One million output tokens costs $0.50 on the cheapest model in the comparison and $50 on the dearest. Same job, a hundred times the price — the 2026 price list, and what the numbers actually buy.
The invoice arrives monthly, and it is the only benchmark nobody argues with. A team running a support bot on GPT-6 Astra pays $50 per million output tokens. The team across the hall, on Gemini 3.8 Flash, pays $3.75. Both models answer the same customer, in the same second, with an answer the customer cannot tell apart. The $46.25 difference is not a quality gap. It is a procurement decision nobody made on purpose.
Published list prices per million tokens, as compiled on this site’s market-data page from provider pricing pages via October 2026 comparisons: at the top, GPT-6 Astra at $10 input and $50 output — the highest-cost flagship in the comparison. Claude Opus 5.5 at $4 and $20. Gemini 4 Argon at $2 and $10, an introductory rate that runs through November 30, 2026 and rises to $4 and $20 from December 1. In the middle, Claude Sonnet 5.5, GPT-6.1 Sol, and Kimi K3.1 Pro all at $2 and $10, with GPT-6.1 Terra at $1 and $5. At the floor, Gemini 3.8 Flash at $0.75 and $3.75, DeepSeek V4.1 Flash at about $0.15 and $0.60 off-peak, and Claude Haiku 5.5 at $0.10 and $0.50 — the lowest in the comparison. The spread from floor to ceiling on output tokens is a hundred to one.
One ratio runs through the whole list: output costs four to five times input at every model named — $4 against $20, $2 against $10, $10 against $50, $0.75 against $3.75, $0.10 against $0.50, $1 against $5. DeepSeek is the single exception at four times, $0.15 against $0.60. That ratio decides where the money goes. A prompt-heavy workload and an answer-heavy workload price nothing alike, and reasoning models that do their thinking in output tokens are the expensive case. The arithmetic is blunt: one million input tokens plus one million output tokens costs $24 on Claude Opus 5.5, $60 on GPT-6 Astra, and $0.60 on Claude Haiku 5.5 — the same token counts, a hundredfold apart. Any cost model that prices the prompt and ignores the answer is pricing the wrong half of the exchange.
The history behind the list is the price war this site covered in October. GPT-4 launched at $30 input and $60 output per million tokens in March 2023; DeepSeek V3 arrived at $0.14 and $0.28 in January 2025, and Nvidia lost $589 billion in market value in a single session on the news. By 2026 the cheapest models sold input tokens for a nickel per million. The list above is where that war has settled — for now.
Price alone says nothing about value, so set it against the scores. The Artificial Analysis Intelligence Index v5.0.1 (CritPt regrade, ~November 5, 2026) puts Gemini 4 Carbon at 60.0, the index leader; Claude Opus 5.5 at 59.1; GPT-6 Astra at 56.9; GPT-6.1 Sol at 54.6; Claude Sonnet 5.5 at 54.5; GPT-6.1 Terra at 51.3; Gemini 4 Argon at 50.4; Kimi K3.1 Pro at 46.5; and DeepSeek V4.1 Flash at 32.6, the top open-weights model. Divide the score by the output price and the ranking inverts: DeepSeek returns 54.3 index points per output dollar, Terra 10.3, Sonnet 5.5 and Sol 5.5, Argon 5.0 at its introductory price, Kimi 4.7, Opus 5.5 3.0, and Astra 1.1. The gap between Astra and Terra is 5.6 index points against a ninefold difference in intelligence per dollar. The dearest model is not nine times smarter. It is nine times dearer per point.
One gap in the table is worth naming. Gemini 4 Carbon leads the index at 60.0, but no published price row exists for it in the comparison — the smartest model on the board cannot be scored per dollar from published prices. Any ranking that claims a winner without a price is ranking half the question.
List prices are the sticker, and three mechanics move the real number. First, the same weights rent at wildly different prices depending on who serves them: Artificial Analysis data shows Llama 3.1 70B ranging from $0.20 to $2.90 per million tokens across providers, with the median near $0.70, and the 405-billion-parameter version from $0.90 to $9.50. Second, caching, batch APIs, and routing cut effective costs — prompt caching alone cuts effective input costs 50 to 90 percent. Third, the enterprise markup: the same DeepSeek R1 that lists at $0.55 input and $2.19 output direct costs $1.35 and $5.40 through Azure OpenAI and Amazon Bedrock, a two- to fivefold premium for compliance, uptime, and a procurement department’s signature. Throughput is the other half of the cost question, and it is measured, not listed: MLPerf Inference v6.1, published September 16, 2026, drew 30 submitting organizations and 486 results, with NVIDIA’s Vera Rubin NVL72 up to 3.7 times the GB300 NVL72 on Qwen3-VL. The prior round had the GB300 NVL72 serving DeepSeek-R1 at 8,064 tokens per second per GPU. MLPerf Endpoints replaces MLPerf Inference for datacenter use in 2027, which means the measurement itself is about to change hands.
Cheapest per token is not cheapest per answer, and the distinction matters more as reasoning models spread. A model that answers in 200 output tokens at $50 per million costs a cent; a model that needs 5,000 output tokens at $3.75 per million costs nearly two cents for the same answer — the expensive token wins the cheap question. Verbose models are expensive models, whatever the price list says. Prices also move weekly: this site’s market-data page refreshes on a weekly cadence because a flagship that led an index in September can be repriced, discounted, or undercut by October, and Gemini 4 Argon’s rise on December 1 is already scheduled. List prices exclude negotiated enterprise discounts, which providers do not publish. The price list is a snapshot, and a snapshot of a market that reprices itself every week is a perishable document.
The implication is practical, not philosophical. Measure cost per completed task, not per token. Build the routing, caching, and measurement that let you switch models when the list moves, because it will move. And read the price list the way a buyer reads any menu: the number that matters is the one on the invoice, and the invoice counts output tokens first.
Sources: Published list prices per million tokens (input / output) and Intelligence Index v5.0.1 scores: inferencebenchmarks.com market-data page, November 9, 2026 (scores: Artificial Analysis Intelligence Index v5.0.1, CritPt regrade ~November 5, 2026; prices: published pricing via Tech Insider, TrendInTech, and Techietory October 2026 comparisons; Gemini 4 Argon introductory pricing: Google Cloud pricing announcement). Score-per-output-dollar figures: arithmetic from the published scores and list prices above. Price-war history — GPT-4 at $30/$60 (March 2023), DeepSeek V3 at $0.14/$0.28, Nvidia $589 billion single-day loss (January 27, 2025), Llama 3.1 provider spreads ($0.20–$2.90 for 70B; $0.90–$9.50 for 405B), caching savings of 50–90%, DeepSeek R1 direct vs Azure/Bedrock pricing ($0.55/$2.19 vs $1.35/$5.40): “The Price of a Token,” InferenceBenchmarks.com Insights, October 2026. MLPerf Inference v6.1 (30 organizations, 486 results, published September 16, 2026; Vera Rubin NVL72 up to 3.7x; MLPerf Endpoints replacing Inference for datacenter in 2027): MLCommons; Nebius v6.1 results. MLPerf Inference v6.0 GB300 NVL72 DeepSeek-R1 server figure (8,064 tokens/second/GPU): NVIDIA v6.0 results, as reported via this site’s market-data page.
