What AI Model Should I Use? The 2026 Guide to ChatGPT, Claude, Gemini, and Grok

Here is the state of play as of July 27, 2026, across all four major assistants: every sub-model, real pricing, and who should actually use what.

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

Every lab now sells a range rather than a single model, and in July the mid-tier openly overtook the top tier on value at three of the four.

Anthropic released Opus 5 and said plainly that it comes close to Fable 5’s intelligence at half the price, while its own charts show Opus 5 beating Fable on eight of thirteen benchmarks. OpenAI’s Sol sits above Terra and Luna, but OpenAI’s own framing is that Terra delivers GPT-5.5-class quality at roughly half the cost. Google shipped a cheaper, more efficient Flash model while its flagship stayed in testing. The old game was “biggest model wins.” The new game is cost per completed task, and the labs are now the ones saying so.

Shift two: the effort dial

In 2025, more intelligence meant switching models. In 2026, it increasingly means turning a dial on the same model.

“Effort” is how long and hard the model is allowed to think before answering. Low effort is fast and cheap. High effort spends more computation reasoning through the problem, checking its work, and exploring alternatives, which produces better answers on hard tasks and costs more in tokens, time, or usage allowance. It is roughly the difference between asking a colleague for a quick take and asking them to go away and work the problem properly.

Claude now exposes a full ladder on Opus 5: low, medium, high, xhigh, and max. OpenAI added a max reasoning setting and an “ultra” mode that farms work out to parallel subagents. Gemini has thinking levels and Deep Think. Grok ships with reasoning on by default and a configurable effort parameter in the API.

The practical consequence is that “which model should I use?” has become “which model, at what effort?” Anthropic says Opus 5 at max effort lands within 0.5 percent of Fable 5’s peak CursorBench score at half the cost per task. OpenAI says Sol at max reasoning comes within one point of Fable 5 on the Artificial Analysis Intelligence Index while finishing tasks in 61 percent less time. Those are vendor-reported numbers, but the shape of the claim is consistent across three labs: the dial is now a genuine cost lever.

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.

Two naming notes. Anthropic’s restricted Mythos 5 sits above Fable 5 but is not generally available. And xAI is now formally SpaceXAI, following its rebrand on July 6 after SpaceX absorbed the company in February. The Grok product name has not changed.

The Claude spotlight: Opus 5 changes the math

What Opus 5 actually is

Anthropic released Claude Opus 5 on July 24, its fourth model in under two months after Mythos 5, Fable 5, and Sonnet 5. It is now the default model on Claude Max and the strongest model available on Claude Pro, and it runs on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry as claude-opus-5.

The headline is price positioning rather than raw capability. Opus 5 costs $5 per million input tokens and $25 per million output, unchanged from Opus 4.8, which makes it exactly half the price of Fable 5. Anthropic’s own comparison across thirteen benchmarks has Opus 5 scoring higher on eight of them. On Frontier-Bench v0.1, an internal run spanning tasks in physics, chemistry, and cryptography, Anthropic reports 43.3 percent for Opus 5 against 33.7 percent for Fable 5, and more than double Opus 4.8’s score at a lower cost per task. On ARC-AGI 3, a novel problem-solving evaluation, it claims roughly three times the next-best model’s score.

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 also calls Opus 5 its most aligned model to date, with the lowest rate of deceptive behavior in its automated safety audit, and its most capable generally available model for scientific research. It remains behind the restricted Mythos 5 on offensive cybersecurity work, which is by design.

The Fable 5 access saga finally resolved

If you followed the Fable story, the ending is worth knowing. Fable 5 came back online on July 1 after an 18-day, government-ordered suspension, with subscription-included access promised only through July 7. Anthropic then extended that window twice, to July 12 and then to July 19, citing demand it described as hard to predict.

The permanent split took effect July 20. On Max plans and premium Team and legacy Enterprise seats, Fable 5 is a standard inclusion at up to 50 percent of weekly usage limits. On Pro plans and standard seats, it runs on pay-as-you-go usage credits at API rates of $10/$50 per million tokens, softened by a one-time $100 credit. Opus 5, Opus 4.8, Sonnet 5, and Haiku stay inside normal plan limits on every tier.

BNC covered the original shutdown in Washington Pulls the Plug on Anthropic’s Most Powerful AI and the two-product Fable/Mythos structure in Anthropic Splits the AI Frontier in Two.

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.

One cost caveat that has not gone away: Sonnet 5 uses an updated tokenizer, and the same text can map to roughly 1.0 to 1.35 times more tokens than under Sonnet 4.6. The rate card looks flat. Your bill may not be. BNC’s launch coverage has the detail: Anthropic Launches Claude Sonnet 5.

AI model comparison chart, source: Anthropic

The four assistants in depth

ChatGPT (OpenAI): the gate is gone

What it is: OpenAI’s consumer assistant, now running the GPT-5.6 family. GPT-5.6 reached general availability on July 9, 2026, thirteen days after the government-gated preview that made it, briefly, the first frontier model released behind a federal access list. BNC covered that launch in GPT-5.6 Launches Under Government Watch.

The three models: Sol is the flagship, Terra the balanced tier, Luna the fast and cheap one. All three carry a 1.05 million-token context window, 128,000 max output tokens, and a February 2026 knowledge cutoff. The gpt-5.6 API alias routes to Sol.

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.

What else shipped July 9: ChatGPT Work, an agentic work surface aimed squarely at Claude Cowork, plus the folding of the Codex desktop app into a new ChatGPT desktop application, a max effort level, an ultra multi-agent setting, and Programmatic Tool Calling. GPT-5.4 retired on July 23. GPT-5.5 remains available.

Signature strengths: breadth, still. Best-in-class voice, ChatGPT Images 2.0, Sora video, Codex, Agent Mode, a large connector directory, and now the strongest work-surface story of the four. OpenAI claims Sol tops the Artificial Analysis Coding Agent Index at 80, ahead of Fable 5, using less than half the output tokens.

Weak spots: the Auto router still decides for you, and on default settings you cannot easily see which variant answered. A Plus subscriber paying for the headline model may get Sol on a hard prompt and something cheaper on a routine one. OpenAI also says Sol’s cyber safeguards block roughly ten times more potentially harmful activity than previous models, which means the same false-positive friction Claude users have complained about. There is a retry-on-a-lower-model option in ChatGPT and Codex.

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

What it is: covered in depth above. The character of the family has not changed: Claude models remain the developer community’s favorite for real-world software engineering and long agentic tasks, and Sonnet-class models remain the workhorse pick for professional writing and analysis. What changed in July is that the best version of that proposition now costs half what it did three weeks ago.

Signature strengths: coding and agentic reliability, meaning finishing multi-step tasks rather than stalling partway. Long-document work. Careful writing. Claude Code remains the most widely adopted command-line coding agent, and Claude Cowork now has a direct OpenAI competitor, which is usually good for users.

Weak spots: no image or video generation to speak of, a smaller consumer feature surface than ChatGPT or Gemini, and the regulatory exposure that Fable’s suspension demonstrated. The Fable access rules also changed four times in six weeks, which is a reasonable thing to be annoyed about if you built a workflow on it. Haiku 4.5 is now conspicuously overdue for a refresh. Anthropic is also reportedly preparing an IPO later this year, which will not change your model choice but will change how much noise surrounds it.

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

What it is: Google’s assistant, woven into Search, Gmail, Docs, Drive, Calendar, Android, and Chrome. On July 21 Google released three new models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, a security-tuned variant limited to governments and trusted partners. It also said it has begun pre-training Gemini 4.

Gemini 3.6 Flash is the interesting one. Google says it uses about 17 percent fewer output tokens than 3.5 Flash, moves its knowledge cutoff forward from January 2025 to March 2026, and scores 49 percent against 37 percent on the DeepSWE coding test. It is priced at $1.50 input and $7.50 output per million tokens, cheaper on output than the model it replaces.

The flagship problem: Gemini 3.5 Pro still does not exist publicly. Sundar Pichai promised it for June at I/O on May 19. A widely reported July 17 target passed without a launch. On July 21, Reuters reported that Google shipped the three cheaper models while offering no new timeline for Pro, saying only that it is testing with partners. That leaves Google as the only one of the four labs without a current-generation flagship in general production, and it means the flagship you can actually use is still Gemini 3.1 Pro.

Signature strengths: if your life runs on Google, nothing else comes close. Gemini reads your Gmail, drafts in your Docs, and lives in your Search results. Add the largest context windows on the market, Veo video, NotebookLM, a strong Deep Research mode, and the best free tier in the business.

Weak spots: the naming is now genuinely confusing, with a 3.6 Flash shipping before a 3.5 Pro. Prices have moved fast. The best reasoning is locked behind the most expensive tiers. And the pattern of announcing ahead of shipping has become the defining story of Google’s AI year.

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

What it is: the assistant with live access to X and the web, now from a company called SpaceXAI following the July 6 rebrand. The flagship is Grok 4.5, released July 8 inside Cursor and opened publicly on July 9. It was trained jointly with Cursor on real developer session data and aimed explicitly at coding and agentic work. Musk’s pitch was that it is an Opus-class model that is faster, more token-efficient, and cheaper.

The pricing supports the claim: $2 input and $6 output per million tokens, with cached input at $0.50, on a 500,000-token context window. Independent trackers place it fourth on the Artificial Analysis Intelligence Index and note extreme token efficiency, roughly 14,000 output tokens per index task against about 67,000 for Opus-class rivals. Above it sits Grok 4 Heavy, a multi-agent system exclusive to the $300-per-month tier. Below it, Grok 4.3 remains available and cheaper at $1.25/$2.50 with a 1 million-token context, and a dedicated coding model, grok-build, runs at $1/$2.

Signature strengths: real-time. DeepSearch pulls live X and web data into cited reports, which makes Grok the strongest of the four for breaking news, market sentiment, and anything where “what are people saying right now” is the actual question. Token efficiency is now a second genuine advantage.

Weak spots: three. Tools are billed separately, at roughly $5 per 1,000 calls each for web search, X search, and code execution, plus $10 per 1,000 for file attachments, so agentic workloads cost meaningfully more than the token price implies. Staged rollouts make it hard to know which model answered you: as of late July, SuperGrok Heavy at $300 is the only consumer tier with confirmed full Grok 4.5 access, with cheaper tiers receiving it in stages. And Grok 4.5 was not available in the European Union at launch, with xAI’s own documentation pointing to later in the month.

What does each AI cost in 2026?

Consumer subscription pricing

US dollars, monthly, as of July 27, 2026. Verify before buying. Several of these changed this month.

Value verdicts: The $20 tier is still where most people should land. Gemini AI Pro is the best pure value on storage and context. Claude Pro is the best for work output, and materially better than it was in June now that Opus 5 is the strongest model on the plan. ChatGPT Plus is the best generalist and now includes Sol. Below $20, Gemini’s free tier and AI Plus still win, though ChatGPT’s free Terra path via Work and Codex is a real competitor. Above $20, the honest test has not changed: the $100-plus tiers only pay for themselves if you hit usage limits weekly or need a gated capability. The one new argument for Claude Max is that it is the only plan where Fable 5 is included rather than metered.

Developer API pricing

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

Which is cheapest? On raw token price, GPT-5.6 Luna and Grok 4.3. But cheapest per token is not cheapest per task, and this is where most budgets go wrong. Grok’s tool fees add real cost to agentic workloads. Gemini’s Pro rates double past 200K input. Sol’s rates jump above 272K input. Sonnet 5’s denser tokenizer turns the same text into more billable tokens. And Grok 4.5’s token efficiency means a nominally pricier model can finish a task for less. Benchmark your own workload. Two teams on the same model can see wildly different bills depending only on effort settings.

Which AI model should you use?

Then apply the second half of the rule: match effort to difficulty. Run everyday tasks at default or medium. Save high, xhigh, and max, and the flagship models, for the tasks where a wrong answer is expensive.

What people get wrong about choosing an AI model

“The most expensive model is the best choice.” This month it is demonstrably false at the top of the market. Anthropic’s own benchmarks put the $5/$25 Opus 5 ahead of the $10/$50 Fable 5 on most coding and knowledge-work tests. Google’s cheap Flash tier outran its own previous flagship on coding. Pay up only when the task demands it.

“The free tiers are toys.” No. Claude’s free tier serves Sonnet 5, Gemini’s free tier runs the same Flash model behind Google’s paid defaults, and ChatGPT’s free users can reach GPT-5.6 Terra through Work and Codex. The 2026 free-versus-paid gap is mostly usage limits and extras, not raw intelligence.

“A cheaper price per token means a cheaper bill.” Not necessarily. Tool fees, long-context surcharges above 200K or 272K input, tokenizer density, and effort settings all move the real number. Cost per completed task is the only figure that matters.

“You’re always getting the model on the label.” Still not guaranteed, and this got worse rather than better. ChatGPT’s Auto router picks variants without telling you unless you dig into settings. Grok’s staged rollouts mean two subscribers on the same plan can get different model versions. Claude routes classifier-flagged requests to Opus 4.8, and OpenAI offers a retry on a lower-capability model when Sol’s safeguards fire.

A note on the governance backdrop

One factor belongs in any 2026 buying decision: frontier models can be switched off, or gated, by policy. Fable 5’s 18-day government-ordered suspension and GPT-5.6’s thirteen-day launch behind a federal access list both happened in the space of a month. Both resolved. Both could recur. The interruption also handed measurable momentum to rivals, including Chinese open-weight models that briefly inherited benchmark top spots while Fable was dark.

None of this makes any single assistant a bad choice. It does mean that if AI access is operationally critical to you, treat sudden model unavailability as a real risk and keep a second assistant warm. This guide takes no view on whether the interventions were wise, only that they happened.

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.

Start free, upgrade to the $20 tier when you hit limits, and re-read the pricing tables before buying anything above that. And date-check everything. In a month where one lab shipped its fourth model in eight weeks and another still has not shipped the flagship it promised in May, the only safe assumption is that this page will need updating again.

FAQ

Is Claude Opus 5 better than Fable 5?

On most measures Anthropic publishes, yes. Opus 5 scored higher than Fable 5 on eight of thirteen benchmarks in Anthropic’s own comparison, including Frontier-Bench and ARC-AGI 3, at half the price. Fable 5 still leads on some long-horizon and software-engineering measures, including SWE-Bench Pro in OpenAI’s comparison table, and remains Anthropic’s most capable publicly available model overall. For most paying users, Opus 5 is now the rational default.

Is GPT-5.6 available to normal ChatGPT users?

Yes, since July 9, 2026. Sol is available in regular ChatGPT chat on Plus, Pro, Business, and Enterprise plans at medium effort and above, with a Sol Pro configuration on Pro and Enterprise. Free and Go users cannot select Sol in plain chat but can reach the mid-tier Terra through ChatGPT Work and Codex.

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

They are tiers of the Claude family. Fable 5 is the top public model at $10/$50 per million tokens. Opus 5 is the new everyday flagship at $5/$25, with Opus 4.8 still available at the same price as a fallback. Sonnet 5 is the near-flagship workhorse and the default on Free and Pro plans at $2/$10 through August 31, then $3/$15. Haiku 4.5 is the cheap, fast tier at $1/$5. Mythos 5 is a restricted cyber-focused sibling of Fable, available only to vetted organizations.

Where is Gemini 3.5 Pro?

Still unreleased as of July 27, 2026. Google promised it for June at I/O in May, a widely reported July 17 target passed without a launch, and on July 21 Google shipped three cheaper Flash models instead while declining to give a new timeline. It is reportedly in limited testing with partners. Gemini 3.1 Pro remains the flagship you can actually use.

What AI model should I use in 2026?

For most people: the free or $20 tier of the assistant that matches their ecosystem, then the mid tier of that family rather than the flagship, then the effort level matched to each task. Gemini for Google users, Claude for coders and writers, ChatGPT for generalists, Grok for real-time information.

Which AI is best for free?

Gemini, narrowly. Its free tier runs the same Flash model behind Google’s paid defaults, with image generation, voice, and a monthly Deep Research allowance. Claude free serves Sonnet 5 and is the strongest free option for writing and coding quality. ChatGPT free now reaches GPT-5.6 Terra through Work and Codex, which is the most underrated free path available.

Which AI is cheapest for developers?

On raw token prices, GPT-5.6 Luna at $1/$6 and Grok 4.3 at $1.25/$2.50. Gemini 3.6 Flash at $1.50/$7.50 and Grok 4.5 at $2/$6 are the value picks a tier up, and Claude Sonnet 5’s introductory $2/$10 runs through August 31. Always model total cost per task: tool fees, long-context surcharges, tokenizer differences, and effort settings routinely outweigh the rate card.

Which AI is best for coding?

Claude, still, but the specific recommendation changed this month. Opus 5 is now the pick for serious work, at half Fable 5’s price and ahead of it on most of Anthropic’s coding benchmarks, with Claude Code the most widely adopted coding agent. GPT-5.6 Sol is a genuine challenger and tops the Artificial Analysis Coding Agent Index on OpenAI’s numbers. Grok 4.5 is the value play, jointly trained with Cursor and dramatically token-efficient at $2/$6.

---

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