Diminishing Returns at the Frontier

1 · The mid-2026 model landscape

Since the February snapshot, OpenAI shipped GPT-5.5 (2026-04-23) and Anthropic shipped Claude Fable 5 + Mythos 5 (2026-06-09). The leading open-weight models are all large sparse Mixture-of-Experts designs — trillion-plus total parameters, only 32–49B active per token.

Frontier (closed-weight)

Open-weight

2 · The aggregate spread is narrow

On the Artificial Analysis Intelligence Index, Claude Fable 5 (65) leads Claude Opus 4.8 (61) and GPT-5.5 (60) by single digits; the best open-weight models trail by only 3–6 points. Paying for #1 buys a small edge over a crowded field.

3 · …but the gap is real on the hardest tasks

Convergence is benchmark-dependent. On the two hardest evaluations the best open-weight models still trail the closed frontier by 8–18 points — and on agentic work Grok 4.3 trails GPT-5.5 by 276 Elo on GDPval-AA.

4 · The cost of intelligence — real, but method-dependent

The cost to reach a fixed capability has fallen sharply, but how much depends entirely on the measurement method. The famous “200×/year” is a token-price figure; a stricter benchmark-cost measure that corrects for reasoning models emitting more tokens gives 5–10×/year, and ~3×/year once hardware and competition are removed. All three are correct.

Concrete data point: by Grok 4 Fast (Sept 2025), an Index>60-tier model cost just $0.20 / $0.50 per million tokens (in/out) [13]. As the MIT authors put it, “much of the more dramatic 100×–1000× price decreases… are likely due to non-technical factors like increased market competition” [11].

5 · KV-cache economics, corrected

The formula is 2 × n_layers × n_kv_heads × head_dim × bytes × tokens. The February doc used the 32-layer (8B) result but labeled it “70B.” Re-derived for a real 70B (80 layers, FP16):

One H200 (141 GB) serving a Q4 70B (~40 GB weights, ~95 GB free for KV) drops from the doc's claimed 270 concurrent users at 4K to a real ~76, to ~2 at 128K, to under one at 1M. You don't escape this with bigger GPUs — you redesign attention: GQA (≈4× vs MHA), MLA (≈5× smaller), and GTA/GLA which match GQA/MLA quality at roughly half the KV cache [14].

6 · What changed since February 2026

7 · Methodology & verification

The web data was gathered by a deep-research harness: parallel search fan-out → source fetch → claim extraction → 3-vote adversarial verification (a claim needs 2 of 3 “refute” votes to be killed) → synthesis. 6 angles · 27 sources · 116 claims extracted · 25 verified · 22 confirmed, 3 killed · 110 agents. KV-cache arithmetic was additionally re-derived by hand.

Confirmed findings

Refuted — killed in verification, NOT used

Caveats

• Never quote a cost-decline number without its method (token-price 50×/200× vs benchmark-cost 5–10× vs algorithmic ~3×).
• NVIDIA H200/B200/B300 VRAM & bandwidth specs did not survive verification; the H200 = 141 GB figure is vendor-published, used as-is.
• Model identities are a June-2026 snapshot; thesis-level conclusions hold regardless of specific versions.

Open questions

• Confirmed NVIDIA H200 / B200 / B300 VRAM + HBM bandwidth.
• The actual published GPT-5.5 API price (existence confirmed; pricing refuted).
• Whether the fastest ~900×/yr price trends persist or revert toward the 50× median.
• How much of the 100–1000× market price decline is competition vs algorithmic progress.

Sources (27 fetched, by quality)

Full appendix with every claim, evidence quote, and vote: frontier-diminishing-returns-2026-research-appendix.md.