2026 AI Model Guide: ChatGPT, Claude, Gemini, Grok

As of July 27, 2026, here’s a breakdown of the current landscape for the four leading AI assistants: a look at each version, their costs, and which users they’re best suited for.

TL;DR: the short answers

• Most people should start free. ChatGPT, Claude, Gemini, and Grok all have credible free tiers. Gemini’s remains the most generous, and Claude’s free tier still runs Sonnet 5.
• For coders and knowledge workers, Claude Opus 5 is the story of the month: Anthropic’s own benchmarks put it ahead of the pricier Fable 5 on most coding and knowledge-work tests, at $5/$25 per million tokens instead of $10/$50.
• For ChatGPT users, the gate is gone. GPT-5.6 Sol is live in the app on Plus and above, and Free and Go users can reach the mid-tier Terra through ChatGPT Work and Codex.
• For people who live in Gmail, Docs, and Drive, Gemini is still the obvious pick, but Google is now the only major lab without a current-generation flagship in production.
• For real-time news and market chatter, Grok is unmatched. Grok 4.5 arrived July 8 at $2/$6, trained jointly with the code editor Cursor, and is the token-efficiency story of the quarter.
• For developers watching API bills, the cheap end has moved: Grok 4.5 at $2/$6, Gemini 3.6 Flash at $1.50/$7.50, and GPT-5.6 Luna at $1/$6 are all credible. Claude Sonnet 5’s introductory $2/$10 ends August 31.
• The general rule for 2026: match the assistant to your ecosystem, match the tier to the task, and match the effort level to the difficulty. Most people stop after step one.

What changed in July, and why it matters

Two structural shifts define this landscape. Both got sharper this month.

Shift one: the flagship is no longer where the value sits

These days, most companies offer several versions of their products instead of just one. Recently, in July, the mid-range options proved to be a better value for money than the most expensive ones in most cases – at three out of four companies, specifically.

Anthropic’s new Opus 5 model is nearly as intelligent as Fable 5, but costs significantly less. In fact, Anthropic’s data shows Opus 5 actually outperformed Fable 5 on most tests. OpenAI’s Sol model is more powerful than Terra and Luna, but they position Terra as offering performance similar to GPT-5.5 at a much lower price. Google released a more affordable and efficient Flash model while continuing to refine its top-tier model. Previously, the focus was on building the largest possible AI model. Now, the key metric is how much it costs to complete a given task, and AI companies are openly acknowledging this shift.

Shift two: the effort dial

In 2025, improving AI involved changing to different models. But by 2026, it was becoming more about fine-tuning a single model to get better results.

The “effort” setting controls how much thinking the AI does before responding. Lower effort means faster, less expensive responses – like getting a quick opinion from a coworker. Higher effort lets the AI spend more time carefully reasoning through the problem, checking its work, and considering different solutions. This results in better answers for complex tasks but uses more resources and takes longer.

Claude’s Opus model now offers a range of processing speeds – from low to maximum. OpenAI has introduced a ‘max’ reasoning setting and an ‘ultra’ mode that distributes tasks to multiple AI agents working in parallel. Gemini allows users to adjust its ‘thinking level,’ including a ‘Deep Think’ option. And Grok automatically uses reasoning and lets developers control the amount of effort it puts into each task through its API.

These days, choosing an AI model isn’t just about *which* one is best, but also about balancing performance with cost. Anthropic claims their Opus 5 model, when running at its highest setting, achieves nearly the same results as Fable 5 – with half the cost per task. OpenAI states their Sol model, at maximum reasoning power, scores almost as well as Fable 5 on a key benchmark, but completes tasks 61% faster. While these figures come from the companies themselves, the trend is clear: you can now actively adjust the effort and cost of these models to find the right balance for your needs.

Here’s how to choose an AI model, and the effort level needed, using Claude

The four families at a glance

As of July 27, 2026. Lineups are changing monthly. Treat this as a snapshot.

Just a quick update on names: Anthropic’s Mythos 5, which has limited access, is more powerful than Fable 5. Also, xAI has officially become SpaceXAI after being brought back under the SpaceX umbrella in February and undergoing a rebrand on July 6th. The name of their product, Grok, remains the same.

The Claude spotlight: Opus 5 changes the math

What Opus 5 actually is

On July 24th, Anthropic launched Claude Opus 5, their latest AI model. This is the fourth new model they’ve released in just two months, following Mythos 5, Fable 5, and Sonnet 5. It’s now used as the standard model for both Claude Max and Claude Pro, offering the highest performance available. You can also access it through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry under the name ‘claude-opus-5’.

The main advantage of Opus 5 isn’t necessarily its raw power, but rather its pricing. It costs the same as the previous version, Opus 4.8 – $5 per million input tokens and $25 per million output – which is half the price of Fable 5. According to Anthropic’s testing on thirteen different measures, Opus 5 performed better than Fable 5 in most cases (eight out of thirteen). Specifically, on tests involving physics, chemistry, and cryptography, Opus 5 scored 43.3% compared to Fable 5’s 33.7%, all while costing less per task. It also achieved a score on the ARC-AGI 3 problem-solving test that was roughly three times higher than any other model.

Four practical details matter more than the benchmark table:

1. Fast mode. Opus 5 can run at about 2.5 times the default speed for double the base price. Useful when wall-clock time is the constraint, expensive if you leave it on.
2. Thinking is on by default, and disabling it now returns an error at xhigh or max effort. If you pinned a model ID and a config, re-test before swapping.
3. Lighter cyber classifiers. Opus 5 permits vulnerability discovery in source code while blocking binary scanning, penetration testing, and exploit generation. Anthropic expects its classifiers to intervene around 85 percent less often than Fable 5’s do, with flagged requests falling back to Opus 4.8. For security-adjacent development work, this is a meaningful reduction in friction compared with Fable.
4. No 30-day data retention. Consistent with prior Opus models, Opus 5 carries no data retention requirement for general access. Fable 5 does retain inputs and outputs as a safety measure, which matters if you have a hard zero-retention policy.

Anthropic says Opus 5 is currently their safest and most reliable model – it showed the least amount of misleading responses during safety testing. It’s also their best publicly available option for researchers doing scientific work. However, a more restricted model called Mythos 5 is still better at tasks related to cybersecurity defense, and that’s intentional.

The Fable 5 access saga finally resolved

For those following the Fable storyline, here’s an update: Fable 5 became available again on July 1st after being temporarily paused for 18 days due to a government request. Initially, access with a subscription was only guaranteed until July 7th. However, Anthropic extended this period twice – first to July 12th, and then to July 19th – because of unexpectedly high demand.

The changes went into effect on July 20th. For Max, Premium Team, and older Enterprise plans, Fable 5 is included as part of your normal usage, up to 50% of your weekly limit. On Pro plans and standard subscriptions, Fable 5 uses pay-as-you-go credits, costing $10 per million tokens ($50 for older accounts), but you’ll receive a one-time $100 credit to help cover the cost. Opus 5, Opus 4.8, Sonnet 5, and Haiku are all available within your regular plan limits, no matter which tier you’re on.

BNC previously reported on the shutdown of Anthropic’s most powerful AI model and the division of its AI development into two products, Fable and Mythos.

The Claude hierarchy, in order

• Claude Fable 5 ($10/$50) is the most capable publicly available Claude and still leads on some long-horizon and software-engineering measures, including SWE-Bench Pro in OpenAI’s own comparison table. It carries the strictest safeguards, retains data for 30 days, and now costs double Opus 5. The case for paying for it has narrowed considerably.
• Claude Opus 5 ($5/$25) is the new center of the lineup and Anthropic’s own recommended starting point for complex agentic coding, long-horizon work, and enterprise tasks.
• Claude Opus 4.8 ($5/$25) is still available and still the automatic fallback target when a safety classifier flags a request on Opus 5 or Fable 5.
• Claude Sonnet 5 is the default on Free and Pro plans, at introductory pricing of $2/$10 through August 31, 2026, then $3/$15. Still the best value at the quality frontier, with one catch covered below.
• Claude Haiku 4.5 ($1/$5) is the cheap, fast tier, and now by some distance the oldest model in the lineup. Fine for classification, extraction, and high-volume simple work.
• Claude Mythos 5 is Fable’s less-safeguarded sibling for offensive-grade cybersecurity work, restricted to vetted organizations in Anthropic’s Project Glasswing program.

Keep in mind that Sonnet 5, while potentially cost-effective overall, uses a new text processing method. This means the same amount of text might be counted as 1.0 to 1.35 times more ‘tokens’ compared to Sonnet 4.6, which could affect your bill. While the listed prices haven’t changed, your actual costs might be higher. You can find more details about the launch of Sonnet 5 in Anthropic’s launch coverage.

AI model comparison chart, source: Anthropic

The four assistants in depth

ChatGPT (OpenAI): the gate is gone

OpenAI’s widely available assistant is now powered by the GPT-5.6 family of models. GPT-5.6 became generally accessible on July 9, 2026, following a limited preview that required government approval – making it the first AI model released with such restrictions. We previously reported on its initial launch in an article titled ‘GPT-5.6 Launches Under Government Watch’.

We offer three models: Sol is our most powerful, Terra provides a good balance of features and price, and Luna is designed for speed and affordability. All three models can process information from a large context window of 1.05 million tokens, generate responses up to 128,000 tokens long, and their knowledge is current up to February 2026. If you use the `gpt-5.6` API, you’ll be accessing the Sol model.

Who gets what: In regular ChatGPT conversations, Plus, Pro, Business, and Enterprise users get Sol at medium effort and above, and Pro and Enterprise additionally get a Sol Pro configuration for the hardest tasks. GPT-5.5 Instant remains the default for fast everyday responses. Free and Go users do not get Sol in plain chat, but they can reach Terra through ChatGPT Work and Codex, which is the most useful and least publicized free path in the market right now.

On July 9th, several updates were released including ChatGPT Work – a new tool designed to compete with Claude Cowork, as well as the consolidation of the Codex desktop app into the updated ChatGPT desktop application. These releases also featured advanced multi-agent capabilities and programmatic tool calling. GPT-5.4 was discontinued on July 23rd, but GPT-5.5 is still accessible.

OpenAI’s strengths include its wide range of capabilities and consistent performance. It boasts leading-edge features like advanced voice technology, improved image generation (ChatGPT Images 2.0), video creation with Sora, the Codex code generation tool, Agent Mode, and a vast library of integrations. Most recently, its performance on complex tasks has significantly improved – OpenAI reports its Sol model leads the Artificial Analysis Coding Agent Index with a score of 80, surpassing Fable 5 while using fewer resources.

Here are some limitations of the system: The automatic routing feature still chooses the model for you, and it’s not immediately clear which model responded to your request with the default settings. A user paying for the most powerful model might sometimes get a less capable one instead, especially with complex requests. OpenAI states that the enhanced safety features in the latest model block significantly more potentially harmful content, but this can also lead to some legitimate requests being incorrectly flagged – a problem similar to what users of a competing service have reported. ChatGPT and Codex offer an option to automatically retry a request using a less powerful model if the first attempt fails.

Claude (Anthropic): the professional’s coding and writing tool

As I’ve detailed previously, the core strengths of these models haven’t shifted. Developers still overwhelmingly prefer Claude for coding and complex, multi-step tasks – essentially, building things. And Sonnet continues to be a solid choice for professional writing and data analysis. The big news this July is simply that we can now offer the best performance at half the price it was just three weeks ago.

Claude excels at coding and consistently completing complex, multi-step projects, even lengthy ones, with careful attention to detail. It’s still the most popular coding tool for command-line use, and its collaborative feature, Claude Cowork, now has a competitor from OpenAI – which generally benefits users through increased innovation and options.

Haiku 4.5 currently lacks the ability to create images or videos, and offers fewer features for everyday users compared to competitors like ChatGPT and Gemini. It also faced regulatory issues, as seen with Fable’s temporary shutdown. The rules for accessing Fable changed frequently – four times in just six weeks – which was frustrating for those relying on it. Furthermore, Anthropic, the company behind Haiku, is planning to go public later this year, meaning you’ll likely hear more about it, though this won’t directly impact how the model performs.

Gemini (Google): the ecosystem play, minus a flagship

Google Assistant is now integrated across many of its services, including Search, Gmail, Docs, Drive, Calendar, Android, and Chrome. On July 21st, Google launched three new versions of its Gemini model: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, a specialized, secure version for government and partner use. Google has also started initial training on Gemini 4.

Gemini 3.6 Flash is particularly noteworthy. Google reports it generates about 17% less text in its responses compared to the previous 3.5 Flash model, has more up-to-date information (its knowledge now extends to March 2026, instead of January 2025), and performs significantly better on coding benchmarks – achieving a score of 49% versus 37%. It also offers lower output costs, priced at $1.50 per million input tokens and $7.50 per million output tokens.

Google’s biggest issue right now is that Gemini 3.5 Pro isn’t available to the public yet. Although announced in May with a promised June release, and then a reported July target, it still hasn’t launched. Recent reports indicate Google has released less powerful versions of Gemini, but there’s no new date for the Pro model, only that it’s still being tested with partners. This leaves Google as the only major AI developer without a current, top-of-the-line model widely available – meaning the most advanced Gemini version most users can access is still the older Gemini 3.1 Pro.

If you’re heavily invested in the Google ecosystem, Gemini is a standout AI assistant. It seamlessly integrates with your existing tools like Gmail and Docs, and appears directly in your Google Search results. Plus, it can process massive amounts of information, create videos with Veo, offers a helpful research feature, and boasts a generous free plan.

There are a few key issues: the product naming is now quite confusing, as a version 3.6 was released before 3.5 Pro. Prices have changed rapidly. The most compelling features are only available in the priciest options. Overall, Google’s AI efforts this year have been characterized by announcing products well before they actually ship.

Grok (SpaceXAI): real-time data at ruthless prices

This AI assistant, now offered by SpaceXAI after a rebranding on July 6th, can access both X (formerly Twitter) and the wider web. Its most powerful version, Grok 4.5, launched on July 8th within the Cursor platform and became publicly available on July 9th. It was specifically designed for coding and automated tasks, trained using real-world developer data alongside Cursor. Elon Musk describes it as a top-tier model – comparable to Opus – that’s faster, uses resources efficiently, and is more affordable.

The model costs $2 for input and $6 for output per million tokens, with a reduced rate of $0.50 for cached input. It can process a context window of 500,000 tokens. Independent evaluations rank it fourth on the Artificial Intelligence Index, highlighting its efficiency – it generates about 14,000 output tokens per task, compared to around 67,000 for similar high-end models. The Grok 4 Heavy system, which is more advanced and available only at the $300/month price point, is the only one that performs better. The Grok 4.3 model is still available and more affordable at $1.25/$2.50 per million tokens, with a 1 million-token context, and a specialized coding model, grok-build, costs $1/$2.

Grok excels at providing up-to-the-minute information thanks to its ability to quickly access and integrate live data from various sources. This makes it particularly strong for tracking breaking news, understanding current market trends, and gauging public opinion. Beyond that, Grok is also now very efficient in how it uses processing power.

Here are some drawbacks: First, using tools like web search, specialized searches, and code execution adds extra costs – around $5 per 1,000 requests for each, and $10 per 1,000 for including files. This means the total cost can be much higher than just the price of the tokens used. Second, it’s difficult to know which version of the model is responding, especially with staggered releases. As of late July, only the most expensive plan at $300 guaranteed full access to Grok 4.5; lower-priced plans were receiving it gradually. Finally, Grok 4.5 wasn’t initially available in the European Union and was expected later in the month.

What does each AI cost in 2026?

Consumer subscription pricing

Prices are in US dollars and current as of July 27, 2026. Please confirm details before making a purchase, as some prices have recently been updated.

Here’s a breakdown of which AI subscription tier offers the best value: For most users, the $20 per month options are ideal. Gemini Pro provides the most storage and context for your money. Claude Pro excels at producing work-related content and has improved significantly with its latest model, Opus 5. ChatGPT Plus is a well-rounded choice and now includes access to Sol. If you’re looking for free options, Gemini and AI Plus are top picks, though ChatGPT’s Terra (through Work and Codex) is also competitive. Spending over $20 per month only makes sense if you consistently reach usage limits or require specific features. A new benefit of Claude Max is that it’s the only plan where you get full access to Fable 5 without usage restrictions.

Developer API pricing

USD per million tokens, input/output, standard rates, July 27, 2026.

When comparing costs, GPT-5.6 Luna and Grok 4.3 currently have the lowest token prices. However, the cheapest price per token doesn’t always translate to the cheapest cost for completing a task – and that’s where many people miscalculate their budgets. Grok adds extra fees when used for complex tasks, Gemini’s Pro pricing increases significantly after 200,000 input tokens, and Sol’s rates go up after 272,000 tokens. Sonnet 5 uses a tokenizer that creates more billable tokens from the same amount of text. Interestingly, even though Grok 4.5 might have a slightly higher price per token, its efficiency can sometimes make it the most cost-effective option overall. It’s important to test these models with your specific tasks, as even using the same model, different teams can end up with very different costs based on how much effort they require.

Which AI model should you use?

Next, remember to adjust your effort level to the task’s difficulty. For routine or simple tasks, use normal or medium settings. Reserve the most powerful settings – high, extra high, and maximum – for tasks where mistakes could be costly.

What people get wrong about choosing an AI model

The idea that you always get what you pay for isn’t true with AI models right now, especially at the high end. Recent tests show that Anthropic’s less expensive Opus 5 often performs better than its more costly Fable 5, and Google’s budget-friendly Flash model even surpassed their previous top-of-the-line option in coding tasks. So, it’s best to only spend extra money when your specific project actually requires a more powerful model.

The idea that free AI models are just basic ‘toys’ isn’t true. Claude’s free version uses a powerful model called Sonnet 5, Gemini’s free tier runs the same technology as its paid options, and even free ChatGPT users have access to advanced capabilities similar to GPT-5.6 Terra. By 2026, the main difference between free and paid AI will likely be how much you can *use* them, rather than a difference in their core intelligence.

Just because a token is cheaper doesn’t automatically mean your overall costs will be lower. Things like platform fees, extra charges for very long inputs, how efficiently the text is processed, and the complexity of your request all impact the final price. Ultimately, the total cost to *complete* what you need is the only thing that really matters.

It’s frustrating that you often don’t know which version of the AI model you’re actually using. This problem hasn’t improved – it’s gotten worse! ChatGPT automatically selects different versions without clearly showing you, unless you manually check the settings. With Grok, even users on the same subscription plan might experience different models. Claude sends certain requests to its most powerful version (Opus 4.8), and OpenAI will sometimes switch you to a less capable model if its safety features are triggered.

A note on the governance backdrop

When making purchasing decisions for 2026 and beyond, it’s important to remember that new AI models can be temporarily disabled or restricted by government policy. We’ve already seen this happen recently with two examples: Fable 5 was shut down for 18 days due to a government order, and the launch of GPT-5.6 was delayed for thirteen days while access was limited. While both issues were resolved, they could easily happen again. These interruptions also benefited competitors – specifically, Chinese AI models briefly performed better in testing when Fable 5 wasn’t available.

This doesn’t mean any particular AI assistant is poor quality. However, if your work heavily relies on AI access, be prepared for occasional outages and consider having a backup assistant ready to go. This guide focuses solely on the fact that these changes occurred, not whether they were good decisions.

How to choose: the simple rule

1. Match the assistant to your ecosystem and your dominant task. Google-centric life, Gemini. Serious coding or writing, Claude. A bit of everything plus voice and media, ChatGPT. Real-time information, Grok.
2. Match the tier and the effort to the difficulty. Default effort and the mid tier for daily work. Flagships and max effort for the few tasks that genuinely deserve them.

You can start with the free plan, and if you reach the usage limits, upgrading to the $20 plan is a good next step. Always double-check the pricing details before considering any more expensive options. Keep in mind that things change quickly – with new models launching rapidly and delays happening, this information may need to be updated soon.

FAQ

Is Claude Opus 5 better than Fable 5?

Generally, yes. Anthropic’s Opus 5 performed better than Fable 5 on eight out of thirteen tests, including key evaluations like Frontier-Bench and ARC-AGI 3, and it costs half as much. While Fable 5 still excels at certain complex tasks – particularly those involving long-term reasoning and software development (like SWE-Bench Pro) – it remains Anthropic’s most powerful model available to the public. For most users who are paying for access, Opus 5 is now the smarter choice.

Is GPT-5.6 available to normal ChatGPT users?

Sol has been available since July 9, 2026. You can access it within regular ChatGPT if you have a Plus, Pro, Business, or Enterprise subscription and are using a medium or higher effort level. A more advanced version, Sol Pro, is specifically for Pro and Enterprise users. If you’re on the free or Go plan, you won’t be able to select Sol directly in chat, but you can access Terra – a slightly less powerful option – through ChatGPT Work and Codex.

What’s the difference between Fable, Opus, Sonnet, and Haiku?

Claude comes in different versions. Fable 5 is currently the most powerful version available to the public, costing $10 or $50 per million tokens. Opus 5 is now the main recommended version, priced at $5 or $25 per million tokens, and the older Opus 4.8 is still available at the same price as a backup. Sonnet 5 is a reliable and powerful option, and is the standard for Free and Pro users, costing $2 or $10 per million tokens until August 31st, after which it will be $3 or $15. Haiku 4.5 is the fastest and most affordable option, at $1 or $5 per million tokens. Mythos 5 is a specialized version of Fable designed for cybersecurity and is only accessible to approved organizations.

Where is Gemini 3.5 Pro?

As of July 27, 2026, this product still hasn’t been released. Google originally announced it would be available in June at their I/O event and a rumored launch date of July 17th came and went. Instead of the anticipated release, on July 21st Google launched three more affordable Flash models without providing an updated timeline for the original product. It’s currently being tested with a limited number of partners. For now, Gemini 3.1 Pro remains Google’s main offering.

What AI model should I use in 2026?

Most people will get the best value by starting with the free or lowest-cost version of an AI assistant that works well with the tools they already use. Then, choose a mid-range option within that family of products instead of the most expensive one. Finally, use the right level of effort for each task. For example, Google users might prefer Gemini, coders and writers could try Claude, ChatGPT is good for everyday tasks, and Grok is best for up-to-the-minute information.

Which AI is best for free?

Gemini just edges out the competition, offering its free users the same powerful Flash model as its paid plans, complete with image and voice capabilities, plus a monthly allowance for in-depth research. Claude’s free version uses Sonnet 5 and is currently the best choice for high-quality writing and coding. Surprisingly, the free version of ChatGPT now utilizes GPT-5.6 Terra through its Work and Codex features, making it a remarkably capable, yet often overlooked, free option.

Which AI is cheapest for developers?

Currently, GPT-5.6 Luna costs $1 to $6, and Grok 4.3 is priced at $1.25 to $2.50. Gemini 3.6 Flash ($1.50/$7.50) and Grok 4.5 ($2/$6) offer good value, while Claude Sonnet 5 has a special introductory price of $2 to $10 through August 31st. Remember to calculate the *total* cost of each tool for your specific task – things like fees for using the tool, costs for long inputs, how the tool breaks down text, and the amount of processing power needed often add up to more than just the listed price.

Which AI is best for coding?

Claude remains a strong choice, but our top recommendation has shifted this month. Opus 5 is now preferred for demanding tasks – it’s half the price of Fable 5 and performs better on most of Anthropic’s coding tests, especially with Claude Code as the leading coding assistant. GPT-5.6 Sol is a strong competitor, ranking highest on OpenAI’s coding benchmark. For the best value, Grok 4.5, trained alongside Cursor, is very efficient and costs just $2/$6.

This guide is updated as the model landscape changes. Figures verified as of July 27, 2026.

2026-07-27 05:32