A current, source-verified comparison of the 13 leading large language models on price, context window, benchmark standing and real-world cost, plus a free interactive calculator to price any workload across every model.
Type in your own token numbers and see instantly what every model on this page would cost you each month, ranked cheapest to most expensive.
Choosing a model in 2026 is no longer about one clear winner. Claude Fable 5 leads on raw intelligence, GPT-5.6 Sol tops coding agents, Gemini 3.1 Pro wins on value, and challengers like Kimi K3 and DeepSeek V4 now sit inside striking distance of the frontier at a fraction of the price. This page compares the 13 leading large language models on the numbers that actually decide a project, explains what input and output tokens mean for your bill, covers how voice changes the cost, and lets you price your own usage in the interactive LLM cost calculator further down.
Every price on this page is quoted per million tokens, so it helps to know what that means. A token is a chunk of text, roughly four characters, or about three quarters of an English word. So how many tokens is a word? Roughly three quarters of one. As a rule of thumb, 1,000 tokens is about 750 words, a single page of text is around 500 tokens, and a 2,000 word document is about 2,700 tokens. Turned into money, generating 1,000 words of output costs about $0.04 on GPT-5.6 Sol, about $0.07 on Claude Fable 5, and about a tenth of a cent on DeepSeek V4. Models do not read letters or words, they read tokens, which is why billing is measured this way. Our guide to how large language models work covers the mechanics in depth.
Input tokens are everything you send the model. That includes your prompt, any documents or data you paste in, and, in a chat, the entire conversation so far, which is resent with every turn. Output tokens are everything the model writes back. The split matters because output is almost always the expensive half, though the multiple varies by model and is worth checking for the one you actually use. It is six times on GPT-5.6 Sol ($5 in, $30 out) and five times on Claude Fable 5 ($10 in, $50 out), but only about twice on DeepSeek V4 ($0.44 in, $0.87 out).
| Counts as input (what you send, cheaper) | Counts as output (what you get back, pricier) |
|---|---|
| Your question or instruction | The written answer |
| Documents, code or data you paste in | Generated reports, code or summaries |
| The full chat history, resent every turn | Step-by-step reasoning the model writes out |
| System prompts and tool definitions | Tool calls the model produces |
The practical consequence is that your bill is driven by the shape of your workload, not just the model you pick. Pasting a long report and asking for a one line answer is cheap. Asking for a long generated document is expensive. Long chat sessions quietly inflate input costs, because the whole history is resent on each turn until it fills the context window. That is a simplification worth qualifying: most providers now offer prompt caching, which bills repeated context at a steep discount (Kimi K3 caches input at $0.30 against $3.00, Qwen at $0.25 against $2.50), so a well-engineered app pays considerably less than the naive arithmetic suggests. That is also why retrieval-augmented generation is so widely used, it sends only the relevant passages instead of everything.
Every price is in US dollars per one million tokens, shown as input / output. The Intelligence Index is the Artificial Analysis composite score for July 2026. The value score is Santage's own price-performance measure, explained in the methodology. If you are used to seeing prices quoted per 1,000 tokens, divide by a thousand: GPT-5.6 Sol's $5 per million is $0.005 per 1,000 input tokens and $0.03 per 1,000 output, Claude Fable 5 works out at $0.01 and $0.05, and DeepSeek V4 at roughly $0.0004 and $0.0009.
| Model | Maker | Type | Context | Price in/out (per 1M) | Intelligence Index | Value score | Best at |
|---|---|---|---|---|---|---|---|
| GPT-5.6 Sol | OpenAI | Closed | 1.05M | $5 / $30 | 58.9 | 5.2 | Agentic, coding, tool-heavy work |
| GPT-5.6 Terra | OpenAI | Closed | 1M-class* | $2.5 / $15 | Not rated | Not rated | Balanced cost and quality |
| GPT-5.6 Luna | OpenAI | Closed | 1M-class* | $1 / $6 | Not rated | Not rated | High-volume, budget |
| Claude Fable 5 | Anthropic | Closed | 1M | $10 / $50 | 60.0 | 3.0 | Hardest coding and reasoning |
| Claude Opus 4.8 | Anthropic | Closed | 1M | $5 / $25 | 53.5 | 5.3 | Complex agentic and writing |
| Claude Sonnet 5 | Anthropic | Closed | 1M | $3 / $15 | 53.4 | 8.9 | Balanced everyday workhorse |
| Claude Haiku 4.5 | Anthropic | Closed | 200K | $1 / $5 | Not rated | Not rated | Low latency, high volume |
| Gemini 3.1 Pro | Closed | 1M | $2 / $12 | 46.5 | 10.3 | Reasoning at value, multimodal | |
| Grok 4.5 | xAI | Closed | 500K | $2 / $6 | 53.8 | 17.9 | Real-time, X-native tasks |
| Muse Spark 1.1 | Meta | Closed | 1M | $1.25 / $4.25 | Not rated | Not rated | Agentic, tool and computer use |
| Kimi K3 | Moonshot | API only | 1M | $3 / $15 | 57.0 | 9.5 | Reasoning, long-horizon coding |
| DeepSeek V4 | DeepSeek | Open weights | 1M | $0.44 / $0.87 | 44.3 | 80.9 | Low-cost reasoning at scale |
| Qwen 3.7 Max | Alibaba | API only | 1M | $2.5 / $7.5 | 56.6 | 15.1 | Strong-value flagship, multilingual |
* GPT-5.6 Terra and Luna sit in the same 1M-class family as Sol, but their exact per-tier context windows are not separately disclosed, so we mark them rather than guess.
A note on Llama. We have left Meta's Llama out of the comparison table rather than risk misleading anyone. The widely reported Llama 5, said to carry a 5 million token context window, could not be confirmed against an official Meta model card or announcement, so we will not publish specifications or a value score we cannot stand behind. What Meta does officially document is the previous generation: Llama 4 Scout, with a context window of up to 10 million tokens, the largest of any widely available model. We will add Llama to the table as soon as an official model card for the current generation exists.
A note on Kimi K3's status. Kimi K3 is Moonshot AI's July 2026 flagship, reported at 2.8 trillion parameters with a 1 million token context window. As of July 19, 2026 it is available through Moonshot's API and app only. Downloadable weights are pledged for July 27 but are not yet public, so we do not classify it as an open-weight model today, and the table lists it as API only. Moonshot positions K3 for advanced reasoning, long-horizon coding and multimodal work, and early benchmark reports place it close to the frontier, but those remain reported benchmark claims rather than settled proof of overall leadership.
What "open weights" actually means. A closed model can only be used through the provider's API: you rent access, and the model itself stays on their servers. An open-weight model has had its trained parameters published, so anyone can download it, run it on their own hardware, adapt it, or host it privately. That is not the same as open source, because the training data and code usually remain private and the licence often restricts commercial use. The practical trade is that you stop paying per token and start paying for GPUs, engineering and operations instead. That is usually cheaper at high volume and better for privacy, but it is not free, which is why an open-weight model is not simply free. Labels also matter here: of the models in this table only DeepSeek V4 is genuinely open weight, published under an MIT licence and downloadable today. Kimi K3 has weights pledged but not yet released, and Alibaba's Qwen 3.7 Max is API only, even though earlier Qwen generations were open. We label each by what is actually true on July 19, 2026, not by reputation.
The chart below shows the blended cost per million tokens for each model that publishes a price, using a 3:1 input-to-output ratio (the typical real-world mix). The spread is enormous: the priciest frontier model costs roughly forty times the cheapest.
The open-weight surge is not just a benchmark story, Chinese models now run nearly half of US enterprise AI traffic, and specialised inference providers are raising at multibillion dollar valuations to serve them.
Raw price only tells half the story. This chart plots each model's intelligence against its blended cost. Models toward the upper left deliver the most quality per dollar.
These are Santage-computed scenarios so you can see real spend, not just per-token rates. Assumptions: a typical chat is 1,000 input and 500 output tokens; a document summary is a 2,000-word input (about 2,700 tokens) with a 200-word output (about 270 tokens). All figures in US dollars.
| Model | 1M in + 1M out | 1,000 chats | Summarise 1,000 docs |
|---|---|---|---|
| GPT-5.6 Sol | $35.00 | $20.00 | $21.60 |
| GPT-5.6 Terra | $17.50 | $10.00 | $10.80 |
| GPT-5.6 Luna | $7.00 | $4.00 | $4.32 |
| Claude Fable 5 | $60.00 | $35.00 | $40.50 |
| Claude Opus 4.8 | $30.00 | $17.50 | $20.25 |
| Claude Sonnet 5 | $18.00 | $10.50 | $12.15 |
| Claude Haiku 4.5 | $6.00 | $3.50 | $4.05 |
| Gemini 3.1 Pro | $14.00 | $8.00 | $8.64 |
| Grok 4.5 | $8.00 | $5.00 | $7.02 |
| Muse Spark 1.1 | $5.50 | $3.38 | $4.52 |
| Kimi K3 | $18.00 | $10.50 | $12.15 |
| DeepSeek V4 | $1.31 | $0.88 | $1.42 |
| Qwen 3.7 Max | $10.00 | $6.25 | $8.78 |
Use the free interactive calculator below to estimate and compare the monthly API cost of GPT-5.6 Sol, Claude Fable 5, Gemini 3.1 Pro, Kimi K3, DeepSeek V4 and every other model on this page for your own usage. Enter your monthly input and output tokens, or pick a preset workload, and the calculator instantly ranks all 13 models from cheapest to most expensive and shows the price-performance leader. Everything runs in your browser; nothing is sent anywhere.
Costs are estimated from list prices in US dollars and exclude cached-input discounts, batch pricing and volume deals. Open-weight models can be self-hosted at infrastructure cost instead of per-token pricing.
Yes, usually a great deal more, and it is billed on a separate meter. When you speak to a model through an API, your audio is charged as audio tokens, priced far above text tokens. A minute of speech you send is roughly 600 input tokens, about one token per 100 milliseconds. A minute of speech the model generates is roughly 1,200 output tokens, because generated audio is tokenised about twice as densely. The economics matter more than they used to, because as we reported, voice AI's missing piece was never the voice, it was the reasoning behind it.
The rates tell the story. OpenAI's realtime audio model is priced at $32 per million input audio tokens and $64 per million output, against $5 and $30 for GPT-5.6 Sol in text. In practice that is about $0.019 for every minute the model listens and about $0.077 for every minute it speaks. Full voice agents, once reasoning and tool calls are added, typically run $0.18 to $0.46 per minute, falling to roughly $0.05 to $0.10 per minute with prompt caching and trimmed tool output.
| Service | What it does | Price (USD) |
|---|---|---|
| OpenAI GPT-Realtime-2 | Speech to speech | $32 / $64 per 1M audio tokens (about $0.019 per minute heard, $0.077 per minute spoken) |
| OpenAI Realtime Whisper | Streaming transcription | About $0.017 per minute |
| OpenAI GPT-4o Transcribe | Batch transcription | $6 per 1,000 minutes |
| OpenAI TTS | Speech output | $15 per 1M characters standard, $30 HD |
| OpenAI GPT-Realtime-Translate | Speech translation | About $0.034 per minute |
| Gemini 3.1 Flash Live | Speech to speech | $0.005 per minute input, $0.018 per minute output |
| Gemini 3.5 Flash TTS | Speech output | $6 per 1M output tokens |
| Deepgram Nova-3 | Transcription | About $0.0048 per minute |
| Deepgram Voice Agent API | End to end voice agent | About $4.50 per hour |
| Deepgram Aura-2 | Speech output | $30 per 1M characters, falling to $27 at volume |
| ElevenLabs Scribe | Transcription | $0.22 to $0.48 per hour depending on plan |
| ElevenLabs Flash v2.5 | Speech output | $50 per 1M characters |
| Typical full voice agent | End to end, with reasoning | $0.18 to $0.46 per minute, or $0.05 to $0.10 cached |
The spread looks enormous, but read it carefully, because these are not like-for-like prices. The three categories bill differently and do different amounts of work. A full speech to speech agent bundles transcription, model reasoning and speech output, and is billed per minute or per audio token. Streaming transcription does only the first of those and is billed per minute. Text to speech is usually billed per character. Setting the cheapest transcription (about $0.0048 a minute) against a full agent at the top of its range (about $0.46 a minute) is close to a ninetyfold gap, but that compares two different services, not one service priced two ways.
The conclusion that does hold is narrower and more useful: if all you need is words on a page, transcription-only is dramatically cheaper than running a full voice agent, and the specialist providers undercut the frontier labs on that specific job. One more trap worth knowing: with an open microphone, a realtime session bills input tokens during silence and hold time too, so dead air costs real money.
If you are not building anything, the maths is completely different. Everything above is API pricing for people shipping products. If you simply want to talk to your computer instead of typing, consumer dictation tools charge a flat subscription and none of the per-minute rates apply. Wispr Flow, one of the most popular, is $15 a month or $12 a month billed annually, with a free tier capped at 2,000 words a week on desktop. Talking to ChatGPT or Claude inside their own apps on a normal subscription is likewise already included, and costs you nothing extra per minute.
If you want the best result regardless of cost, Claude Fable 5 is the current intelligence leader, with GPT-5.6 Sol a close and cheaper second that wins on agentic coding. If you are cost-sensitive, DeepSeek V4 delivers most of the quality of far pricier models for well under a dollar per million tokens, and Gemini 3.1 Pro is the best value among the big Western labs. If you need an open model you can host yourself, Kimi K3 is the closest to the frontier, while Llama 5 offers an unmatched 5 million token context for long-document and privacy-sensitive work. For high-volume, latency-sensitive tasks, Claude Haiku 4.5, GPT-5.6 Luna and Muse Spark 1.1 keep costs low.
Most teams overpay, and usually for the same handful of reasons. If your API costs are climbing faster than your usage, these are the levers that actually move the number, roughly in order of how much they save for how little effort.
The story of mid-July 2026 is the open-weight surge. Moonshot's Kimi K3, a 2.8 trillion parameter model, launched on July 16 and immediately ranked fourth on the Artificial Analysis Intelligence Index, ahead of Claude Opus 4.8 and GLM-5, with downloadable weights pledged for July 27 but not public as of July 19. OpenAI's GPT-5.6 family reached general availability on July 9 across its Sol, Terra and Luna tiers, and Meta priced its closed Muse Spark 1.1 at $1.25 input. Claude Fable 5 returned to general availability on July 1 after an 18-day export-control suspension. Google, meanwhile, has rebuilt Gemini from scratch for its next flagship, which is why 3.5 Pro carries no official price yet. We refresh this page on the 1st and 15th of each month, and whenever a major model ships.
Artificial Analysis ranks Claude Fable 5 first on its Intelligence Index at 60. GPT-5.6 Sol follows at 58.9, and Kimi K3 reaches 57 at roughly a third of Fable's price. That is how narrow the quality gap at the top has become. Artificial Analysis Intelligence Index, July 2026
Every price and specification on this page is compiled from primary sources wherever one exists, meaning the providers' own pricing pages, developer documentation and model cards, cross-checked against the Artificial Analysis and BenchLM leaderboards. Where only a secondary source exists, we say so in the notes rather than presenting it as confirmed. Llama 5's reported 5 million token context is the one figure in this edition that rests on a secondary source, and it is flagged accordingly. Where a figure cannot be verified against a primary source, we mark it rather than estimate. Benchmark figures are attributed to their source and dated; we do not run our own benchmarks.
The value score is Santage's own analysis. It divides a model's Artificial Analysis Intelligence Index by its blended price per million tokens, where blended price weights input and output 3:1 to reflect typical real usage. A higher score means more measured intelligence per dollar. We show it only for models that have a published Intelligence Index and a public price; others are marked "not rated." The worked cost examples use the stated token assumptions and list prices. These derived numbers are labelled as Santage analysis throughout and should not be read as vendor figures.
How we chose these 13 models. We include the current flagship from each major frontier lab, the cheaper tiers from those labs that teams actually deploy at volume, and the strongest challengers from outside those labs, including the leading genuinely open-weight model. We exclude models superseded by a newer version from the same lab, models with no published price, and models far enough down the index that they are not realistic choices for the workloads this page addresses. This is an editorial selection, not an exhaustive census, and we revise it each cycle as models ship.
What the value score is, and is not. It is a price-efficiency ratio, not a measure of quality. A high score means a model returns more measured benchmark performance per dollar. It does not mean the model is better. DeepSeek V4 scores highest on value while sitting near the bottom of the raw Intelligence Index. If you need the strongest possible output and cost is secondary, read the Intelligence Index column and ignore the value column entirely.
On comparing like with like. The prices here are first-party list prices for API access, in US dollars, excluding cached-input discounts, batch rates and negotiated volume deals. Open-weight models are a different commercial shape: self-host them and there is no per-token fee but you carry GPU, engineering and operations cost, or buy them from a third-party host and pay that host's margin. Llama 5 is marked self-host for exactly this reason, and its true cost is not zero. Read any comparison between a hosted API price and a self-hosted model as indicative rather than exact.
Dating. Every price, context window, benchmark score and availability status was compiled on July 19, 2026. Where a figure carries its own effective date we say so: Claude Sonnet 5's introductory rate runs to August 31, 2026, and Kimi K3's weights are pledged for July 27, 2026. Leaderboard positions move between our refreshes, so Intelligence Index figures reflect the July 2026 table rather than a permanent ranking.
| Model | Notes and caveats |
|---|---|
| GPT-5.6 Sol | Requests over 272K input tokens are billed at $10 / $45. |
| GPT-5.6 Terra | GPT-5.6 mid tier. Per-tier context not separately disclosed. |
| GPT-5.6 Luna | GPT-5.6 economy tier. Per-tier context not separately disclosed. |
| Claude Fable 5 | Leads the Artificial Analysis Intelligence Index (60). |
| Claude Opus 4.8 | Index figure reflects the Opus 4.x line. |
| Claude Sonnet 5 | Introductory $2 / $10 through Aug 31, 2026, then $3 / $15. |
| Claude Haiku 4.5 | Fast, low-cost current-generation tier. |
| Gemini 3.1 Pro | Officially priced flagship. Gemini 3.5 Pro is not yet officially priced. |
| Grok 4.5 | Current xAI flagship per official xAI developer docs. Grok 4 (256K context) is the prior generation. Price, context and index now all refer to the same model. |
| Muse Spark 1.1 | Meta's closed frontier model, in public preview. |
| Kimi K3 | API and app only as of July 19, 2026. Downloadable weights are pledged for Jul 27 but are not yet public, so it is not an open-weight model today. Cached input $0.30. |
| DeepSeek V4 | Genuinely open weights, published under an MIT licence and downloadable. V4 Pro rates ($0.435 / $0.870). |
| Qwen 3.7 Max | Proprietary and API only. Earlier Qwen generations shipped open weights, but the 3.7 Max tier is closed and cannot be downloaded. Cached input $0.25. |