Over the course of September, new models came from Anthropic, OpenAI, Google and xAI, and Chinese vendors are keeping pace. We compared 16 models on price, coding performance and fit for everyday business work, and added a framework for choosing a model by task.
Why “just buy the most expensive one” no longer works
Picture the CIO of a logistics company. One development team pays for an AI assistant on its own initiative, another is trying a different one, marketing uses a chatbot for copy, and finance uses nothing because nobody knows where company data is allowed to go. Invoices arrive from three places, and nobody can say exactly what AI costs the company or what it delivers.
Many companies end up here, because new models arrive faster than most teams can track and properly evaluate them. Within three weeks of September, Google released Gemini 3.8 Flash, OpenAI shipped GPT-6 Astra followed by the smaller Sol and Luna, xAI launched Grok 4.7 and Anthropic released Claude Opus 5.5.
Many teams once went straight for the most capable model, assuming better output would justify the price. That logic no longer holds. Prices for the same number of tokens now differ by up to a hundred times, and on some tasks a pricier model delivers no better result. Each table lists its source so you can verify the figures, which matters because prices and results change very quickly.
LLM comparison for business: which models are worth your attention
The market has split into three groups. US vendors (Anthropic, OpenAI, Google, xAI) top the performance leaderboards, and their flagship (frontier) models cost more than most others in this overview. France’s Mistral focuses on open models and positions its European base as an advantage. And Chinese vendors offer open-weight models that are mostly far cheaper and come close to the leaders on many tasks.
TAB. 1 Flagship models by vendor (as of September 24, 2026)
| Vendor | Model | Released | Context (tokens) | Weights |
|---|---|---|---|---|
| Anthropic | Claude Opus 5.5 | Sep 22, 2026 | 1 million | closed |
| OpenAI | GPT-6 Astra | Sep 3, 2026 | 1.05 million | closed |
| Gemini 3.8 Flash | Sep 2, 2026 | 1 million | closed | |
| xAI | Grok 4.7 | Sep 21, 2026 | 500,000 | closed |
| Mistral | Mistral Medium 3.5 | Apr 28, 2026 | 256,000 | open |
| DeepSeek | DeepSeek V4.1 Flash | Sep 10, 2026 | 1 million | open |
| Alibaba | Qwen3.8 Max | Aug 3, 2026 | 1 million | open |
| Moonshot AI | Kimi K3 | Jul 16, 2026 | 1 million | open |
| Zhipu AI | GLM-5.3 | Aug 18, 2026 | 1 million | open |
| MiniMax | MiniMax M3 | Jun 1, 2026 | 1 million | open |
Context is how much text a model can take in at once: one million tokens is roughly 750,000 English words, enough for the code of a smaller application plus its documentation. Open weights mean you can download the model and run it on your own infrastructure. That isn’t always open source in the strict sense. Licenses vary from model to model.
Most vendors offer smaller, cheaper variants next to the flagship. Anthropic offers the cheaper Claude Sonnet 5 and Haiku 4.5, while Claude Fable 5.1 costs even more than Opus. OpenAI offers GPT-6 Sol and Luna alongside Astra. Watch the names: GPT-5.6 Terra is not part of the GPT-6 family, it belongs to the previous generation.
Token prices, and why the price list doesn’t tell the whole story
Models are billed per token. Input is what you send the model, output is what it generates. Every vendor charges several times more for output than for input.
TAB. 2 Price per million tokens in USD, from official vendor price lists (as of September 24, 2026)
| Model | Input | Cached input | Output |
|---|---|---|---|
| Claude Fable 5.1 | 10 | 0.25 | 50 |
| Claude Opus 5.5 | 4 | 0.20 | 20 |
| Claude Sonnet 5 | 2 | 0.20 | 10 |
| Claude Haiku 4.5 | 1 | 0.10 | 5 |
| GPT-6 Astra | 10 | 1 | 50 |
| GPT-6 Sol | 2 | 0.20 | 10 |
| GPT-6 Luna | 0.10 | 0.01 | 0.50 |
| Gemini 3.8 Flash | 0.75 | 0.075 | 3.75 |
| Gemini 3.5 Flash-Lite | 0.30 | 0.03 | 2.50 |
| Grok 4.7 | 2 | 0.50 | 6 |
| Mistral Medium 3.5 | 1.50 | 0.15 | 7.50 |
| DeepSeek V4.1 Flash | 0.30 | 0.006 | 1.20 |
| Qwen3.8 Max | 1.65 | 0.21 | 4.95 |
| Kimi K3 | 3 | 0.30 | 15 |
| GLM-5.3 | 1.40 | 0.26 | 4.40 |
| MiniMax M3 | 0.30 | 0.06 | 1.20 |
Prices come from the official price lists: Anthropic, OpenAI, Google, xAI, Mistral, DeepSeek, Alibaba Cloud, Moonshot AI, Z.ai and MiniMax. Gemini 3.8 Flash has an introductory price until December 31, 2026, after which it doubles to 1.50 and 7.50 USD. DeepSeek lists its peak-hour price and charges half off-peak. Qwen prices follow Alibaba Cloud’s global price list and differ by region; in Singapore, for example, they are 2 and 6 USD. OpenAI, xAI and MiniMax charge a higher rate once a single request exceeds 272,000, 200,000 and 512,000 tokens respectively. The table shows cache reads; some vendors bill cache writes separately.
The gap in list prices is large. Output costs 0.50 USD per million tokens for GPT-6 Luna and a hundred times as much for GPT-6 Astra or Claude Fable 5.1. Even so, list prices don’t show what it costs to complete a task.
Prompt caching can cut the bill substantially. When repeated requests include the same context, such as a project’s source code or company guidelines, the vendor stores that context for a while and bills later reads at a fraction of the price. For Claude Opus 5.5 that’s a twentieth of the regular input price, for DeepSeek a fiftieth. In coding, where an agent reads the same code over and over, cache can make up most of the input tokens. Moonshot AI reports a share above 90% for coding on its own API; other tools and workloads will look different.
Token usage matters too, since a cheaper model may use more tokens to complete the same task, or it may need a second attempt. That’s why it pays to calculate the cost of a finished task, not of a million tokens.
Coding performance: what benchmarks measure and what they don’t
Coding is easier to measure than writing, because code either passes the tests or it doesn’t. Scores still depend on the test design, the agent tool and the runtime setup.
For a long time, the yardstick was SWE-bench Verified, five hundred real tasks from open-source projects on GitHub. But top-tier models now score 95 to 99% on it, and Vals.ai archived its leaderboard on September 1, 2026, because it no longer tells models apart. Vendors still like to quote it in their announcements.
Three other measures separate models better today:
- Terminal-Bench 4.0 gives an agent command-line tasks, from configuring a server to fixing a build. Vals.ai runs it for all models in the same environment.
- SWE-Bench Pro from Scale AI contains harder tasks. The table shows its public hard set from open-source repositories; commercial code sits in a private subset. The agent tool a model works in has a big effect on the result.
- Artificial Analysis Intelligence Index combines results from ten evaluations into a single number and works for general comparison, not just code.
TAB. 3 Results on independent leaderboards as of September 22 and 23, 2026
| Model | Terminal-Bench 4.0 | SWE-Bench Pro, hard set | Intelligence Index |
|---|---|---|---|
| Claude Opus 5.5 | 61.6% | not listed | 58 |
| Claude Fable 5.1 | 49.5% | 92.2% | 53 |
| Claude Sonnet 5 | 8.1% | 88.2% | 38 |
| GPT-6 Astra | 57.1% | 90.2% | 53 |
| GPT-6 Sol | not listed | not listed | 48 |
| GPT-6 Luna | not listed | not listed | 37 |
| Gemini 3.8 Flash | 13.1% | 58.8% | 41 |
| Grok 4.7 | 28.3% | not listed | 46 |
| GLM-5.3 | 25.3% | 84.3% | 45 |
| Qwen3.8 Max | 24.8% | not listed | 45 |
| Kimi K3 | 12.6% | 88.2% | 44 |
| DeepSeek V4.1 Flash | 11.6% | not listed | 39 |
| MiniMax M3 | not listed | not listed | 29 |
| Mistral Medium 3.5 | not listed | not listed | 14 |
Sources: Vals.ai (Terminal-Bench 4.0, standardized run), Scale Labs (SWE-Bench Pro V2, public hard set), Artificial Analysis. “Not listed” means the model isn’t on that leaderboard yet. We didn’t estimate or fill in anything. The SWE-Bench Pro figure belongs to a model inside a specific agent tool and setting (Claude Code, Codex or mini-swe-agent), not to the model alone.
The table shows Claude Opus 5.5 and GPT-6 Astra in the lead, but their order depends on who measured and how. Anthropic’s own announcement puts Opus 5.5 at 66.4% on Terminal-Bench, while the independent Vals.ai run shows 61.6%. So always check whether a number comes from the vendor or from an independent source.
Notice also how the numbers diverge across benchmarks. Kimi K3 scores the same as Claude Sonnet 5 on SWE-Bench Pro, yet reaches a fraction of the leaders’ score on Terminal-Bench. No benchmark represents every kind of coding work.
Before standardizing on a model, build your own test. Give two or three candidates ten to twenty real tasks from recent sprints, then compare both the results and the cost.
What a single real task costs
To make it concrete, we priced two example tasks. The first is an agent fixing a bug in an existing repository: it reads the code, finds the cause, writes the fix and a test. The second is generating a thousand product descriptions for an online store.
TAB. 4 Example cost of two tasks in USD
| Model | Agent bug fix | 1,000 product descriptions |
|---|---|---|
| Claude Fable 5.1 | 5.45 | 20.00 |
| GPT-6 Astra | 6.80 | 20.00 |
| Claude Opus 5.5 | 2.36 | 8.00 |
| Kimi K3 | 2.04 | 6.00 |
| Grok 4.7 | 1.66 | 2.80 |
| Claude Sonnet 5 | 1.36 | 4.00 |
| GPT-6 Sol | 1.36 | 4.00 |
| Mistral Medium 3.5 | 1.02 | 3.00 |
| GLM-5.3 | 1.01 | 2.02 |
| Qwen3.8 Max | 1.00 | 2.31 |
| Gemini 3.8 Flash | 0.51 | 1.50 |
| MiniMax M3 | 0.24 | 0.51 |
| DeepSeek V4.1 Flash | 0.14 | 0.51 |
| GPT-6 Luna | 0.07 | 0.20 |
We assume the bug fix uses 2 million input tokens, 90% of them from cache, and 60,000 output tokens, spread over dozens of requests, none of which crosses the threshold for the higher rate. One description means a 500-token prompt and 300 tokens of text. This is our illustrative assumption, not a measurement; real usage varies with the tool, the codebase and the task. Prices follow TAB. 2, excluding tax, batch discounts and cache-write fees.
The table shows the cost of one attempt, not of finished work. A cheap model that fixes the bug on the second try, or not at all, ends up costing more than an expensive one that gets it right the first time. And the developer’s time spent reviewing the result often costs more than the tokens themselves.
For product descriptions, a lower-cost model is often sufficient, and at scale the difference between 0.20 and 20 USD per thousand items becomes material. Product specs and claims still need checking with any model, because any of them can make a mistake.
For development teams there’s a second route: a subscription. Claude Team costs 25 USD per user per month, or 125 USD with higher limits for heavy use, and includes Claude Code. ChatGPT Business has the same price tiers and includes Codex. Paid annually, both come out a fifth cheaper. Gemini Code Assist starts at 19 USD with an annual commitment. For a team that works with an agent all day, a subscription is usually more predictable than paying per token.
Office, marketing, finance: where a cheaper model will do
Outside software development, many routine tasks usually call for low cost and consistent output rather than flagship-level capability. The differences between models are smaller here than in coding.
Office and documents. Meeting summaries, email triage, draft replies, turning an invoice into a spreadsheet. GPT-6 Luna, Gemini 3.5 Flash-Lite and Claude Haiku 4.5 are good options. On simple tasks the quality gap with flagship models is usually small, while the price is several times lower, and a hundred times lower for GPT-6 Luna compared with Astra. Before rolling one out, test it on a sample of your own documents in your own language.
Marketing. For product descriptions, ad copy variants or translations, volume decides, so a cheap model makes sense. For brand-sensitive work, such as an expert article or a client presentation, check whether candidates keep your terminology and tone across several drafts. Stronger models tend to hold the tone better, keep facts straight and slip into clichés less often.
Finance. Finance teams use models to analyze statements, review contracts and prepare management materials, so accuracy and data location both matter. For more complex analysis, Claude Opus 5.5, GPT-6 Sol or Gemini 3.8 Flash make sense. For routine extraction of figures from documents, a smaller model will do, as long as someone checks the output.
For many companies, model choice isn’t the main problem. Teams use AI independently, and no one tracks what it costs or what it delivers. In an earlier article we wrote about why companies often don’t measure the benefits of AI.
Chinese and open models: where you save and where you take risks
Chinese models are serious competition today. GLM-5.3 and Qwen3.8 Max score 45 on the Artificial Analysis index, more than Gemini 3.8 Flash, and cost a fraction of US frontier models. DeepSeek V4.1 Flash costs 0.30 USD per million input and 1.20 USD per million output tokens; in our overview only GPT-6 Luna is cheaper.
For a business, though, the question isn’t whether they’re good enough but where the data will flow. When you call a Chinese vendor’s API directly, your data is processed on its servers and governed by its terms. For a bank, an insurer or an energy company that’s usually unacceptable; for an internal tool working with public data, maybe not.
Open weights offer a third way, because Kimi K3, GLM-5.3, DeepSeek, Qwen and MiniMax can all be downloaded and run on your own hardware or with a European infrastructure provider. On your own hardware the data never leaves the company; with a European provider it stays in the region you chose and under the contract you signed. Some of these models are also available on AWS Bedrock, where they run in regions you select.
Self-hosting has operational costs: a large model needs powerful GPUs, regular updates and staff to maintain them. Licenses vary from model to model, and Kimi K3, Qwen3.8 Max and MiniMax M3 don’t use standard open-source licenses. The MiniMax M3 license also restricts commercial use, and larger companies need separate permission from the vendor. The European alternative is Mistral, which offers open models as well as a hosted API of European origin.
Direct, through OpenRouter, or through the cloud?
You can pay for a model directly with the vendor, through an aggregator such as OpenRouter, or through the cloud platform your company already uses. The routes differ in extra cost, how fast new versions arrive and where the data is processed.
TAB. 5 Three ways to access models
| Route | When it makes sense | Extra cost | Data |
|---|---|---|---|
| Vendor’s direct API | You need the newest version right away and work with one or two models | none | per vendor terms |
| OpenRouter | You’re testing several models, want a fallback during outages, or want to route by price | 5.5% on credit purchases, minimum 0.80 USD, no markup on tokens | OpenRouter doesn’t store prompts by default, but each provider has its own rules |
| AWS Bedrock, Google Vertex AI, Microsoft Azure | Your company is already in the cloud and needs unified procurement, access management and region choice | depends on platform and region | in the chosen region under your cloud contract; per AWS, Google and Microsoft, not used to train models |
OpenRouter is most useful during evaluation, when one API key lets a team compare hundreds of models on its own tasks without managing ten vendor accounts. It adds no markup on tokens; its fee is 5.5% on credit purchases, with a minimum of 0.80 USD.
Watch the data, though. OpenRouter itself doesn’t store prompts by default, but it forwards the request to a provider whose rules may differ. You can guard against that with a setting that excludes providers that store data or train on it. Make that restriction part of your default account policy and review it as provider rules change.
Large companies often use cloud platforms because they fit existing contracts, security rules and accounting processes. Availability still varies by model and platform. Claude Opus 5.5 was on AWS, Google Cloud and Microsoft Foundry from day one, while cloud catalogs of smaller and Chinese models tend to be narrower, often without the newest versions.
How to choose: a decision framework
Choosing a model comes down to five questions. Answer them for each task separately, not for the whole company at once.
- How sensitive is the data? Personal data, trade secrets and regulated data narrow the choice more than anything else.
- How complex is the task? An agent working independently in code, or multi-step analysis, needs a top-tier model, while triage and summaries can make do with a cheaper one.
- What is the volume? At thousands of requests a day, unit cost carries more weight; at dozens, output quality does.
- What does a mistake cost? If a person reviews the output, you can afford a cheaper model. If the output goes straight to a client, a stronger model pays off.
- Who will run it? Running an open model yourself saves on token fees but requires staff time and hardware.
TAB. 6 Decision matrix by task type
| Task | Data sensitivity | Recommended class | Example models |
|---|---|---|---|
| Agentic development, large codebases | medium | top tier | Claude Opus 5.5, GPT-6 Astra |
| Everyday coding, code review | medium | balanced model | Claude Sonnet 5, GPT-6 Sol, Gemini 3.8 Flash |
| Analysis and finance | high | balanced model via cloud or self-hosted | Claude Opus 5.5, GPT-6 Sol, GLM-5.3 on your own infrastructure |
| Office work, email, summaries | low to medium | low-cost model | GPT-6 Luna, Gemini 3.5 Flash-Lite, Claude Haiku 4.5 |
| Bulk copy and translation | low | lowest-cost model | GPT-6 Luna, DeepSeek V4.1 Flash, MiniMax M3 |
| Data that must not leave the company | very high | open model on your own infrastructure | GLM-5.3, Qwen3.8, Mistral |
One model rarely covers a whole company, so a sensible setup has two or three: a powerful one for development and demanding analysis, a cheap one for volume work and possibly an open one for data that must stay in-house. The five questions above decide which model goes where.
Our development team has moved to Claude Code as its primary tool. We chose it as a complete development tool, not for the model alone: its agent works directly in the repository, runs tests and proposes changes for a developer to approve.
If you’re working out how to bring AI into development or business processes, take a look at our services. We’ll help you choose models for specific tasks, connect them to your systems and define where your data is processed. You’ll find related topics in our articles on AI agents, the MCP protocol and how not to overpay for LLMs.
FAQ
What is the best AI model for coding? According to independent leaderboards as of September 2026, Claude Opus 5.5 and GPT-6 Astra lead. On Terminal-Bench 4.0 as measured by Vals.ai, Opus 5.5 scores 61.6% and Astra 57.1%. For everyday coding, the cheaper Claude Sonnet 5 or GPT-6 Sol is often enough.
How much does it cost to use AI models through an API? Prices range from 0.10 USD per million input tokens for GPT-6 Luna to 10 USD for GPT-6 Astra and Claude Fable 5.1. Output costs several times more. The real cost depends mostly on cache usage and the number of attempts per task.
Is it safe to use Chinese AI models in a business? It depends on where the model runs. Through a Chinese vendor’s API, data is processed on the vendor’s servers. Open-weight models such as GLM-5.3, Qwen or DeepSeek, however, can run on your own infrastructure or in a European cloud, where data stays in the environment you choose.
Is OpenRouter worth it for companies? For testing and comparing models, yes: one account gives you access to hundreds of models with no markup on token prices. With sensitive data, you need to restrict which providers requests can go to, because each has its own data retention rules.
What is the difference between GPT-6 Astra, Sol and Luna? Astra is the high-capability, high-price variant for complex coding and agentic tasks. Sol is the balanced mid-tier, and Luna the low-cost option for volume work. GPT-5.6 Terra, which is sometimes confused with them, belongs to the previous generation.
Is one AI model enough for a company? Usually not. For most companies, a setup of two or three models chosen by task complexity, volume and data sensitivity makes sense.
Should we pay for AI per token or by subscription? For developers who work with AI all day, a subscription such as Claude Team or ChatGPT Business, from 20 USD per user per month billed annually, is usually more predictable. Automated processes and applications pay per token through the API.
