New AI models in summer 2026: a short roundup

GPT-5.6, Claude Opus 5, Gemini 3.6 Flash and Grok 4.5 landed within a few weeks of each other. The names and benchmark tables are crowded again. This time I paid less attention to "which one is smartest" and more to how each company is splitting models by the kind of work they are meant to do.

This is not a hands-on comparison. The notes below are based on official announcements and product positioning. I do not want to declare a winner without running the models against the same tasks, prompts and budget.

Workload tiers instead of one model

OpenAI split GPT-5.6 into three tiers: Sol as the flagship, Terra as the balanced model for everyday work, and Luna as the fastest and most affordable option. On July 30, OpenAI cut Luna's price by 80% and Terra's by 20%.

The interesting part is not another benchmark win for Sol. Routing hard decisions to an expensive model and well-specified work to a cheaper one is now part of the product design. I wrote about Sol, Terra and Luna separately before. The price cut makes that workload-based approach much more practical.

A coding agent can use Sol to resolve uncertainty and write the plan, then hand mechanical implementation and test loops to Luna. Splitting the workflow is starting to matter more than picking one model for every step.

Claude Opus 5 brings the heavy model closer to daily work

Anthropic released Claude Opus 5 on July 24. The company positions it as a daily model that approaches Fable 5 capability at half the price. It is now the default on Claude Max and the strongest option on Claude Pro.

That positioning is notable because the Opus name has usually felt like the expensive model you open for the hardest job. Opus 5 moves the heavy model closer to the everyday desk. Anthropic emphasizes careful verification, root-cause analysis and keeping the thread through long tasks.

I still would not choose a default by raw capability alone. As I wrote in my Claude Sonnet 5 note, the default model does not need to win every benchmark. It needs to produce a good enough result without breaking the flow. Opus 5 makes sense for difficult bugs, vague requirements and multi-step refactors. Sonnet still looks like the steadier default.

Gemini grows the Flash family while Pro waits

Google announced Gemini 3.6 Flash, 3.5 Flash-Lite and the security-focused 3.5 Flash Cyber on July 21. In the same announcement, Google said 3.5 Pro was still being tested with partners and was not yet broadly available.

The clearest signal from Google is specialization. 3.6 Flash is the workhorse for coding, knowledge work and multimodal tasks. Flash-Lite is for speed and volume. Flash Cyber is paired with the CodeMender agent for security work. There is no single Gemini for every job anymore.

That separation makes sense in production. A system classifying a million documents does not need the same model or cost profile as one reviewing a critical vulnerability. Keeping Pro in partner testing is also a useful reminder: a model name can be public before the product is ready for broad use.

Grok 4.5 enters the engineering race

xAI launched Grok 4.5 on July 16 for coding, agentic tasks and knowledge work. According to xAI's own figures, it is served at 80 tokens per second and costs $2 per million input tokens and $6 per million output tokens. It is available through Grok Build, Cursor and the xAI API.

Treating Grok as only the chat model inside X is no longer enough. The 4.5 story is directly about software engineering, tool use and completing work end to end. I am still cautious here: the speed and benchmark numbers come from the vendor. I would not make Grok 4.5 my default before running the same small, measurable task set across my own codebase.

How I would route the work

I read the current market as a routing table rather than a leaderboard:

• Ambiguous architecture, difficult debugging and long research: GPT-5.6 Sol or Claude Opus 5.

• Daily coding, document work and regular agent loops: GPT-5.6 Terra, Claude Sonnet 5 or Gemini 3.6 Flash.

• High-volume work with easy-to-check outputs: GPT-5.6 Luna or Gemini 3.5 Flash-Lite.

• A strong alternative for coding experiments: Grok 4.5, starting with a small and measurable evaluation set.

• Security analysis: a specialist model can help, but findings and production changes still need human verification.

This list will move. Pricing, speed, access and quotas can change within weeks. The selection method should stay stable: define the cost of error and the verification path first, then pick the model. That was also the point of measuring AI investment by completed work rather than tokens.

The model race has quietly become a workflow race

The common theme is not better conversation. Nearly every vendor is talking about long tasks, tool use, token efficiency, cost and adjustable effort. The model is no longer the whole product. It is one part of a system that also needs routing, tools, permissions, tests and human approval.

My short summary of summer 2026 is that searching for one "best model" is becoming less useful. A better setup has a default model, an escalation path for difficult work and a cheap tier for volume. Irreversible actions such as deleting files, writing to production, publishing content and making payments should stay with a human, regardless of how capable the model becomes.