The LLM guide: ChatGPT, Gemini, Claude, DeepSeek… models, token prices and limits

An LLM, or large language model, is an artificial intelligence program trained on vast amounts of text to predict, word by word, the most likely continuation of a text. It is the engine behind ChatGPT, Gemini, Claude, DeepSeek and Meta AI in WhatsApp: you give it an instruction and it drafts, summarises, translates, sorts or answers.

Whether you run a business in Buea, Douala or Yaoundé, or follow one from Houston, London or Toronto, the question is no longer whether these tools are useful. It is which one to pick, what it will cost each month, and what it will do with your data.

This guide answers with the figures on the providers’ own pages, converted into FCFA and euros. Prices move fast, and some already have an announced end date, so every table shows, under its rows, the date its figures were checked. Read the date before the number.

Why does the choice of model matter for a business?

Because three things vary enormously from one model to the next: the price, the quality in your language, and what happens to your data.

Price. In our tables, the most expensive model charges a hundred times more per output token than the cheapest. For occasional use, the gap is a few cents. For an assistant handling hundreds of messages a day, it runs to tens of thousands of FCFA a month.

Language. Public leaderboards rest mostly on tests in English. That sounds like good news for an English-speaking business, but your customers do not write textbook English: they write Cameroonian English, Pidgin, French, or all three in one message. Only a test on your own messages will tell you how a model copes.

Your data. Depending on the plan, what you type may or may not be used to train the provider’s models. The difference lies less in the model than in the contract: personal plan, business plan or API.

A word on terms. The API is the technical access, billed per token, that a developer connects to your software, your website or your WhatsApp. A subscription is the app you open in a browser or on your phone, billed by the month. The same model often comes in both, with different limits and rules.

Who makes which models, and which are offered in Cameroon?

This guide follows ten providers. Five are American: OpenAI (ChatGPT and the GPT models), Anthropic (Claude), Google (Gemini), Meta (Meta AI and the Muse models) and xAI (Grok), which its own pages now call SpaceXSI. One is French: Mistral AI. Four are Chinese: DeepSeek, Alibaba Cloud (Qwen), Moonshot AI (Kimi) and Z.ai (GLM).

Their models fall into two families:

• closed models run only with their publisher or the hosting companies it works with: you rent access, you never receive the model;
• open-weight models publish their “weights”, the parameters learned in training: you can download them and run them on your own servers, within the terms of their licence. DeepSeek, Moonshot, Z.ai and Mistral do so for the models in our tables, except Mistral Large 4, whose weights Mistral announces for the end of October 2026, and Kimi K3.1 Pro, whose weights Moonshot announces by 4 November; Alibaba, Google, Meta and OpenAI also publish some open-weight models.

The next table gives each publisher’s own description of its model (not our opinion) and where it sits in the price range. “Preview” marks a model the publisher does not yet call stable.

What each model is for, by its publisher’s positioning and its price

What each model is for, by its publisher’s positioning and its price

Model	How the publisher positions it
Claude Opus 5.5Anthropic	Anthropic’s suggested starting point: long-running agentic coding and knowledge work
Claude Sonnet 5.5Anthropic	The best combination of speed and intelligence, according to Anthropic
Claude Haiku 5.5Anthropic	High-volume, latency-sensitive work: classification, extraction, routing
GPT-6 AstraOpenAI	OpenAI’s most capable model, for the most demanding work
GPT-6.1 SolOpenAI	Near-Astra performance at a lower cost, according to OpenAI
GPT-6 LunaOpenAI	OpenAI’s most efficient model, for focused, high-volume tasks
Gemini 4 ArgonGoogle	Google’s flagship: complex coding, maths, science and multi-step agents
Gemini 3.8 FlashGoogle	Google’s most intelligent Flash model, with a free API tier
Muse Spark 1.3Meta	Meta’s model for agents and coding, with a cheaper variant paid for with your data
Grok 4.7xAI	xAI’s most capable model, the one it recommends for code and chat
Mistral Large 4 previewMistral AI	Mistral AI’s most capable model, in public preview; open weights announced
DeepSeek V4.1 FlashDeepSeek	Very low cost, half price off-peak, open weights under MIT

Which ones are offered in Cameroon?

Always check the publisher’s own page, never a forum. As of 21 October 2026:

• OpenAI lists Cameroon among its supported countries, for ChatGPT and for its API;
• Anthropic lists it for the Claude API and in its supported regions policy, which also covers Claude.ai and Claude Code and applies from 12 November 2026;
• Google lists it for the Gemini API and for the Google AI Plus, Pro and Ultra plans;
• xAI publishes no country list in its documentation.

For the other publishers in this guide, check availability when you sign up.

What is a token, and how many does a text take?

A token is the unit a model reads and writes: a short word, part of a longer word, a punctuation mark. Everything is billed in tokens: the ones you send, the input (your question, your instructions, attached documents, the conversation so far), and the ones the model produces, the output.

How many tokens per word? There is no universal rule: each publisher splits text its own way, with its own “tokenizer”. Three reference points, with their method:

• On this blog’s articles. The tokens in 12 pairs of articles, in English and French (17,489 words in English, 19,189 in French), were counted with seven tokenizers: OpenAI’s two, o200k_base and cl100k_base, and those published for Qwen3.8-27B, DeepSeek V4 Pro, Gemma 4 31B, Mistral Small 4 and gpt-oss-120b. Result: 1.25 to 1.31 tokens per word in English, 1.44 to 1.69 in French.
• At Anthropic. According to its documentation, a million tokens is roughly 555,000 words with the tokenizer of current Claude models.
• At Google. Its documentation counts about 4 characters per token, and 100 tokens for 60 to 80 English words.

Take these as an order of magnitude, not a rule. Our article pairs are adaptations, not word-for-word translations: the French versions run about 10% longer, and each French word takes more tokens. A bilingual business pays a little more for its French conversations.

Five mechanisms then make up the bill:

• Output costs more than input. In our tables, from 3 to 6 times more depending on the model, most often 4 to 5 times. A short, precise answer is the first saving.
• Reasoning is billed as output. Models that “think” before answering produce reasoning tokens you do not always see, which OpenAI, Anthropic and Google bill as output. GPT-6 Luna, Claude Haiku 5.5 and DeepSeek V4.1 Flash reason by default. On GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, GLM-5.3, GLM-5.3-Flash and Kimi K3 reasoning cannot be switched off, and Claude Sonnet 5.5 only lets you turn off up-front thinking.
• Cached input costs far less. When every request starts with the same block — your instructions, your catalogue, your manual — most providers keep it in memory for a limited time, five minutes or an hour at Anthropic for example, and charge it again at a fraction of the normal price: that is the “Cached input” column in the price table. Writing that block to the cache can cost a little more the first time.
• Batch processing is half price. If the answer can wait — product descriptions to write, invoices to sort overnight — the batch APIs of OpenAI, Anthropic and Google charge half the normal price.
• Very long requests cost more with several providers, above a threshold shown under the price table.

The context window is how many tokens a single request can hold, question, documents and history included. Most recent models accept around a million tokens, several hundred thousand words; output is capped far lower.

Context size, maximum output and model weights

Context size, maximum output and model weights

Model	Context	Max output	Weights
Claude Opus 5.5Anthropic	1M	128K	Closed
Claude Sonnet 5.5Anthropic	1M	128K	Closed
Claude Haiku 5.5Anthropic	1M	128K	Closed
GPT-6 AstraOpenAI	1.05M	128K	Closed
GPT-6.1 SolOpenAI	1.05M	128K	Closed
GPT-6 LunaOpenAI	1.05M	128K	Closed
Gemini 4 ArgonGoogle	1.05M	65.5K	Closed
Gemini 3.8 FlashGoogle	1.05M	65.5K	Closed
Muse Spark 1.3Meta	1.05M	not published	Closed
Grok 4.7xAI	500K	not published	Closed
Mistral Large 4 previewMistral AI	1M	not published	Open · licence not stated
DeepSeek V4.1 FlashDeepSeek	1M	384K	Open · MIT

A large context does not mean you should fill it: every token you send is paid for, on every request.

What do the main models cost?

Here are the public API prices for the models this guide follows, in dollars per million tokens. A million tokens is about 760,000 to 800,000 words of English on our measurements.

API prices, in US dollars per million tokens

API prices, in US dollars per million tokens

Model	Input	Cached input	Output
Claude Opus 5.5Anthropic	$4.00	$0.20	$20.00≈ 11,721 FCFA
Claude Sonnet 5.5Anthropic	$2.00	$0.10	$10.00≈ 5,860 FCFA
Claude Haiku 5.5Anthropic	$0.10	$0.01	$0.50≈ 293 FCFA
GPT-6 AstraOpenAI	$10.00	$1.00	$50.00≈ 29,302 FCFA
GPT-6.1 SolOpenAI	$2.00	$0.10	$10.00≈ 5,860 FCFA
GPT-6 LunaOpenAI	$0.10	$0.01	$0.50≈ 293 FCFA
Gemini 4 ArgonGooglePrice until 30 November 2026	$2.00	$0.10	$10.00≈ 5,860 FCFA
Gemini 3.8 FlashGooglePrice until 31 December 2026	$0.75	$0.075	$3.75≈ 2,198 FCFA
Muse Spark 1.3Meta	$1.25	$0.15	$4.25≈ 2,491 FCFA
Grok 4.7xAI	$2.00	$0.50	$6.00≈ 3,516 FCFA
Mistral Large 4 previewMistral AI	$1.36	$0.14	$4.18≈ 2,450 FCFA
DeepSeek V4.1 FlashDeepSeek	$0.30	$0.006	$1.20≈ 703 FCFA

Read the notes under the table: some prices have an end date, shown under the model’s name (Gemini 4 Argon’s doubles on 1 December 2026); DeepSeek charges half price outside its peak hours; and Mistral Large 4, in public preview, may still change its terms.

What about the lesser-known models?

Qwen, GLM and Kimi get less attention than ChatGPT, yet GLM-5.3-Flash was among the four most used models on OpenRouter in the week to 22 October, as you will see below. Their prices, from cheapest to dearest output:

API prices of other models, in US dollars per million tokens

API prices of other models, in US dollars per million tokens

Model	Input	Cached input	Output
Qwen3.8-FlashQwen (Alibaba Cloud)	$0.15	not published	$0.47≈ 275 FCFA
GLM-5.3-FlashZ.ai	$0.15	$0.03	$0.50≈ 293 FCFA
Mistral Small 4Mistral AI	$0.15	$0.015	$0.60≈ 352 FCFA
Gemini 3.1 Flash-LiteGoogle	$0.25	$0.025	$1.50≈ 879 FCFA
Qwen3.7-PlusQwen (Alibaba Cloud)	$0.40	$0.08	$1.60≈ 938 FCFA
DeepSeek V4 ProDeepSeek	$1.32	$0.044	$3.96≈ 2,321 FCFA
GLM-5.3Z.ai	$1.40	$0.26	$4.40≈ 2,579 FCFA
Qwen3.8-MaxQwen (Alibaba Cloud)	$2.00	not published	$6.00≈ 3,516 FCFA
Kimi K3.1 ProMoonshot AI	$2.00	$0.20	$10.00≈ 5,860 FCFA
Kimi K3Moonshot AI	$3.00	$0.30	$15.00≈ 8,791 FCFA

Open-weight models are also chosen by their licence. Read it before any commercial use:

Context, maximum output and licence of open-weight models

Context, maximum output and licence of open-weight models

Model	Context	Max output	Weights
DeepSeek V4.1 FlashDeepSeek	1M	384K	Open · MIT
DeepSeek V4 ProDeepSeek	1M	384K	Open · MIT
Mistral Large 4 previewMistral AI	1M	not published	Open · licence not stated
Mistral Medium 3.5Mistral AI	256K	not published	Open · Modified MIT
Mistral Small 4Mistral AI	256K	not published	Open · Apache 2.0
Kimi K3.1 ProMoonshot AI	1.05M	131K	Open · licence not stated
Kimi K3Moonshot AI	1.05M	131K	Open · Kimi K3 License
GLM-5.3Z.ai	1M	128K	Open · glm-5.3
GLM-5.3-FlashZ.ai	1M	128K	Open · MIT

What would a WhatsApp assistant cost a Douala distributor?

Take an example: a drinks distributor in Douala receives 400 WhatsApp messages a day, in English, French and Pidgin, 26 days a month. An assistant drafts each reply; a member of staff checks it and sends it. Each request sends a fixed block of 3,000 tokens — the catalogue, delivery zones and rules — then 500 tokens for the customer’s own message, and gets 250 tokens back.

Estimated monthly cost of 10,400 requests, each 3,500 tokens in and 250 out

Estimated monthly cost of 10,400 requests, each 3,500 tokens in and 250 out

Model	Per month (USD)	FCFA	EUR
GPT-6 LunaOpenAI	$2.13	1,249 FCFA	€1.90
Claude Haiku 5.5Anthropic	$2.13	1,249 FCFA	€1.90
Mistral Small 4Mistral AI	$2.81	1,646 FCFA	€2.51
GLM-5.3-FlashZ.ai	$3.02	1,768 FCFA	€2.69
DeepSeek V4.1 FlashDeepSeek	$4.87	2,852 FCFA	€4.35
Gemini 3.1 Flash-LiteGoogle	$5.98	3,505 FCFA	€5.34
Gemini 3.8 FlashGooglePrice until 31 December 2026	$15.99	9,371 FCFA	€14.29
Mistral Large 4 previewMistral AI	$22.31	13,073 FCFA	€19.93
Claude Sonnet 5.5Anthropic	$39.52	23,160 FCFA	€35.31
GPT-6.1 SolOpenAI	$39.52	23,160 FCFA	€35.31
Gemini 4 ArgonGooglePrice until 30 November 2026	$39.52	23,160 FCFA	€35.31
Claude Opus 5.5Anthropic	$79.04	46,321 FCFA	€70.62

At this volume, the cheapest small models cost nearly twenty times less than Claude Sonnet 5.5 or GPT-6.1 Sol. With them, the monthly API bill even stays below the price of a single ChatGPT Plus or Claude Pro subscription.

That figure is a floor, for four reasons:

• it assumes the fixed block is read from the cache on every request; in practice the cache expires after a few quiet minutes, and writing to it costs a little more;
• reasoning tokens are not counted, yet reasoning cannot be switched off on GPT-6.1 Sol, Claude Opus 5.5 or GLM-5.3-Flash, Claude Sonnet 5.5 only lets you turn off up-front thinking, and GPT-6 Luna, Claude Haiku 5.5 and DeepSeek V4.1 Flash reason by default;
• tax, your bank’s exchange fees and paid tools such as web search come on top;
• the cost of a project lies mostly in integration, human review and maintenance, covered in our guide to AI in SMEs.

And an assistant that answers from your documents?

A second example: an internal assistant answers staff questions from a procedures manual of about 20,000 words in French, some 30,000 tokens, attached to every question. Count 1,500 questions a month, 1,000 tokens of question and history, and 800 tokens of answer.

Estimated monthly cost of 1,500 requests, each 31,000 tokens in and 800 out

Estimated monthly cost of 1,500 requests, each 31,000 tokens in and 800 out

Model	Per month (USD)	FCFA	EUR
GPT-6 LunaOpenAI	$1.20	703 FCFA	€1.07
Claude Haiku 5.5Anthropic	$1.20	703 FCFA	€1.07
DeepSeek V4.1 FlashDeepSeek	$2.16	1,266 FCFA	€1.93
Gemini 3.8 FlashGooglePrice until 31 December 2026	$9.00	5,274 FCFA	€8.04
Mistral Large 4 previewMistral AI	$13.36	7,827 FCFA	€11.93
Claude Sonnet 5.5Anthropic	$19.50	11,428 FCFA	€17.42
GPT-6.1 SolOpenAI	$19.50	11,428 FCFA	€17.42
Gemini 4 ArgonGooglePrice until 30 November 2026	$19.50	11,428 FCFA	€17.42
Claude Opus 5.5Anthropic	$39.00	22,856 FCFA	€34.84

Here the cache makes the bill: without it, the same work would cost about five times more on Claude Sonnet 5.5 or GPT-6.1 Sol. Well-built requests, fixed block first and question second, save more than the choice of model.

Which usage limits apply, on the API and in subscriptions?

There are two worlds: the API, where limits protect the provider and your budget, and subscriptions, where they ration use.

On the API: tiers that rise with your spending

Each provider puts your account in a tier that sets how many requests and tokens you can send per minute, and often how much you can spend per month. Your tier rises with your payment history. Past a limit, the API answers with a 429 error: your application must know how to wait and retry.

• OpenAI has three paid tiers, which rise with your total credit purchases, plus a free tier for allowed countries. Each has a monthly usage cap, and you can set your own spend limits (OpenAI’s limits).
• Anthropic places your organisation in a tier based on its history; a new organisation may start lower, in “evaluation”. At your tier’s monthly cap, the API stops until the 1st of the next month. Tokens read from the cache do not count toward the input limit on most models (Anthropic’s limits).
• Google applies limits per project, not per key, resets the daily quota at midnight Pacific time and, depending on your billing history, also caps spending per ten-minute window (Gemini’s limits).
• xAI sets your tier by cumulative spend since 1 January 2026 and never moves it down; cached tokens count toward its per-minute limit (xAI’s limits).
• DeepSeek limits how many requests can run at once per account; an increase is free on request (DeepSeek’s limits).

We do not copy these figures into a table: they depend on your tier and change often. Each provider’s page, listed at the end of this guide, is the reference.

In subscriptions: caps per session and per week

Consumer apps almost never publish a fixed message count. Claude counts usage per five-hour session, with a weekly cap. Google describes “compute-based” limits that refresh every five hours up to a weekly cap. OpenAI gives estimates per five-hour window for some tools.

Consumer subscriptions

Consumer subscriptions

Plan	Price	What you get	Limits
ChatGPT FreeOpenAI	Free	GPT-6 Luna and GPT-5 Thinking Mini; limited GPT-5.6 Terra in Work and Codex on desktop	Unlimited text chats, subject to abuse guardrails; limited uploads, images and voice
ChatGPT GoOpenAI	$8.00/ month≈ 4,688 FCFA	The Free models, with more messages, uploads, images and voice	Unlimited text chats, subject to abuse guardrails; this plan may include ads
ChatGPT PlusOpenAI	$20.00/ month≈ 11,721 FCFA	GPT-6 Sol and GPT-6 Luna in chat; GPT-6 Astra and GPT-6.1 Sol in Work and Codex, not in chat	OpenAI’s Codex estimate: 15 to 160 GPT-6.1 Sol messages per 5 hours; weekly limits may apply
Claude FreeAnthropic	Free	Help with everyday questions	Limited usage, with no fixed message count
Claude ProAnthropic	$20.00/ month≈ 11,721 FCFA	At least 5x the usage of Free per 5-hour session	A cap per 5-hour session, plus a weekly limit
Google AI PlusGoogle	$4.99/ month≈ 2,924 FCFA	2x the usage of the free plan, 400 GB of storage	Compute-based limits, not quantified
Google AI ProGoogle	$19.99/ month≈ 11,715 FCFA	4x the usage of the free plan, a one-million-token context, 5 TB	Compute-based limits, not quantified
Mistral Vibe ProMistral AI	$14.99/ month≈ 8,785 FCFA	Mistral’s assistant (France): more messages and searches, all-day coding	Fair usage, under Mistral’s terms
X Subscription PlusxAI	$8.00/ month≈ 4,688 FCFA	Grok, Cursor and X; the subscription’s base usage	Premium gives 4x this usage ($30), Super 15x ($100), Ultra 50x ($200)
Meta AI in WhatsAppMeta	not published	Meta’s assistant inside WhatsApp, with incognito chats	Features not yet available to every user, according to Meta

For a team, business plans cost a little more per person but change the rules on your data, as the data section below explains:

Team and business plans

Team and business plans

Plan	Price	What you get	Limits
ChatGPT BusinessOpenAI	$25.00/ month per user≈ 14,651 FCFA	A ChatGPT team workspace, from 2 users	The same five-hour estimates as Plus
Claude TeamAnthropic	$25.00/ month per user≈ 14,651 FCFA	More usage than Pro, in a team workspace	From 2 to 150 seats
Mistral Vibe TeamMistral AI	$24.99/ month per user≈ 14,645 FCFA	Higher limits, up to 30 GB of storage per user, data export	Fair usage, under Mistral’s terms
Microsoft 365 Copilot BusinessMicrosoft	$25.20/ month per user≈ 14,768 FCFA	Copilot, as an add-on to a Microsoft 365 Business plan	Some advanced features are billed by usage on top

Other assistants are within reach too. Microsoft says Copilot is free with a Microsoft account. DeepSeek offers its assistant on the web and as an app, and Moonshot AI offers its own at kimi.com. And Google’s AI Mode answers right inside Search: our guide to Google AI Mode shows how to use it.

Free plans: what do you really pay?

Nothing in money, often something in data. The Gemini API’s free tier is used to improve Google’s products, the paid tier is not (Gemini pricing). Meta’s “Contributor” mode sells its models at a steep discount in exchange for the right to train future models on your content (Meta pricing). And free apps, like personal subscriptions, may use your conversations for training unless you opt out.

Which model performs best?

That depends on what you measure, and three public leaderboards measure three different things. Read on 23 October 2026, except Arena (21 October):

• Capability, according to Artificial Analysis. This analysis firm runs its own battery of tests — agent tasks, coding, scientific reasoning, general knowledge — and turns them into an index. Its summary puts Claude Opus 5.5 and Claude Sonnet 5.5 at their highest setting on top, Opus 5.5 first, with Meta’s Muse Wonder Full, which Meta offers only in its Muse Code subscriptions, tied with Sonnet 5.5; next comes Moonshot’s Kimi K3.1 Pro, the best placed of the models whose weights are announced; GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 follow, tied. Artificial Analysis states that its index is mostly an English-language test suite (its method).
• Preference, according to Arena. People put one question to two anonymous models and vote for the better answer; the 20 October leaderboard combines nearly 9 million votes. Gemini 4 Argon is first there, with a score still marked “preliminary”. GPT-6 Astra, well placed at Artificial Analysis, is only 34th there, somewhere between 21st and 54th given the margin of error.
• Adoption, according to OpenRouter. This platform, which gives access to many models through one API, ranks them by tokens processed. In the week to 22 October, the most used were low-cost models: DeepSeek V4.1 Flash first, under both its names, then a free MiniMax model and GLM-5.3-Flash. Usage reflects price and developer habits, not quality.

Why do these leaderboards disagree? Because an agent task in English, a vote on a chat answer and a usage volume do not measure the same thing. And none measures your case: a customer asking in Pidgin what a crate costs delivered to Molyko.

The only reliable method takes four steps:

1. Collect 100 real customer messages, in English, French and Pidgin, with names and numbers removed.
2. Write the answer you would accept for each.
3. Have two or three models answer, one small and one mid-range, with the same instructions.
4. Count the correct answers, the wrong prices, the inventions and the total cost. Keep the cheapest that reaches your bar.

What happens to your data?

The general rule, as most publishers write it themselves: personal plans may train models on your conversations unless you opt out; most paid APIs and business plans do not by default. In detail:

• OpenAI may train its models on ChatGPT conversations; you can opt out, and temporary chats are not used. Its business plans and API are not used for training by default (ChatGPT data controls, API).
• Anthropic lists model training as “Opt-out” for Claude Free, Pro and Max, and as none by default for Team and Enterprise (Claude plans). API data is never used for training without your express permission (Anthropic).
• Google uses Gemini app chats to improve its services while the activity setting is on, with human review of some of them (Gemini privacy). For the API, see above: free, yes; paid, no.
• xAI may train Grok on your content unless you untick “Improve the model” in the settings (xAI FAQ); its API never trains on your requests without explicit permission (API security).
• Mistral AI trains on Vibe conversations unless you opt out, on the Team plan too, where an administrator can switch it off for everyone; Enterprise is opted out by default, and the API has a separate setting to switch off (Mistral).
• Meta offers “incognito” chats with Meta AI in WhatsApp, which are not kept once closed (Meta AI in WhatsApp).
• DeepSeek writes that data from its services is collected, processed and stored in the People’s Republic of China and used to train its models; a right to opt out of that training is among the rights it says “may be available” to you (privacy policy).

In Cameroon, Law No. 2024/017 of 23 December 2024 relates to personal data protection. Before you send customer data to an AI service hosted abroad, have a lawyer check what it requires in your case.

If your customers are in the European Union, diaspora customers in France or Belgium included, the GDPR can apply to you even with no establishment there. OpenAI can process data in Europe, for 10% more on models released on or after 5 March 2026 (OpenAI pricing). Our guide to AI in SMEs covers these laws in more depth.

How should you choose for your needs and budget?

A simple rule: start with the cheapest model that passes your test, and move up only for the cases where it fails.

Which dates should you watch in the coming months?

• By 31 October 2026: Mistral says it will publish the weights of Mistral Large 4.
• By 4 November 2026: Moonshot says it will publish the weights of Kimi K3.1 Pro, naming no licence.
• 31 October 2026, 5 pm in Douala (1 November in Beijing): DeepSeek starts billing the web searches and page fetches that coding tools such as Claude Code or Codex make through its API.
• 12 November 2026: Anthropic’s supported regions policy applies; Cameroon is on its list.
• 30 November 2026: last day of Gemini 4 Argon’s current price, which doubles the next day.
• 31 December 2026: last day of Gemini 3.8 Flash’s current price, which doubles the next day, and of Microsoft’s discount on the Microsoft 365 Copilot Business add-on.
• No date: Qwen3.7-Plus shows a “limited-time” discount with no end date.
• 22 September 2027: the earliest retirement date for Claude Opus 5.5, according to Anthropic. No model lasts forever: plan to switch.

How can Nematek help?

Our AI and automation service starts from a task your team already does — reminders, reports, copy-paste between two systems — and automates it, AI assistants included, without overturning everything else.

AI consulting and development designs something custom: an agent that chains actions, with guardrails suited to the use case, or an assistant that answers from your own documents in your team’s language, connected to your systems through APIs. The model, the data used and the controls are chosen during scoping, with privacy, connectivity and in-house skills in mind.

To run an open-weight model yourself, our cloud, DevOps and DevSecOps service designs and operates architectures on AWS, Azure, Google Cloud or OCI, with containers, Kubernetes, monitoring and backup.

Our IT training covers AI engineering, with a trainer who comes to your premises, in English or French.

See also Automatisation, BI & Intégrations Nematek, our automation, BI and integration offering. To scope your first use case, describe it in a free assessment request.

How is this guide kept up to date?

The figures in this guide are not written into the text: they come from one data file per provider, shared by the English and French versions. Every month, and on each of the dates listed above, we reopen each provider’s official page and compare every row.

Under each table, “Checked” gives the date of the oldest check among its rows, so a figure never looks newer than it is. When a price changes, the old one is kept in our data, the “Updated” date moves, and a line is added to the update history at the bottom of the page. Something looks out of date? Tell us through the contact page.

Official pricing and limits pages

Official pricing and limits pages

Provider	Pricing	Limits	Checked
AnthropicUnited States	Pricing	Limits	23 Oct 2026
OpenAIUnited States	Pricing	Limits	23 Oct 2026
GoogleUnited States	Pricing	Limits	23 Oct 2026
MetaUnited States	Pricing	Limits	23 Oct 2026
xAIUnited States	Pricing	Limits	23 Oct 2026
Mistral AIFrance	Pricing	Limits	23 Oct 2026
DeepSeekChina	Pricing	Limits	23 Oct 2026
Qwen (Alibaba Cloud)China	Pricing	Limits	23 Oct 2026
Z.aiChina	Pricing	not published	23 Oct 2026
Moonshot AIChina	Pricing	Limits	23 Oct 2026