State of LLM Pricing — November 2026

1. Frontier intelligence cost $5.14 per million tokens on November 1, 2026, the same as on October 1. The LLM Price Index is the equal-weight blended price across 10 current flagship models, one per major lab. It held the same ten models at the same ten list prices for the whole of October. That makes October the first full calendar month without a move since the index's daily as-published record began on August 1. By November 1 the level had been unchanged for 32 days, the longest stretch on that record. It stood 12.5% above its first reading of February 23.
2. Two labs launched flagships cheaper than the models holding their seats, and neither entered the index. Google's Gemini 4 Argon (October 13) and Moonshot's Kimi K3.1 Pro (October 23) both went on sale at $4.00 blended. The models in those seats cost $4.50 (Gemini 3.1 Pro) and $6.00 (Kimi K3). Each seat follows a lab's flagship line by name, and both models launched under new names. With both seated, the November 1 index would have read $4.89, 4.9% lower.
3. Every generally available successor launched cheaper or at the same price. Claude Haiku 5.5 launched 90% below Claude Haiku 4.5, GPT-6.1 Terra 56% below GPT-5.6 Terra, and Kimi K3.1 Pro 33% below Kimi K3. Two previews were priced above the lines they extend: Mistral Large 4, 40% above Large 3, and Meta's Muse Wonder, at four times Muse Spark 1.3.
4. Gemini 4 Argon is cheap only until November 30. Its $2/$10 is an introductory price. From December 1 Google's rate card prints $4/$20, or $8.00 blended, 78% above Gemini 3.1 Pro. At that price, with Kimi K3.1 Pro also seated, the index would read $5.29 rather than $4.89.
5. GPT-4-class capability still cost $0.090/Mtok, 57.1x below the flagship ceiling. The floor did not move in October. None of the month's launches had an externally measured GPQA Diamond score, so none could challenge DeepSeek V4 Flash on DeepInfra.

The index entered October at $5.14/Mtok on the ten-lab basket and left it at $5.14. It did not move once:

Google and Moonshot both launched new flagships in October, generally available and below the price of the model in their seat:

For the second month running, no generally available successor from the index's labs launched above the model it followed. The cuts were deepest below the top tier:

The cheapest model with an externally measured GPQA Diamond score of 70 or better, roughly GPT-4-class general reasoning, cost $0.090/Mtok on November 1, exactly as on October 1. It is still DeepSeek V4 Flash on DeepInfra, a host's copy of a model DeepSeek stopped selling on its own API in September. Ceiling ÷ floor held at 57.1x.

No October release could challenge it. Every new model needs an externally measured GPQA Diamond score to count, and none of the month's launches had one by November 1. Particle AI's Particle 1 publishes a score of 90.4, but it is the company's own, so it does not count. The fallbacks did not change either. If DeepInfra withdrew its listing, the floor would move to GPT-OSS 20B on Fireworks at $0.128. The cheapest qualifying model a lab sells itself is still GPT-5.6 Luna at $0.45. One October retirement closes a chapter of the series: o4-mini, which set the first floor of the record at $1.93 on March 1, was among the models OpenAI shut down on October 23. Full series: the Intelligence Cost Curve.

On November 1 the ten constituents averaged $0.0676 per point of measured intelligence, the same as on October 1. The models, prices and scores were all the same, with all ten scored. Mistral Large 3 remained the best value at $0.0140 per point, and GPT-6 Astra the costliest at $0.2356. Treat these as a ranking with wide error bars, not a measurement. Each score averages whichever benchmarks a model has been measured on, and the sets differ.

Low launch prices came with end dates. Gemini 4 Argon's $2/$10 runs through November 30. On October 28 Google announced Gemini 4 Carbon at the same introductory price, with the same $4/$20 to follow. Google is making Carbon available only to a subset of cyber defenders and trusted partners, and it was on none of Google's rate cards. The tracker does not list it, because nobody can buy it at that price today. Gemini 3.6 Flash and 3.8 Flash keep their introductory $0.75/$3.75 through December 31, and both are scheduled to double on January 1, 2027. Mistral Large 4 had three prices in its first month on our records. It launched on October 6 at $0.68/$2.09. On October 12 and 19 that price showed beside a struck-through $1.36/$4.18, and on October 26 $1.36/$4.18 stood alone. From October 29 it was $0.70/$2.10, with nothing struck through.

OpenAI closed its long tail. On October 23 OpenAI shut down eight of the models we track: GPT-3.5 Turbo, GPT-4 (0613), GPT-4 Turbo, GPT-4.1 nano, o1, o1-pro, o3-mini and o4-mini. On October 14 it gave notice that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on January 14, 2027. Others cleared out too. On October 10 Alibaba retired five of the models we track, among them Qwen3-Max. Google deprecated Gemini 3.5 Flash and 3.7 Flash on October 8 and now routes their calls to 3.6 Flash and 3.8 Flash. Upstage retired Solar Pro 3 on October 29. On October 31 Mistral deprecated Mistral Large 3 and Medium 3.5, naming Large 4 as their replacement, and set their retirement for April 30, 2027.

A deprecation is not a price change, but it can work like one, because a team pinned to a retiring model has to move to whatever its replacement costs. OpenAI's own suggested replacements cut the bill for its old large models and raised it for its small ones. GPT-4, GPT-4 Turbo, o1 and o1-pro users are pointed to GPT-5.6 Sol, which costs a fraction of their price at its promotional $4/$20. GPT-3.5 Turbo ($0.75 blended) and o4-mini ($1.93) are pointed to GPT-5.6 Terra ($4.50), six and 2.3 times as much. GPT-4.1 nano ($0.175) is pointed to GPT-5.6 Luna ($0.45), 2.6 times as much.

Cache reads at 5% are spreading. Until September no Anthropic or OpenAI model priced cache reads below a tenth of its input rate. October's launches made 5% common at the top. Gemini 4 Argon ($0.10 against $2.00) and GPT-6.1 Terra ($0.05 against $1.00) launched there, at the level Claude Opus 5.5 and GPT-6.1 Sol set in September. Muse Wonder's preview price is also 5%. Claude Haiku 5.5 and Kimi K3.1 Pro launched at 10%. By October 7 Anthropic had halved Claude Sonnet 5.5's cache reads to 5% ($0.20 → $0.10).

Outside new launches, few list prices moved on dates we can place in October. The clearest was a promotion ending: Upstage's Solar Mini 4 doubled to $0.10/$0.40 on October 23. The tracker grew to 303 model rows across 41 providers, with Cloudflare, StepFun, Moondream, Apodex, ScaleDown and Particle AI new to it. But the number of models listed as on sale fell from 220 to 209, as more models left sale than arrived.

This edition is fixed at November 1. The index held $5.14 with the same ten constituents through November 6. On November 7 Gemini 4 Argon took Google's seat and Kimi K3.1 Pro took Moonshot's by dated rebalance, which moved the index to $4.89 (methodology). Outside the basket:

• November 2: Meituan released LongCat-2.5 at $0.30/$1.20 ($0.525 blended). That is a limited-time price and the only one Meituan prints. LongCat-2 left pay-as-you-go the same day.
• November 2: Mistral Large 3 and Medium 3.5 were off Mistral's API price list. Mistral's own endpoint on OpenRouter still sells Large 3 at $0.50/$1.50.
• November 4: Tencent released Hy4 at $0.417/$1.2505, half its list price until December 31 (Beijing time).
• November 5: Thinking Machines released Inkling 2 Small at $0.30/$1.00 ($0.475), 10% below Inkling Small.
• November 6: Sakana announced that Namazu rises from $0.95/$4.00 to $1.90/$6.00 on December 1, a 71% increase blended.

Already on the calendar: Claude Sonnet 4.5 retires on November 30. Gemini 4 Argon, now in Google's seat, moves to $4/$20 on December 1. OpenAI shuts down GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3 and o3-pro on December 11. Gemini 3.6 Flash and 3.8 Flash double on January 1, 2027, and the GPT-4o and GPT-4.1 models above shut down on January 14.

Every figure above is computed from the ModelPriceWatch tracker as it stood on November 1, 2026: 303 model rows (209 then listed as on sale) across 41 providers, and 23,869 price snapshots from February 2024 to November 1, 2026. Each price is traced to the provider's official pricing page. The full dataset is public: