# LLM API Pricing Comparison 2026: Every Major Provider per Million Tokens, Cached Input, and What a 40-Turn Conversation Costs

> LLM API pricing compared as of 30 October 2026: the flagship and cheap model of OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral, Moonshot, Alibaba, Zhipu, Groq, Fireworks and Together per million tokens, with cached input and cache-write rates, the three list prices that repeat across providers, the Gemini rises on 1 December and 1 January, and what the same 40-turn conversation costs on every model, from $0.016 on Qwen3.7-Flash to $5.46 on GPT-6 Astra and Claude Fable 5.1.

_Gaurav Dadhich · 2026-10-30_

\# LLM API Pricing Comparison 2026: Every Major Provider per Million Tokens, Cached Input, and What a 40-Turn Conversation Costs \*Published 30 October 2026. Every price below was read on 30 October 2026 from the provider's own pricing page or documentation, linked beside each table, and every price is as of that date.\* On 30 October 2026, the flagship models of the major LLM APIs cost between $0.66 and $10 per million input tokens and between $1.98 and $50 per million output tokens, and the cheapest model each provider sells costs between $0.03 and $0.25 for input and between $0.10 and $1.50 for output. \[OpenAI's GPT-6 Astra\](https://developers.openai.com/api/docs/pricing) and \[Anthropic's Claude Fable 5.1\](https://platform.claude.com/docs/en/about-claude/pricing) sit at the top at $10 input and $50 output. \[Alibaba's Qwen3.7-Flash\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) sits at the bottom at $0.03 and $0.13 for requests up to 32,000 tokens. Between them, three list prices keep repeating across companies that do not share a cap table: $2 and $10 (Claude Sonnet 5.5, GPT-6.1 Sol, Kimi K3.1 Pro, and Gemini 4 Argon until 30 November), $0.10 and $0.50 (Claude Haiku 5.5 and GPT-6 Luna), and $0.15 and $0.60 (DeepSeek's \`deepseek-flash\` off-peak, Mistral Small 4, and GPT-OSS 120B on Groq, Fireworks and Together). When two models list at the same price, the cache decides which one is cheaper, because a chat or agent request resends its whole history and most of what you pay for on a long conversation is cached input. A cache hit costs 2% of the input price on \`deepseek-flash\`, 2.5% on Claude Fable 5.1, 5% on Claude Sonnet 5.5 and GPT-6.1 Sol, 10% on most other models, 18.6% on GLM-5.3 and 50% on Groq. In the worked example below, the same 40-turn conversation costs $5.46 on both GPT-6 Astra and Claude Fable 5.1 with no caching, and $1.43 against $1.08 with a warm cache, purely because Astra reads its cache at four times Fable's rate. Each provider has its own page on this site with every model, every multiplier and the method behind its numbers: \[Claude\](https://www.maximem.ai/blog/claude-api-pricing-2026), \[Gemini\](https://www.maximem.ai/blog/gemini-api-pricing-2026), \[DeepSeek\](https://www.maximem.ai/blog/deepseek-api-pricing-2026), \[Mistral\](https://www.maximem.ai/blog/mistral-api-pricing-2026), \[Kimi K3\](https://www.maximem.ai/blog/kimi-k3-api-pricing-2026), \[Qwen\](https://www.maximem.ai/blog/qwen-api-pricing-2026), \[GLM\](https://www.maximem.ai/blog/glm-api-pricing-2026), \[Groq\](https://www.maximem.ai/blog/groq-api-pricing-2026), \[Fireworks vs Together\](https://www.maximem.ai/blog/fireworks-vs-together-ai-pricing-2026) and \[OpenRouter\](https://www.maximem.ai/blog/openrouter-pricing-2026). This page puts the flagship and the cheap model of each on one table, prices one conversation on every row with the same token assumptions those pages use, and keeps the numbers identical across all of them. ## LLM API price per million tokens: flagship and cheap model of every major provider | Provider | Model | Role | Input | Cached input | Cache write | Output | |---|---|---|---|---|---|---| | OpenAI | GPT-6 Astra (\`gpt-6-astra\`) | Flagship | $10.00 | $1.00 | $12.50 (30 min) | $50.00 | | OpenAI | GPT-6.1 Sol (\`gpt-6.1-sol\`) | Mid | $2.00 | $0.10 | $2.50 (30 min) | $10.00 | | OpenAI | GPT-6 Luna (\`gpt-6-luna\`) | Cheap | $0.10 | $0.01 | $0.125 (30 min) | $0.50 | | Anthropic | Claude Fable 5.1 | Flagship | $10.00 | $0.25 | $12.50 (5 min) | $50.00 | | Anthropic | Claude Opus 5.5 | Mid | $4.00 | $0.20 | $5.00 (5 min) | $20.00 | | Anthropic | Claude Sonnet 5.5 | Mid | $2.00 | $0.10 | $2.50 (5 min) | $10.00 | | Anthropic | Claude Haiku 5.5, prompts up to 100K | Cheap | $0.10 | $0.01 | $0.125 (5 min) | $0.50 | | Google | Gemini 4 Argon, to 30 Nov 2026 | Flagship | $2.00 | $0.10 | none (implicit only) | $10.00 | | Google | Gemini 4 Argon, from 1 Dec 2026 | Flagship | $4.00 | $0.20 | none (implicit only) | $20.00 | | Google | Gemini 3.8 Flash, to 31 Dec 2026 | Cheap | $0.75 | $0.075 | none (implicit) | $3.75 | | Google | Gemini 3.8 Flash, from 1 Jan 2027 | Cheap | $1.50 | $0.15 | none (implicit) | $7.50 | | Google | Gemini 3.1 Flash-Lite | Cheapest | $0.25 | $0.025 | none | $1.50 | | xAI | Grok 4.7 | Flagship | $2.00 | not listed | not listed | $6.00 | | DeepSeek | \`deepseek-v4-pro\`, off-peak / peak | Flagship | $0.66 / $1.32 | $0.022 / $0.044 | none | $1.98 / $3.96 | | DeepSeek | \`deepseek-flash\`, off-peak / peak | Cheap | $0.15 / $0.30 | $0.003 / $0.006 | none | $0.60 / $1.20 | | Mistral | Mistral Large 4 (Public Preview) | Flagship | $0.70 | $0.07 | none | $2.10 | | Mistral | Mistral Medium 3.5 | Highest-priced | $1.50 | $0.15 | none | $7.50 | | Mistral | Mistral Small 4 | Cheap | $0.15 | $0.015 | none | $0.60 | | Mistral | Ministral 3 3B | Cheapest | $0.10 | $0.01 | none | $0.10 | | Moonshot | Kimi K3.1 Pro | Flagship | $2.00 | $0.20 | $2.00 (5 min) | $10.00 | | Moonshot | Kimi K3 | Previous flagship | $3.00 | $0.30 | $3.00 (5 min) | $15.00 | | Alibaba | Qwen3.8-Max, Singapore | Flagship | $2.00 | $0.25 | $2.50 | $6.00 | | Alibaba | Qwen3.8-Flash, Singapore | Cheap | $0.15 | $0.016 | $0.20 | $0.47 | | Alibaba | Qwen3.7-Flash, Singapore, up to 32K | Cheapest | $0.03 | $0.003 | $0.0375 | $0.13 | | Zhipu (Z.ai) | GLM-5.3 | Flagship | $1.40 | $0.26 | none | $4.40 | | Zhipu (Z.ai) | GLM-5.3-Flash | Cheap | $0.15 | $0.03 | none | $0.50 | | Groq | GPT-OSS 120B | Largest self-serve | $0.15 | $0.075 | none | $0.60 | | Groq | GPT-OSS 20B | Cheap | $0.075 | $0.037 | none | $0.30 | | Fireworks | GPT-OSS 120B | Open model | $0.15 | $0.015 | none | $0.60 | | Together | GPT-OSS 120B | Open model | $0.15 | no cached rate | none | $0.60 | US dollars per million tokens, standard tier, as of 30 October 2026. Sources: \[OpenAI pricing\](https://developers.openai.com/api/docs/pricing) and \[models\](https://developers.openai.com/api/docs/models); \[Anthropic pricing\](https://platform.claude.com/docs/en/about-claude/pricing); \[Gemini API pricing\](https://ai.google.dev/gemini-api/docs/pricing); \[xAI models\](https://docs.x.ai/developers/models); \[DeepSeek Models & Pricing\](https://api-docs.deepseek.com/quick\_start/pricing); \[Mistral pricing\](https://docs.mistral.ai/inference/pricing); \[Kimi API Platform pricing\](https://platform.kimi.ai/docs/pricing/chat); \[Alibaba Cloud Model Studio pricing\](https://www.alibabacloud.com/help/en/model-studio/model-pricing); \[Z.ai pricing\](https://docs.z.ai/guides/overview/pricing); Groq's \[GPT-OSS 120B\](https://console.groq.com/docs/model/openai/gpt-oss-120b) and \[GPT-OSS 20B\](https://console.groq.com/docs/model/openai/gpt-oss-20b) pages; \[Fireworks serverless pricing\](https://docs.fireworks.ai/serverless/pricing); \[Together model catalog\](https://docs.together.ai/docs/serverless/models). Alibaba does not print cache prices for Qwen3.8-Max and Qwen3.8-Flash, so those two cache rows are the rates \[OpenRouter's endpoint data\](https://openrouter.ai/api/v1/models/qwen/qwen3.8-max/endpoints) lists for Alibaba's own endpoints on 30 October. "Flagship" follows each provider's own label where it gives one: OpenAI calls GPT-6 Astra "our most capable model for the most demanding work" on its \[models page\](https://developers.openai.com/api/docs/models), xAI calls Grok 4.7 "our flagship model for code and everything else", and Moonshot's model list calls Kimi K3 its previous-generation flagship, which is why K3.1 Pro, at a third less on every line, is the Kimi model to price first. Groq does not train models; GPT-OSS 120B and 20B are OpenAI's open-weight models and the only production text models Groq sells on a self-serve account, as our \[Groq guide\](https://www.maximem.ai/blog/groq-api-pricing-2026) explains. Mistral cut Large 4's price in late October: on 28 October its pricing page and \[model card\](https://docs.mistral.ai/models/mistral-large-4-0) listed $1.36 input, $0.14 cached and $4.18 output, and on 30 October both list $0.70, $0.07 and $2.10, so guides that captured the earlier rate overstate it by almost double. xAI's models page lists Grok 4.7 at $2 and $6 with no cached-input rate, so Grok gets no cache-warm figure below. ### The same price at three different companies Read the table by price instead of by provider and the market looks much smaller. OpenAI and Anthropic now list their top, middle and bottom models at identical rates: Astra and Fable 5.1 at $10 and $50, Sol and Sonnet 5.5 at $2 and $10, Luna and Haiku 5.5 at $0.10 and $0.50. Moonshot's K3.1 Pro joins the $2 and $10 tier, and so does Gemini 4 Argon for one more month. At the bottom, DeepSeek, Mistral and every host of GPT-OSS 120B charge $0.15 and $0.60. Ties like these are broken by three lines that most comparison tables leave out: the cache-read rate, any cache-write fee, and the multipliers for time of day, region and prompt length. For a single short request none of them matter much. For a conversation they decide the bill, which is what the next table shows. ## What a 40-turn conversation costs on every major LLM API Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same assumptions as every one of our per-provider pricing guides, so the figures line up across all eleven pages. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because the API is stateless and each request carries every exchange before it. The whole conversation consumes 476,000 input tokens to produce 14,000 output tokens. Reasoning tokens are left out, so a model that thinks before it answers adds to every output figure. | Model, price as of 30 October 2026 | 40 turns, no caching | 40 turns, cache warm | 10,000 conversations a month, cache warm | |---|---|---|---| | Qwen3.7-Flash, Singapore | $0.016 | $0.004 | $40 | | Groq GPT-OSS 20B | $0.040 | $0.023 | $226 | | Mistral Ministral 3 3B | $0.049 | $0.008 | $82 | | Claude Haiku 5.5 | $0.055 | $0.014 | $142 | | OpenAI GPT-6 Luna | $0.055 | $0.014 | $142 | | Qwen3.8-Flash, Singapore | $0.078 | $0.018 | $182 | | GLM-5.3-Flash | $0.078 | $0.024 | $239 | | DeepSeek \`deepseek-flash\`, off-peak | $0.080 | $0.013 | $130 | | DeepSeek \`deepseek-flash\`, peak | $0.160 | $0.026 | $260 | | Mistral Small 4 | $0.080 | $0.019 | $186 | | GPT-OSS 120B on Fireworks | $0.080 | $0.018 | $185 | | GPT-OSS 120B on Groq | $0.080 | $0.046 | $457 | | GPT-OSS 120B on Together (no cached rate) | $0.080 | $0.080 | $798 | | Gemini 3.1 Flash-Lite | $0.140 | not modelled | not modelled | | DeepSeek V4 Pro, off-peak | $0.342 | $0.052 | $520 | | DeepSeek V4 Pro, peak | $0.684 | $0.104 | $1,040 | | Mistral Large 4 | $0.363 | $0.077 | $771 | | Gemini 3.8 Flash, to 31 December 2026 | $0.410 | $0.111 | $1,106 | | Gemini 3.8 Flash, from 1 January 2027 | $0.819 | $0.221 | $2,213 | | GPT-6.1 Terra | $0.546 | $0.120 | $1,198 | | GLM-5.3 | $0.728 | $0.210 | $2,100 | | Mistral Medium 3.5 | $0.819 | $0.207 | $2,073 | | Qwen3.8-Max, Singapore | $1.04 | $0.252 | $2,517 | | Grok 4.7 | $1.04 | no cached rate listed | n/a | | Claude Sonnet 5.5 | $1.09 | $0.240 | $2,396 | | GPT-6.1 Sol | $1.09 | $0.240 | $2,396 | | Kimi K3.1 Pro | $1.09 | $0.274 | $2,742 | | Gemini 4 Argon, to 30 November 2026 | $1.09 | $0.387 | $3,871 | | Gemini 4 Argon, from 1 December 2026 | $2.18 | $0.774 | $7,742 | | Kimi K3 | $1.64 | $0.411 | $4,113 | | Claude Opus 5.5 | $2.18 | $0.479 | $4,791 | | Claude Fable 5.1 | $5.46 | $1.08 | $10,842 | | OpenAI GPT-6 Astra | $5.46 | $1.43 | $14,250 | Our arithmetic on the rates in the first table. Rows are ordered by uncached cost. "Cache warm" follows each provider's own caching rule, the same way each per-provider guide models it: - \*\*Anthropic and OpenAI\*\* cache from turn 1, since the first request is already 2,150 tokens and OpenAI's minimum is 1,024 according to its \[prompt caching guide\](https://developers.openai.com/api/docs/guides/prompt-caching). Each turn writes its new tokens at 1.25 times the input price and reads everything before them, and every reply is assumed to arrive inside the 5-minute (Claude) or 30-minute (OpenAI) window. - \*\*Alibaba's explicit cache\*\* works the same way, with a write at 125% of input. - \*\*DeepSeek, Groq, Z.ai, Fireworks and Kimi\*\* read the previous request's prompt at the cached rate from turn 2 and bill the 500 new tokens at the normal input rate. Kimi now bills a 5-minute cache write, but at the input rate, so it adds nothing here. Mistral is the same with its cached count rounded down to 64-token blocks. - \*\*Google's implicit cache\*\* starts once the previous request passes 4,096 tokens, which is turn 6 here, according to its \[caching page\](https://ai.google.dev/gemini-api/docs/caching). On Gemini 4 Argon, a request that adds fewer than 2,050 tokens makes Gemini "re-process the last 2,050 tokens of the previous request at the standard input rate", which every turn of a chat does. The caching page lists no implicit minimum for Flash-Lite, so we left its warm figure out instead of guessing. None of the caches is guaranteed, and Google's page says only that it passes on savings "if your request hits caches", so treat the warm column as the best case and the uncached column as the ceiling. ### Five things the conversation table shows \*\*The cheapest LLM API for a chat is not the cheapest per token.\*\* Qwen3.7-Flash wins on both columns, at $0.016 uncached and $0.004 warm, but it stays that cheap only while every request is under 32,000 tokens; above that Alibaba bills the whole request at $0.10 and $0.40, and our \[Qwen guide\](https://www.maximem.ai/blog/qwen-api-pricing-2026) shows the uncached 100-turn version of this conversation costing $0.206, almost half of \`deepseek-flash\` off-peak at $0.425. Among the cheap models without a tier boundary, DeepSeek \`deepseek-flash\` off-peak at $0.013 warm beats Claude Haiku 5.5 and GPT-6 Luna at $0.014 even though it lists higher on both input and output, because a $0.003 cache hit is less than a third of their $0.01. \*\*Caching reorders the bottom of the table.\*\* Groq's GPT-OSS 20B is the cheapest uncached model after Qwen3.7-Flash, at $0.040, and falls behind eight other models once the cache is warm, because Groq's cache discount is 50% where the others are 80% to 98%. GPT-OSS 120B shows the same effect across hosts: identical $0.080 uncached on Groq, Fireworks and Together, then $0.046, $0.018 and $0.080 warm, a 4.4-fold spread on the same weights. \*\*Same list price, different bill.\*\* Claude Sonnet 5.5 and GPT-6.1 Sol come out identical to the cent, $1.09 and $0.240, because their input, output, cache-read and cache-write rates are all the same. Kimi K3.1 Pro matches them uncached and costs $0.274 warm, because its cache hit is $0.20. Gemini 4 Argon matches them uncached and costs $0.387 warm, because of the 2,050-token rule, and from 1 December it costs $0.774, more than Claude Opus 5.5. \*\*At the top, the cache read decides it.\*\* GPT-6 Astra and Claude Fable 5.1 cost the same $5.46 without caching. With caching, Fable's $0.25 read against Astra's $1.00 makes Fable 24% cheaper per conversation, which at 10,000 conversations a month is $10,842 against $14,250. \*\*The bill grows faster than the conversation.\*\* Going from 40 to 100 turns multiplies turns by 2.5 and the uncached bill by roughly 5.2 on every model without a context tier, because each turn carries every turn before it; a 100-turn conversation sends 2,690,000 input tokens. Our per-provider guides carry the 100-turn columns. ## Cache read and cache write prices, compared | Provider and model | Cache hit as % of input | Cache write fee | Cache lifetime | |---|---|---|---| | DeepSeek \`deepseek-flash\` | 2% | none | best effort | | DeepSeek V4 Pro | 3.3% | none | best effort | | Claude Fable 5.1 | 2.5% | 1.25x (5 min), 2x (1 hour) | 5 minutes or 1 hour | | Claude Opus 5.5 and Sonnet 5.5 | 5% | 1.25x, 2x | 5 minutes or 1 hour | | GPT-6.1 Sol and Terra | 5% | 1.25x (30 min), 2.5x (12 hours) | 30 minutes or 12 hours | | GPT-6 Astra and Luna | 10% | 1.25x, 2.5x | 30 minutes or 12 hours | | Claude Haiku 5.5 | 10% | 1.25x, 2x | 5 minutes or 1 hour | | Gemini 4 Argon and 3.8 Flash | 5% and 10% | none (implicit) | not guaranteed | | Mistral, all models | 10% | none | not guaranteed | | Kimi K3.1 Pro and K3 | 10% | 1x (5 min), 2x (1 hour) | 5 minutes or 1 hour | | Qwen3.7-Flash, explicit cache | 10% | 1.25x | 5 minutes | | Qwen3.8-Max (Alibaba endpoint on OpenRouter) | 12.5% | 1.25x | 5 minutes | | GLM-5.3 | 18.6% | none | not published | | GLM-5.3-Flash | 20% | none | not published | | Groq GPT-OSS models | 50% | none | 2 hours | | Fireworks GPT-OSS 120B | 10% | none | minutes to hours | | Together GPT-OSS 120B | no discount | none | n/a | Percentages are our division of the cached rate by the input rate in the first table. Write multipliers are from \[Anthropic's pricing page\](https://platform.claude.com/docs/en/about-claude/pricing), \[OpenAI's caching guide\](https://developers.openai.com/api/docs/guides/prompt-caching) and the \[Kimi pricing page\](https://platform.kimi.ai/docs/pricing/chat); lifetimes for Qwen, Groq and Fireworks are from the caching documentation cited in our per-provider guides. A cache write fee only costs money when the cache is not read again, so for a conversation that keeps going it is small: on Claude Sonnet 5.5 the write premium (the extra quarter of the input price) comes to about 4.5% of the warm 40-turn bill. A cache that expires before the user replies is the expensive case, because the whole history is written again at the higher rate. The lifetime column matters more than it looks for anything with a human in the loop: a 5-minute cache does nothing for a user who comes back after lunch, and Groq's 2-hour cache does nothing for one who comes back tomorrow. ## The cheapest LLM API in 2026, by what you need The cheapest model depends on whether your requests repeat a long prefix, where the inference has to run, and how long a single request gets. On the 30 October rates: - \*\*Cheapest overall for short requests:\*\* Qwen3.7-Flash at $0.03 input and $0.13 output, up to 32,000 tokens per request, in Alibaba's Singapore region. - \*\*Cheapest for long, cache-heavy conversations without a context tier:\*\* DeepSeek \`deepseek-flash\` off-peak, with cache hits at $0.003 and no write fee. Peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays) double every rate. - \*\*Cheapest from a US frontier lab:\*\* Claude Haiku 5.5 and GPT-6 Luna, tied at $0.10 and $0.50. Haiku 5.5 charges five times as much on any request over 100,000 tokens, and OpenAI moves every model to a higher rate above 272,000 tokens. - \*\*Cheapest output-heavy model:\*\* Ministral 3 3B, whose output costs the same $0.10 as its input. - \*\*Cheapest on an inference cloud:\*\* Groq's GPT-OSS 20B at $0.075 and $0.30, half of GPT-OSS 120B on every line. - \*\*Cheapest per conversation for an open model on a host:\*\* GPT-OSS 120B on Fireworks, at $0.018 warm, because Fireworks prices its cached input at a tenth of Groq's. Cheapest per token is not cheapest per task. A model that needs a second attempt, a longer prompt or more reasoning tokens to reach the same answer costs more than its rate card suggests, so price the two or three cheapest candidates on your own prompts before committing a workload. ## Flagship LLM API prices compared At the top end the spread is narrower than the names suggest. Seven flagships list at $2 input or less: DeepSeek V4 Pro at $0.66 off-peak, Mistral Large 4 at $0.70, GLM-5.3 at $1.40, and Gemini 4 Argon (until 30 November), Grok 4.7, Kimi K3.1 Pro and Qwen3.8-Max at $2. On output, Mistral Large 4 charges $2.10, Grok 4.7 and Qwen3.8-Max charge $6, GLM-5.3 charges $4.40 and the $2 tier charges $10. Above that, Claude Opus 5.5 lists at $4 and $20, and the two most expensive models, GPT-6 Astra and Claude Fable 5.1, at $10 and $50. For a 40-turn chat with no caching, the order is DeepSeek V4 Pro off-peak at $0.342, Mistral Large 4 at $0.363, GLM-5.3 at $0.728, Qwen3.8-Max and Grok 4.7 at $1.04, Kimi K3.1 Pro and Gemini 4 Argon at $1.09, Kimi K3 at $1.64, Claude Opus 5.5 at $2.18, and Astra and Fable 5.1 at $5.46. With the cache warm the two cheapest flagships stay in front, at $0.052 for V4 Pro and $0.077 for Large 4, and Large 4 is a Public Preview model, which our \[Mistral guide\](https://www.maximem.ai/blog/mistral-api-pricing-2026) covers alongside the rest of Mistral's line-up. Two caveats sit under every flagship row. Reasoning is billed as output: Google's pricing page labels its output column "including thinking tokens", DeepSeek bills "the total number of input and output tokens", and our guides document the same for Claude, Kimi, Qwen and Groq. Kimi K3.1 Pro and GLM-5.3 cannot switch reasoning off, so their real conversation cost sits above the table by however much they think. And a token is not the same unit on every model, because each provider uses its own tokenizer; the table gives every model the same token counts, which is the only way to compare rate cards, and your own traffic will tokenize somewhat differently on each. ## Prices that change in the next 90 days | Date | Change | Source | |---|---|---| | 31 October 2026 | Kimi K2 series closes to newly registered users; Moonshot plans to retire all K2 models by the end of December 2026 | \[Kimi pricing\](https://platform.kimi.ai/docs/pricing/chat) | | 1 November 2026 | DeepSeek starts billing Search and Fetch calls made from DeepSeek Harness, Claude Code and Codex at $0.75 per 1,000 successful requests | \[DeepSeek pricing\](https://api-docs.deepseek.com/quick\_start/pricing) | | 1 December 2026 | Gemini 4 Argon doubles to $4.00 input, $0.20 cached and $20.00 output | \[Gemini API pricing\](https://ai.google.dev/gemini-api/docs/pricing) | | 1 January 2027 | Gemini 3.8 Flash and 3.6 Flash double to $1.50 input, $0.15 cached and $7.50 output; cache storage goes from $0.50 to $1.00 per million tokens per hour | \[Gemini API pricing\](https://ai.google.dev/gemini-api/docs/pricing) | | No end date published | Qwen3.7-Plus sells at a limited-time 20% discount on its $0.40 and $1.60 list price | \[Alibaba Cloud pricing\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) | The two Gemini increases are the ones that move rankings. On 1 December, the Argon conversation goes from $1.09 to $2.18 uncached, the same as Claude Opus 5.5, and from $0.387 to $0.774 warm, more than Opus. On 1 January, Gemini 3.8 Flash goes from $0.111 to $0.221 warm, which puts it above GPT-6.1 Terra ($0.120) and just under Claude Sonnet 5.5 ($0.240). A team that budgets for 2027 on Gemini's October prices will be wrong by a factor of two from the first invoice. ## Multipliers that change the bill: peak hours, US-only, long context and batch | Multiplier | Who charges it | What it does | |---|---|---| | Peak hours | DeepSeek | Every rate doubles 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays | | US-only or regional processing | Anthropic 1.1x (Claude 4.6 and later); OpenAI +10% for models released on or after 5 March 2026; Together 1.25x; Fireworks 1.5x | Applies to every token category, cache included | | Long context | Claude Haiku 5.5 5x above 100K tokens; Gemini 3.1 Pro $4 and $18 above 200K; OpenAI higher rates above 272K; Qwen3.7-Flash $0.10 and $0.40 above 32K | The higher rate applies to the whole request | | Batch and async | 50% off on Anthropic, OpenAI (Flex and Batch), Google, Alibaba (models that support batch) and Fireworks | Results within hours, not seconds | | Speed tiers | OpenAI Fast at 2x; Fireworks Fast about 1.5x; Kimi K3.1 Pro Highspeed at 2x input and output | Pay more for lower latency | Sources: each provider's pricing page linked in the first table. OpenAI's page has a short and long context toggle without printing the threshold; OpenRouter's \[endpoint data\](https://openrouter.ai/api/v1/models/openai/gpt-6-astra/endpoints) for OpenAI's own endpoint lists the higher rates from 272,000 prompt tokens, at twice the input price and 1.5 times the output price. The US-only row is where one provider's pages disagree with each other. Fireworks' \[serverless pricing table\](https://docs.fireworks.ai/serverless/pricing) lists Kimi K3 (US) at $4.50 input, $0.45 cached and $22.50 output, which is 1.5 times its $3.00, $0.30 and $15.00 standard rate, while the \[Kimi K3 model page\](https://fireworks.ai/models/fireworks/kimi-k3) describes the US-only endpoint as "+10%". We use the pricing table, which is what billing follows elsewhere on the docs site, and at those rates the 40-turn conversation costs $2.46 uncached and $0.617 warm on Fireworks, against $2.05 and $0.514 on \[Together's\](https://docs.together.ai/docs/serverless/models) US-only endpoint at 1.25 times list. If you must keep inference in the United States, Together is 16.7% cheaper on every line of the models both serve. Data residency also cuts the other way. DeepSeek, Moonshot, Alibaba and Zhipu are Chinese companies, and each of our per-provider guides covers where requests are processed and which entity you contract with. The same open weights (GPT-OSS, GLM-5.3, Kimi K3, DeepSeek V4.1 Flash) are sold by US hosts at the same list price, so for most teams the choice between a Chinese provider and a US host comes down to caching and residency rather than the headline rate. ## Hosts and routers: the same model at different prices Open-weight models are sold by their makers and by inference clouds, and the hosts converge on one list price and compete on everything around it. GPT-OSS 120B costs $0.15 input and $0.60 output on Groq, Fireworks and Together alike; the cached rate is $0.075 on Groq, $0.015 on Fireworks and not offered on Together. Kimi K3 costs $3.00, $0.30 and $15.00 on Moonshot, Fireworks and Together. GLM-5.3 costs $1.40, $0.26 and $4.40 on Z.ai and Together. The bill is decided by caching, batch discounts, US-only multipliers and speed, which our \[Fireworks vs Together comparison\](https://www.maximem.ai/blog/fireworks-vs-together-ai-pricing-2026) sets out model by model. \[OpenRouter\](https://openrouter.ai/pricing) sits on top of all of them. It passes through each provider's price "without any markup", according to its \[FAQ\](https://openrouter.ai/docs/faq), and charges a fee when you buy credits: 5.5% on Standard and 8% on Business. Bring-your-own-key traffic is free up to $25,000 of list-price inference a month and costs 5% after that. On a warm 40-turn Claude Sonnet 5.5 conversation the Standard fee adds about $0.013, and at 10,000 conversations a month it adds about $132; the fee never changes as a percentage, so it matters in absolute terms only on expensive models. Our \[OpenRouter guide\](https://www.maximem.ai/blog/openrouter-pricing-2026) covers routing, which can send a request to a host with a different price from the one you expected. Behind the prices sits model churn, which is part of the cost of a cheap row. Kimi's cheaper K2 models close to new accounts on 31 October, Groq moved its Llama models to enterprise-only contracts in August, and Mistral lists Large 4 as a Public Preview model that may be retired before reaching general availability. A model that disappears in three months costs a migration, and that cost does not appear on any rate card. ## LLM API free tiers in 2026 | Provider | Free tier on 30 October 2026 | |---|---| | Google | Gemini 3.8 Flash and 3.1 Flash-Lite are free of charge on the free tier; Gemini 4 Argon has no free tier | | Alibaba | A free quota for each eligible Qwen model, valid 90 days, in the Singapore region only on the international site | | Zhipu (Z.ai) | GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash are free on input, cached input, storage and output | | Groq | Free plan on GPT-OSS 120B and 20B at 30 requests a minute, 1,000 a day, 8,000 tokens a minute and 200,000 tokens a day | | OpenRouter | More than 25 free models, at 50 requests a day on the Free plan | | Mistral | Free mode for new accounts, with an allowance Mistral does not publish | Sources: the pricing pages linked above, plus Groq's \[rate limits page\](https://console.groq.com/docs/rate-limits). Anthropic, OpenAI, DeepSeek, Moonshot and xAI list no free rate on the pricing tables we read. A free tier holds a prototype, not a product. On Groq, turn 13 of the conversation above sends 8,150 tokens, which is more than the free plan's 8,000 tokens a minute on its own, and 40 turns move 490,000 tokens without caching, more than twice the free daily allowance. Use free tiers to compare outputs on your own prompts, then move to paid usage with a spend limit set on the first day. ## How to estimate your own LLM API bill A request costs (cached input tokens times the cached rate) plus (cache-write tokens times the write rate) plus (other input tokens times the input rate) plus (output and reasoning tokens times the output rate), with any peak, regional, long-context or batch multiplier applied on top. For a conversation, the input on turn n is the system prompt plus every earlier turn, so add up the turns of a typical session as the table above does, then multiply by sessions per month. Four inputs decide most of the answer: the length of the system prompt and tool definitions, how many turns a session runs, how often a cache is still warm when the next turn arrives, and how many reasoning tokens the model spends per reply. Caching is the first lever, and the table shows how far it goes: it cuts the 40-turn bill by roughly 70% to 85% on most models. It has two limits. A cache cuts the price of the repeated tokens; it does not stop them being sent, so the bill still grows with the square of the conversation length, and it expires, so a user who comes back tomorrow pays to write the whole history again. The other lever is to stop resending the whole transcript and send only what the next turn needs. We build \[Maximem Synap\](https://www.maximem.ai/synap), a memory layer whose retrieval results are merged, deduplicated, ranked on relevance, recency and confidence, and trimmed to a token budget you set, in place of the full history. The conversation length at which that beats caching, and where it does not, is worked through in \[how to reduce LLM token costs in long conversations\](https://www.maximem.ai/blog/reduce-llm-token-costs-long-conversations), and the context limits themselves are in \[LLM context window sizes in 2026\](https://www.maximem.ai/blog/llm-context-window-sizes-2026). ## Why other LLM pricing comparisons disagree Most of the pages that rank for "LLM API pricing comparison" priced their tables between May and August 2026, and the market has moved under them since. They lead with GPT-5 rows where OpenAI's current family is GPT-6, Claude Opus 5 and Haiku 4.5 where Anthropic now sells Opus 5.5 and Haiku 5.5 at lower rates, and DeepSeek V3 where DeepSeek's API serves \`deepseek-flash\` and V4 Pro. Few of them list cached input at all, and none of the five we read prices a multi-turn conversation, which is the number that decides a chat or agent bill. A comparison table is only as current as its last read of each provider's page, which is why every row on this page carries the date it was read and links to the page it came from. ## Frequently asked questions ### What is the cheapest LLM API in 2026? As of 30 October 2026, Alibaba's Qwen3.7-Flash is the cheapest at $0.03 per million input tokens and $0.13 per million output tokens for requests up to 32,000 tokens. For long conversations with caching, DeepSeek \`deepseek-flash\` off-peak is cheapest among models without a context tier, with cache hits at $0.003 per million. From the US frontier labs, Claude Haiku 5.5 and GPT-6 Luna tie at $0.10 input and $0.50 output. ### Which LLM API is the most expensive? OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 both list at $10 per million input tokens and $50 per million output tokens as of 30 October 2026. With caching, Astra costs more for a conversation, because it reads its cache at $1.00 per million against Fable 5.1's $0.25. ### How much does a chatbot conversation cost on an LLM API? A 40-turn conversation with a 2,000-token system prompt, 150-token messages and 350-token replies costs from $0.016 (Qwen3.7-Flash) to $5.46 (GPT-6 Astra or Claude Fable 5.1) without caching, and from $0.004 to $1.43 with a warm cache, at 30 October 2026 prices. Claude Sonnet 5.5 and GPT-6.1 Sol both cost $1.09 uncached and $0.240 cached. ### Is Claude cheaper than OpenAI? At list price they are now identical at three tiers: Claude Fable 5.1 and GPT-6 Astra at $10 and $50, Claude Sonnet 5.5 and GPT-6.1 Sol at $2 and $10, and Claude Haiku 5.5 and GPT-6 Luna at $0.10 and $0.50 per million tokens. With caching, Fable 5.1 is cheaper than Astra and the other two pairs cost the same, as of 30 October 2026. ### Is DeepSeek cheaper than GPT and Claude? On cached conversations, yes: a 40-turn chat on \`deepseek-flash\` costs $0.013 off-peak with caching, against $0.014 on Claude Haiku 5.5 and GPT-6 Luna and $0.240 on Claude Sonnet 5.5 and GPT-6.1 Sol. Without caching, Haiku 5.5 and Luna are cheaper ($0.055 against $0.080), and DeepSeek's peak hours double its rates. ### Which LLM API prices are going up? Gemini 4 Argon doubles to $4 input and $20 output on 1 December 2026, and Gemini 3.8 Flash and 3.6 Flash double to $1.50 and $7.50 on 1 January 2027, according to Google's pricing page on 30 October 2026. Kimi's K2 models close to new users on 31 October 2026. ### Do Groq, Fireworks and Together charge the same for open models? On list price, mostly yes: GPT-OSS 120B costs $0.15 input and $0.60 output on all three as of 30 October 2026. Cached input differs, at $0.075 on Groq, $0.015 on Fireworks and no cached rate on Together, so a 40-turn conversation costs $0.046, $0.018 and $0.080 respectively with the cache warm. ### Does OpenRouter mark up model prices? No. OpenRouter passes through provider prices without markup and charges a 5.5% fee on credit purchases on its Standard plan and 8% on Business, as of 30 October 2026. Bring-your-own-key usage is free up to $25,000 of list-price inference a month and costs 5% above that. ## Sources All retrieved 30 October 2026. - OpenAI: \[API pricing\](https://developers.openai.com/api/docs/pricing), \[models\](https://developers.openai.com/api/docs/models), \[prompt caching\](https://developers.openai.com/api/docs/guides/prompt-caching) - Anthropic: \[Claude API pricing\](https://platform.claude.com/docs/en/about-claude/pricing) - Google: \[Gemini Developer API pricing\](https://ai.google.dev/gemini-api/docs/pricing), \[context caching\](https://ai.google.dev/gemini-api/docs/caching) - xAI: \[models and pricing\](https://docs.x.ai/developers/models) - DeepSeek: \[Models & Pricing\](https://api-docs.deepseek.com/quick\_start/pricing) - Mistral: \[API pricing\](https://docs.mistral.ai/inference/pricing), \[Mistral Large 4 model card\](https://docs.mistral.ai/models/mistral-large-4-0), \[changelog\](https://docs.mistral.ai/resources/changelogs) - Moonshot AI: \[Kimi API Platform pricing\](https://platform.kimi.ai/docs/pricing/chat) - Alibaba Cloud: \[Model Studio model pricing\](https://www.alibabacloud.com/help/en/model-studio/model-pricing) - Z.ai: \[pricing\](https://docs.z.ai/guides/overview/pricing) - Groq: \[models\](https://console.groq.com/docs/models), \[GPT-OSS 120B\](https://console.groq.com/docs/model/openai/gpt-oss-120b), \[GPT-OSS 20B\](https://console.groq.com/docs/model/openai/gpt-oss-20b), \[rate limits\](https://console.groq.com/docs/rate-limits) - Fireworks AI: \[serverless pricing\](https://docs.fireworks.ai/serverless/pricing), \[Kimi K3 model page\](https://fireworks.ai/models/fireworks/kimi-k3), \[pricing\](https://fireworks.ai/pricing) - Together AI: \[serverless model catalog\](https://docs.together.ai/docs/serverless/models), \[pricing\](https://www.together.ai/pricing) - OpenRouter: \[pricing\](https://openrouter.ai/pricing), \[FAQ\](https://openrouter.ai/docs/faq), and the endpoints API for \[Qwen3.8-Max\](https://openrouter.ai/api/v1/models/qwen/qwen3.8-max/endpoints), \[Qwen3.8-Flash\](https://openrouter.ai/api/v1/models/qwen/qwen3.8-flash/endpoints) and \[GPT-6 Astra\](https://openrouter.ai/api/v1/models/openai/gpt-6-astra/endpoints) Worked examples are our own arithmetic on those rates, using the same method as our per-provider pricing guides.

---

Source: [https://www.maximem.ai/blog/llm-api-pricing-comparison-2026](https://www.maximem.ai/blog/llm-api-pricing-comparison-2026)