The AI model war, round two: Gemini 3.1 Pro, GPT-5.6, Claude Fable 5, Opus 4.8, and Grok 4.5

The last comparison already feels old

My previous article compared Gemini 3 Pro, GPT-5.1, and Claude 4.5 Opus. At the time, those names felt like products from the same generation, so putting them side by side made sense.

That title already has a little history on it. Gemini has moved to 3.1 Pro. OpenAI released GPT-5.6 on July 9. Anthropic announced Claude Fable 5, and Grok 4.5 arrived in July. GPT-5.6 now comes in Sol, Terra, and Luna tiers, while Claude has both Fable 5 and Opus 4.8. These are more than decimal-point updates.

Model names now have short lives. You can spend time getting used to one version and find the next one waiting in the product interface.

Gemini is starting to look like a complete Google workspace

What stood out to me about Gemini 3 Pro was its visual understanding and multimodal work. With Gemini 3.1 Pro, Google's direction is clearer. It is not confined to a chat window; it appears in the Gemini API, Vertex AI, the Gemini app, and NotebookLM.

That difference is practical. When I need to inspect a UI screenshot, organize a specification, analyze a document, or hold a large set of material in one context, the model is only half the story. The other half is whether it connects naturally to the tools I already use.

Google has an advantage here that is hard to ignore. Gemini does not always give me the most distinctive answer, but it often becomes the convenient choice when a task involves documents, images, NotebookLM, or other Google services.

GPT is becoming the middle layer of a workflow

During the GPT-5.1 period, OpenAI emphasized adaptive reasoning, a no-reasoning mode, prompt caching, and developer tools such as apply_patch and shell. GPT-5.6 moves further toward long-running work, tool use, documents, spreadsheets, research, and computer control.

The GPT-5.6 naming scheme is different too. Sol is the flagship, Terra balances capability and cost, and Luna favors speed and price. Add the reasoning-effort settings and users are choosing more than a model name. They are choosing how much time and money a task should consume.

That makes GPT feel less like a model that answers questions and more like a layer that decides when to think longer and when to act. It can use tools, inspect the current state, and choose the next step.

Claude's update changed more than the version number

Claude Opus 4.5 already focused heavily on coding, agents, and computer use. Opus 4.6 improved long-context work, code review, and large codebases. Versions 4.7 and 4.8 continued to improve long-running tasks, browser agents, and tool calls.

The larger change came in June, when Anthropic released Claude Fable 5. Anthropic describes Fable 5 as its most capable broadly available model, aimed at difficult reasoning and long-running agentic work. Safety classifiers made it possible to release that level of capability to general users.

There is an easy-to-miss detail. For some high-risk security or biology requests, Fable 5 may refuse or hand the request to Opus 4.8. Choosing Fable 5 does not guarantee that every response comes from Fable 5. That matters when you test behavior or calculate cost.

I have never thought of Claude as the model that is always fastest. I think of it as the model most willing to stay with a complicated problem. Now I would add this: Fable 5 may be Anthropic's first choice for the hardest work, while Opus 4.8 remains an important fallback when safety restrictions apply.

Fable 5 was briefly withdrawn in mid-June and returned on July 1. Its release story carries a practical reminder: availability depends on policy and deployment choices as well as technical capability.

Grok 4.5 is not here to fill out the list

xAI released Grok 4.5 on July 8, 2026, positioning it for coding, agentic tasks, and knowledge work. The company also emphasized its integrations with Cursor and Grok Build. The pitch is direct: do not stop at chat; help build the thing.

xAI reports 80 tokens per second and API pricing of $2 per million input tokens and $6 per million output tokens. It also describes Grok 4.5 as an engineering model that uses fewer tokens. Those numbers are attractive, but they do not prove that Grok has won.

Each company uses different benchmarks, tools, prompts, and test procedures. Official numbers tell us where a vendor wants to position a model. They are not a fair global ranking.

The real contest has left the chat window

The clearest change in this round is that models rarely appear alone. Gemini is moving into Google's document and research tools. GPT is moving into Codex, browsers, and professional work. Claude uses Fable 5 for longer and harder tasks, with Opus 4.8 handling restricted requests. Grok enters development through Grok Build and Cursor.

We used to ask which model gives the best answer. That question is now too small. The practical questions are which model leaves me the least cleanup, which one understands the current context, and which one is easiest to correct after a failure.

Capability, tool connections, context length, speed, and price are now difficult to separate. Switching models often means switching an entire way of working.

How I would choose today

For documents, images, and data inside Google's ecosystem, I would think of Gemini first. Its advantage is not that every answer is best. It is that the model connects easily to the material already in front of me.

For breaking down requirements, using tools, general coding, and agent workflows, I would look at GPT. OpenAI is clearly putting the model inside an environment where it can act.

For a large codebase, long debugging sessions, or work that benefits from careful self-review, I would choose between Claude Fable 5 and Opus 4.8. Fable 5 suits longer, harder, asynchronous work. Opus 4.8 is the steadier general option and may take over requests that Fable 5 will not handle.

For a fast prototype in Cursor or Grok Build, Grok 4.5 deserves a place on the shortlist. Its price and speed look competitive. The remaining questions need to be answered in real projects.

The next round will not be decided by one score

I expect the next differences to show up in boring details that directly affect work: whether a model can continue for hours, whether it loops after an error, whether it avoids breaking existing features, and whether it can explain what went wrong.

Those details may never fit neatly into a benchmark table, but they determine whether an independent developer trusts a model with important work.

I no longer want to pick a model that is supposed to be the permanent winner. It is more practical to give different models different jobs and leave the final decision to a person. That is less satisfying than naming a champion, but it is much closer to how I work every day.