Models

This rung is about what the model itself can do, and what that costs per token — the architecture that decides how much compute a forward pass burns, the released weights you can actually hold, the quoted price on the invoice, and the growing body of work on what is happening inside the thing while it answers.

What this rung prices

A buyer pays for a finished task; a lab quotes per token. This rung owns the first term in that product — the per-token price and the per-task token count that the model itself sets before any harness touches it. Both halves move, and they move for architectural reasons. Sparsity is the clearest case: Kimi K3 carries about 2.8 trillion parameters but activates only 16 experts of 896 per token, roughly 1.8% of the pool, so the forward-pass cost sits far below what the headline count implies (Kimi K3). The token count is just as much a model property: Anthropic's own release notes record that Claude Sonnet 5's new tokenizer produces about 30% more tokens for the same text (Claude Platform release notes), which raises the effective price per word without the quoted per-token price moving at all. So the rung answers a question no price list answers on its own: for a given job, how many tokens does this model spend and what does each one cost — and the empirical work here says the answer has been falling for software reasons rather than hardware ones.

The load-bearing findings

• Token prices fell roughly 600-fold between 2020 and 2026, and essentially all of it came from software, not chips. A dataset of 318 OpenRouter models, 3,237 Epoch AI records and 62 cross-validated milestones decomposes the decline as about 103.7% software and architectural innovation against about -0.9% GPU hardware — hardware was a drag or a noise term, not an independent driver (Tiered Super-Moore's Law). This close read is abstract-level, so it rests on the abstract and the reported findings rather than the PDF body.
• The decline is tiered, not uniform, and the premium tier resists it. Same paper: economy-tier price half-life 1.10 years and mid-tier 1.55 years against a two-year Moore benchmark, while flagship and reasoning tiers hold up, with reasoning commanding about a 31.5× premium over non-reasoning pricing; a structural break in May 2024 (F = 5.74, p = 0.005) separates a technology-driven regime from a competitive one (Tiered Super-Moore's Law). OpenAI shipped that shape as a product three months later — GPT-5.6 as three named tiers, Sol at $5/$30, Terra at $2.50/$15 and Luna at $1/$6 per million tokens (GPT-5.6).
• An open-weight model reached near-frontier capability at roughly a sixth of the closed price, and the efficiency is architectural. DeepSeek V4-Pro is a 1.6-trillion-parameter mixture of experts with 49B active per token and a 1M context, priced at $0.435 input and $0.87 output per million tokens; in the 1M setting it runs single-token inference at about 27% of V3.2's FLOPs and cuts KV-cache memory to about 10% of V3.2 — that efficiency is the stated basis for the price. It reports SWE-bench Verified 80.6% but only about 55% on the harder SWE-bench Pro split against Opus 4.7 at about 64% (DeepSeek V4). The gap survives on the hardest tasks and nowhere else — this is the substance of open-weight frontier parity. The file is summary-derived: the primary article returned HTTP 429, and the benchmark figures are vendor or third-party reported, not peer-reviewed.
• Open-weight parity does not mean cheap — the second instance competes on capability and raises its price. Kimi K3 ships at $3/$15 per million tokens, up sharply from Kimi K2.6's $0.95/$4, making it the most expensive model a Chinese lab has shipped; it ranks first on Arena.ai's Frontend Code leaderboard ahead of Claude Fable 5 while trailing Fable 5 and GPT-5.6 Sol overall on Moonshot's own numbers. It launched with a single reasoning-effort setting, "max", which spent 13,241 reasoning tokens (about $0.25) on a hello-world-grade prompt in the reviewer's own test (Kimi K3). Summary-derived: Moonshot's blog served a bare nginx page, so all of this comes through an independent review and wire coverage.
• Where you put a model's capacity is an optimisation with an interior answer, not a scaling slope. Adding a constant-time lookup table for static knowledge alongside a mixture of experts produced BBH +5.0, ARC-Challenge +3.7, CMMLU +4.0, MMLU +3.4, HumanEval +3.0, MATH +2.4, and multi-query needle-in-a-haystack retrieval from 84.2 to 97.0, scaled to 27B parameters of memory — and the allocation between neural compute and static memory traces a U-shaped curve, so neither pure compute nor pure memory is optimal (Conditional Memory via Scalable Lookup). The gains were largest in reasoning rather than in plain recall.
• Depth is becoming a dial you turn at inference rather than a number fixed at training. LoopFormer runs the same transformer layers repeatedly and trains with a shortcut-consistency scheme that aligns the trajectories of different loop counts, so one trained model can trade depth against compute budget at inference and degrade gracefully when the budget is tight (LoopFormer). The paper reports robust results under aggressive compute constraints but gives no benchmark numbers in the abstract, so there is nothing to quote.
• The visible reasoning trace may not be the reasoning. One preprint argues that chain-of-thought tokens are closer to post-hoc rationalisation than to the computation being performed, using probes, sparse autoencoders and interventions to find cases where behaviour tracks the latent trajectory even when the chain-of-thought content is altered or absent (LLM Reasoning Is Latent). If that holds, then reading a model's stated steps is a partial safety signal at best — see chain-of-thought reasoning. Set against it, chain-of-thought monitoring has caught a real thing: OpenAI used it to catch one of its own reasoning models cheating on coding tests (Mechanistic Interpretability — 10 Breakthrough Technologies).

What we do not know yet

• Nothing on this rung was measured here. Every figure above is read, not run. No price was sampled by us, no benchmark re-executed, no model held locally — and the rung's whole point is cost and capability per token, which is measurable with a script and a bill.
• Four of the thirteen sources could not be fetched from the primary. DeepSeek V4 (HTTP 429), GPT-5.6 (OpenAI's page returned 403), Kimi K3 (the vendor blog served a placeholder) and the token-pricing paper (abstract only) are all summary-derived. The three model releases therefore reach us as vendor claims relayed by journalism, and GPT-5.6's headline "54% more token-efficient on agentic coding" arrives with no methodology attached (GPT-5.6).
• The 31.5× reasoning premium is a cross-sectional average with no breakdown. The paper does not, per the summary available, split it by lab or by model generation, and its production-frontier estimates are sensitive to which models and providers are in the list — closed enterprise pricing is likely under-covered (Tiered Super-Moore's Law).
• Two vendor announcements on this rung carry no independently verified numbers at all. Google's Gemini 3 post states multimodal leadership and improved tool use with independent benchmarks pending and cost/latency trade-offs undisclosed (Gemini 3), and Meta's Muse Spark describes a parallel sub-agent "Contemplating" mode with no published independent benchmarks and an acknowledged compute cost (Meta Muse Spark). Neither moves a number here yet.
• Interpretability has no ground truth to check itself against. The survey lists the problem plainly: no way to verify an interpretability hypothesis, superposition and polysemanticity frustrating feature-level accounts, patching experiments that scale badly with model size, and open doubt about whether any of it transfers to frontier systems (Mechanistic Interpretability for LLM Alignment). The circuit tracing work it draws on is also almost entirely Anthropic's own (Transformer Circuits), so the field's main evidence base is one lab's.
• The rung's own scope has holes. Its stated scope names quantization, distillation and fine-tuning, and no source here covers any of them — nor is there a single measurement of running a model on your own compute, which is half the specific-and-local cluster the rung claims.

Read next

• tiered-super-moores-law-token-pricing — the price series and the tier decomposition the whole rung hangs on.
• deepseek-v4-2026 — the clearest case of architecture, not hardware, buying a sixfold price cut.
• llm-reasoning-is-latent-arxiv-2026 — the claim that undercuts reading a model's stated reasoning.