GPT-5.6 family — Sol, Terra, Luna — per its own API changelog (read 2026-09-24). This settles the launch date; Jul 16 is when the Token Price Index recorded the rows. (raw)
  • Jul 9 — [Analysis] Benedict Evans publishes "Ways to Think About Token Pricing": inference gross margin 40–50% excluding training; four open questions; the mobile-carrier analogy. (essay)
  • Jul 8 — [Preprint] Satoshi Matsuoka (RIKEN R-CCS) posts "Memory Scarcity, Open Models, and the Restructuring of the AI Industry, 2026–2030": cost in $/PB of bandwidth; the depreciation conveyor (incumbent cost gap 3.2× → 1.9× → 3–4×, modelled); scenario weights with commodity and oligopoly tied at 25%. (preprint)
  • Jul 8 — [Measured, sibling rung serving] DigitalOcean measures one H200 serving a 70B open model at $20.32 → $0.45 per million output tokens from batch 1 to 128. (raw)
  • Jul 8 — [Preprint, sibling rung harnesses] The Harness Effect: an orchestration swap cuts cost per task $0.21 → $0.12 at equal quality. (raw)

June

  • Jun 10 — [Preprint] Chitral Patil posts "Beyond Per-Token Pricing": $0.21–$15.25 per million output tokens on one H100 setup at $6.98/GPU-hr; 17.5–36.3× for one model from idle to full load; every surveyed calculator treats utilisation as fixed; releases vllm-cost-meter. (preprint)
  • Jun 10 — [Preprint] Quanyan Zhu posts "AI Tokenomics": tokens spent ≠ value produced. Theoretical, no measured figures. (preprint)
  • Jun 10 — [Preprint] Styles & Miller post "WorkBench Revisited": best-agent completion 43% → 98%, harmful actions 26% → 1.9%; open-weight models "drastically lowered" the cost of a once-proprietary performance level while frontier costs stayed flat (v2 Jul 1). (preprint)
  • Jun 9 — [Preprint] STAGE-Claw: a 40-task agent benchmark across 11 frontier models; per-task cost reportedly $0.17–$6.55 (search-surfaced, unconfirmed). (preprint)

May

  • May 25 — [Report, sibling rung serving] Epoch AI's serving-capacity model: capacity ~3.4×/yr vs demand ~10×/yr. Tokens per second, no dollars. (raw)
  • May 19 — [Preprint] "The Economics of AI Inference … Inference-Cost Phillips Curve" (Stockholm University): US slope 0.087 (s.e. 0.021) on 2022–2026 data; size term λ̄ = 0.18 stipulated. Filed speculative. (preprint)

April

  • Apr 24 — [Analysis] Digital Applied's Q2-2026 provider matrix: one model (Llama 4 70B), list prices spread ~6×, Together $0.65 batch to Cerebras $4.20. (analysis)
  • Apr 7 — [Report, sibling rung serving] vLLM/AMD: splitting prefill from decode on the same eight MI300X GPUs lifts SLO-meeting requests 26 → 73 per 100 — 2.5× goodput. (raw)

March

  • Mar 30 — [Preprint] Mingdeng Du posts "Tiered Super-Moore's Law": ~600× price decline, tier half-lives (economy 1.10 yr, mid 1.55 yr, flagship unfittable), a May-2024 break from technology- to competition-driven decline, ~103.7% of the cost fall attributed to productivity and −0.9% to GPU hardware. Price series ends Mar 28. (preprint)

February

  • Feb 23 — [Preprint] Litowitz, Polson & Sokolov post "Photons = Tokens": the token as a physical quantity; 326 TWh of projected 2028 US AI energy ≈ 6.5×10¹⁷ tokens a year ≈ 225,000 tokens per person per day. (preprint)

2025

April

  • Apr 17 — [Preprint] Erol, El, Suzgun, Yuksekgonul & Zou post "Cost-of-Pass" (revised 2026-02-26): cost per correct answer; the best available cost on hard maths-style tasks "roughly halved every few months". (preprint)

March

  • Mar 12 — [Data] Epoch AI publishes "LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks": median 50×/yr (9×–900×), GPT-4-level science milestone 40×/yr, post-2024 median 200×/yr. (Epoch AI)

Undated

  • Modular, LLM Inference Handbook — a living vendor reference with no publication date; read 2026-08-06 (landing page and metrics chapter). (raw)