Log inGet started
← Level 1
Level 1· Lektion 11 von 13

Model landscape 2026: which AI model for what

An honest overview of Claude, GPT and Gemini in 2026. Which model class is enough for which task, no hype, with concrete prices.

Most people reach for the most expensive model for everything. That is like pulling the sports car out of the garage for every trip to the bakery. It works, but it costs ten times as much and does nothing for you while picking up bread.

In 2026 every big provider has at least three or four model classes. The expensive one for hard problems, a middle one for everyday work, a cheap one for bulk. Once you understand which is built for what, you save money and often get better results too, because a fast model on a simple task overengineers less.

This lesson sorts the three big families. No benchmarks, no tables with 40 columns. Just the question: when is what enough.

Claude: Opus, Sonnet, Haiku

Anthropic tiers by size. The names Opus, Sonnet and Haiku have been around since Claude 3 in March 2024, they describe the size class, not the generation.

The current top is Opus 5 since 24 July 2026 (5 dollars per million input tokens, 25 dollars output), the high-end workhorse for problems where you really need deep reasoning, long chains of inference, difficult code architecture. Two things about it are new: it has a one million token context window as standard, and it thinks by default without you switching that on. Its predecessor Opus 4.8 stays available and costs the same. Below that, Sonnet 5 since 30 June 2026, the middle class that is enough for 90 percent of real work, at an introductory price of 2 and 10 dollars until 31 August 2026, then 3 and 15. And at the bottom Haiku 4.5 (1 and 5 dollars), fast and cheap for simple classification, summaries, short answers.

An instructive footnote, because it shows how fast this turns. Anthropic launched Claude Fable 5 on 9 June 2026 (10 dollars input, 50 dollars output), a class above Opus. Three days later, on 12 June, it was gone, worldwide for everyone, because a US export control following a reported jailbreak barred access for foreign nationals and Anthropic pulled the plug for everyone rather than check nationalities in real time. On 30 June the US authority lifted the control again, and since 1 July Fable is back globally, including in the EU. One detail still matters for European users though: Fable and its sibling model Mythos count as Covered Models with a mandatory retention of at least 30 days, there is no zero-data-retention option, not even in EU regions. Where that matters, Opus 4.8 is the calmer choice. The real lesson is not "take the strongest model", it is "build so that losing a single model does not knock you over". More on that in a later level.

My practical advice: start with Sonnet. If the answer is noticeably too shallow, go up a step. If you are sorting a thousand emails a day, go down to Haiku. You do not have to guess, you can just try it, the difference is visible the moment you read the output.

GPT: Instant, Thinking, Pro and the Codex branch

In 2026 OpenAI splits by thinking effort rather than by size. The default for normal chat has been GPT-5.5 Instant since May 2026, including in the free version (with a usage quota there). Fast, answers instantly, good for normal conversations and most everyday questions.

If you want it to "think longer", there is the Thinking tier, GPT-5.5 Thinking in the paid tiers. Since 9 July 2026 the GPT-5.6 generation sits above that, and it is split into three roles rather than sizes: Sol for the hardest work, explicitly including coding and agentic tasks, meaning models that independently plan several steps and use tools. Terra for balanced everyday work. Luna for fast and cost-sensitive jobs. On paid plans you reach Sol through the reasoning settings.

Interesting for anyone who codes: the Codex branch. If you search for "chatgpt codex" you land here, and if you actually want to set it up, From ChatGPT to Codex walks you there in ten steps. The 5.6 generation has arrived there too, Terra and Luna are available in Codex. The principle stays what it was: if you have code generated all day long, the smaller, faster variant is often the smarter choice, because you do not hit the limit after three hours. On 30 July 2026 OpenAI turned the price dial again on exactly that, Luna got 80 percent cheaper, Terra 20.

A side note, because many people trip over it: older models get retired. GPT-4.5 went on 27 June 2026, o3 disappears on 26 August 2026. If your workflow is still pinned to an old model name, check it in time.

Gemini: Pro, Flash, Flash-Lite

Google tiers similarly to Anthropic, just with different names. The pattern is easy to remember: Pro for heavy reasoning and coding, Flash for fast everyday work, Flash-Lite for high volume at a small price.

The newest model is Gemini 3.6 Flash since 21 July 2026, Google's declared workhorse: one million tokens of context, better at coding, and up to 17 percent leaner on token use than its predecessor 3.5 Flash, at 1.50 dollars input and 7.50 dollars output. The same day brought 3.5 Flash-Lite for cheap bulk operation and 3.5 Flash Cyber, the latter a specialist model for finding security holes that only governments and selected partners get.

One quirk worth knowing: there is no Pro in Gemini 3.5. Google announced it and then delayed it, so the Pro tier still sits with the generation before. With Google it pays to check which tier actually exists in the current generation, rather than inferring availability from the pattern. Next to the Flash line there is also Gemini Omni, a multimodal model that turns arbitrary inputs into arbitrary outputs, so text, image, audio in combination.

If you already sit deep in the Google universe, with Drive, Docs, Gmail, Gemini is often the obvious choice, simply because the integration is already there.

The honest truth

The differences between the three families are smaller than the marketing wants you to believe. On an average task you will barely notice a difference between Sonnet, GPT-5.5 and Gemini Flash. The big lever is not which brand, but which tier within a brand.

Where it really diverges is on the heavy stuff. Deep code architecture, long independent work across many steps, problems that need real multi-step thinking. That is where you pay for the top models, and that is where it pays off too.

A second point that often gets missed: the model alone does not make a good assistant. What makes the difference is context. A Sonnet that knows your project beats an Opus that guesses blindly. This is exactly where something like a codebase-memory-mcp comes in, a tool that gives the model persistent memory about your code. But that is a story for a later level.

For now this is enough: do not reflexively grab the most expensive one. Start in the middle, go up when needed, down when you can.

Quiz

1. You want to automatically sort a thousand customer emails per day into categories. Which model class fits best?

  • A) Opus 5
  • B) Haiku 4.5 or Gemini Flash-Lite
  • C) GPT-5.6 Sol

The answer is B. Simple classification at high volume is exactly the case for the cheap, fast tier. A flagship model would be pure waste of money.

2. What do the names Opus, Sonnet and Haiku describe?

  • A) The generation of the model
  • B) The release year
  • C) The size class within a generation

The answer is C. The tier names have existed since Claude 3 and stand for big, middle, small, not for the generation.

3. You have code generated agentically all day and do not want to hit the usage limit. What is often the smarter choice?

  • A) GPT-5.6 Sol
  • B) GPT-5.6 Luna
  • C) GPT-4.5

The answer is B. Luna is the fast, cost-sensitive tier of the 5.6 generation and it is available in Codex, so you do not hit the limit after three hours. C is out anyway, GPT-4.5 was retired on 27 June 2026.

Source

The specific model names, versions and prices in this lesson are checked against the official provider documentation (as of 31 July 2026):

On the July releases in detail:

On the Fable 5 footnote (launched 9 June 2026, suspended worldwide on 12 June after a US export control, lifted on 30 June, back globally including the EU on 1 July 2026):

Model names and prices change fast. If you read this lesson months later, check the versions via the links above.

You're reading without an account. Login saves your progress so you can pick up where you left off. Log in →
← Git for AI 3, the GitHub MCP sData protection for AI beginne→