Last updated: October 31, 2026 · Refresh cadence: WEEKLY · Trend tracking starts with this edition except where a prior-period figure is shown.
AI Inference Costs & Benchmark Scores — Market Data
Inference pricing now moves faster than the benchmarks that are supposed to justify it. A flagship that led an index in September can be repriced, discounted, or undercut by October, and the cheapest model on a leaderboard is rarely the cheapest model on an invoice once output tokens are counted.
The tables below keep the two questions separate: what models score, and what they charge. The score table uses the Artificial Analysis Intelligence Index as reported in public summaries; the price table uses published per-million-token list prices. Neither table, on its own, answers which model to run — that answer lives in the ratio between them.
Benchmark scores: Artificial Analysis Intelligence Index
Index version changed from v4.3.2 to v5.0 between editions; scores below are v5.0 and are not directly comparable to the prior edition's scores.
| Figure | Value | Trend vs prior period | Source |
|---|---|---|---|
| Claude Opus 5.5 (Anthropic) | 58.4 | Prior (v4.3.2, October 2026): 57.6 ▲ Up on the new v5.0 scale; index version changed, not a like-for-like move | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| Claude Sonnet 5.5 (Anthropic) | 54.2 | Prior (v4.3.2, October 2026): 56.0 ▼ Down on the new v5.0 scale; index version changed, not a like-for-like move | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| GPT-6 Astra (OpenAI) | 56.9 | Prior (v4.3.2, October 2026): 52.7 ▲ Up on the new v5.0 scale; index version changed, not a like-for-like move | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| Gemini 4 Argon (Google) | 50.4 | Prior (v4.3.2, October 2026): 52.6 ▼ Down on the new v5.0 scale; index version changed, not a like-for-like move | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| GPT-6.1 Sol (OpenAI) | 54.5 | Prior (v4.3.2, October 2026): 51.8 ▲ Up on the new v5.0 scale; index version changed, not a like-for-like move | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| Gemini 4 Carbon (Google) | 59 | Index leader ● Baseline (v5.0) | Artificial Analysis Intelligence Index v5.0 (FAQ) |
| GPT-6.1 Terra (OpenAI) | 51.0 | No prior-period figure from the same source ● Baseline (v5.0) | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| Kimi K3.1 Pro (Moonshot AI) | 46.0 | No prior-period figure from the same source ● Baseline (v5.0) | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| DeepSeek V4.1 Flash (open weights) | 32.5 | Top open-weights model ● Baseline (v5.0) | Artificial Analysis Intelligence Index v5.0, as reported via Modelgrep leaderboard |
| Terminal-Bench 4.0 — Claude Opus 5.5 | 64.85% (agentic command-line tasks, October 2026) | Prior staged value 64.8% ● Flat (rounding only) | Terminal-Bench 4.0 public leaderboard, Oct 28, 2026 snapshot, as mirrored via BenchLM |
| Terminal-Bench 4.0 — Claude Sonnet 5.5 | 61.82% | Prior staged value 61.8% ● Flat (rounding only) | Terminal-Bench 4.0 public leaderboard, Oct 28, 2026 snapshot, as mirrored via BenchLM |
| Terminal-Bench 4.0 — GPT-6 Astra | 58.18% | Prior staged value 58.2% ● Flat (rounding only) | Terminal-Bench 4.0 public leaderboard, Oct 28, 2026 snapshot, as mirrored via BenchLM |
Published list prices per 1 million tokens (input / output)
| Figure | Value | Trend vs prior period | Source |
|---|---|---|---|
| Claude Opus 5.5 | $4 / $20 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via Tech Insider leaderboard comparison |
| Claude Sonnet 5.5 | $2 / $10 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via Tech Insider leaderboard comparison |
| GPT-6 Astra | $10 / $50 — highest-cost flagship in the comparison | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via Tech Insider leaderboard comparison |
| GPT-6.1 Sol | $2 / $10 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via TrendInTech comparison |
| Gemini 4 Argon | $2 / $10 introductory through November 30, 2026, then $4 / $20 | Scheduled rise from December 1, 2026 ▲ Rising (scheduled) | Google Cloud pricing announcement, via Tech Insider |
| Gemini 3.8 Flash | $0.75 / $3.75 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via TrendInTech comparison |
| Claude Haiku 5.5 | $0.10 / $0.50 — lowest in the comparison | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via TrendInTech comparison |
| DeepSeek V4.1 Flash | About $0.15 / $0.60 off-peak; open weights | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via Techietory October 2026 comparison |
| GPT-6.1 Terra | $1 / $5 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via October 2026 comparison |
| Kimi K3.1 Pro | $2 / $10 | No prior-period figure from the same source ● Baseline (October 2026) | Published pricing, via comparison |
MLPerf Inference: measured throughput, not list prices
| Figure | Value | Trend vs prior period | Source |
|---|---|---|---|
| MLPerf Inference v6.0 — NVIDIA Blackwell Ultra, DeepSeek-R1 (offline) | 9,821 tokens/second/GPU | Prior (v5.x): 5,842 ▲ Up (across benchmark rounds) | NVIDIA MLPerf Inference v6.0 results, as reported via this coverage |
| MLPerf Inference v6.0 — NVIDIA Blackwell Ultra, Llama 3.1 405B (offline) | 271 tokens/second/GPU | Prior (v5.x): 224 ▲ Up (across benchmark rounds) | NVIDIA MLPerf Inference v6.0 results, as reported via this coverage |
| MLPerf Inference v6.0 — GB300 NVL72, DeepSeek-R1 (server, per GPU) | 8,064 tokens/second/GPU | Prior-round (v5.1): 2,907 ▲ Up | NVIDIA MLPerf Inference v6.0 results, as reported via this coverage |
| MLPerf Inference v6.0 — GB300 NVL72, Llama 3.1 405B (server, per GPU) | 259 tokens/second/GPU | Prior-round (v5.1): 170 ▲ Up | NVIDIA MLPerf Inference v6.0 results, as reported via this coverage |
| MLPerf Inference v6.0 — GB300 NVL72 system, DeepSeek-R1 (server) | 575,580 tokens/second (72-GPU system, Nebius submission) | New benchmark round; not comparable to per-GPU v5.x figures ● Baseline (October 2026) | Nebius MLPerf Inference v6.0 results summary |
| MLPerf Inference v6.0 — software-only gain claimed | 2.77x throughput gain (server; 1.68x offline) vs v5.1 on unchanged GB300-NVL72 hardware | Prior: performance six months earlier, same hardware ▲ Up (vendor claim) | NVIDIA MLPerf Inference v6.0 results, as reported via this coverage |
Methodology
Included: the Artificial Analysis Intelligence Index as republished in public rankings, published per-token list prices as summarized from provider pricing pages, and MLPerf Inference results as published by submitters. Excluded: negotiated enterprise discounts, which providers do not publish, and any single-vendor benchmark that is not an MLPerf or Artificial Analysis result. Per-GPU MLPerf figures are derived by submitters by dividing system throughput by accelerator count and are labelled as such in the source; system-level v6.0 figures are not directly comparable to per-GPU v5.x figures. v6.0 per-GPU figures are compared across benchmark rounds, not within one round.
Trend labels: “Baseline (October 2026) — trend tracking starts with the next refresh” means no prior-period figure from the same source was available this session, so no movement is claimed. Where a prior figure is shown, the trend compares the current figure with that prior figure from the same source.
Disclaimer
Figures on this page are aggregated public estimates compiled for information only. They are not investment advice, professional advice, or an offer to buy or sell any security or asset. Figures change as sources revise or republish them; verify any figure against its linked source before relying on it.
This domain is offered by Prime Domain Capital. Contact only at info@primedomaincapital.com.
Changelog
- October 2026: First edition, published as a staged page. Figures compiled from the public sources listed in SOURCES.md, with trend indicators and baseline labels as shown above. Next refresh follows the WEEKLY cadence in SOURCES.md; the sitemap lastmod for this page is to be updated on every refresh once deployed.
- October 31, 2026: Artificial Analysis scores updated to Intelligence Index v5.0 (version change from v4.3.2 stated on-page); four new score rows added (Gemini 4 Carbon, GPT-6.1 Terra, Kimi K3.1 Pro, DeepSeek V4.1 Flash); Terminal-Bench 4.0 rows confirmed against the Oct 28, 2026 leaderboard snapshot; two new price rows added (Terra, Kimi); all staged list prices reconfirmed, including Gemini 4 Argon introductory pricing through Nov 30, 2026; MLPerf rows updated to v6.0 per-GPU figures. No staged row dropped.