Four Flagship AI Models in One Month: Which Should You Actually Use?

In a span of nineteen days, four companies shipped flagship AI models that, in a quieter year, would each have owned the news cycle for a month. Anthropic opened the run on June 30 with Claude Sonnet 5. OpenAI followed with the GPT-5.6 family on July 9. Elon Musk's xAI — which formally rebranded as SpaceXAI on July 6, completing its fold into SpaceX, according to Yahoo Finance — released Grok 4.5 on July 16. And China's Moonshot AI closed the cluster in mid-July with Kimi K3, a 2.8-trillion-parameter model whose open weights anyone can download. If you are a normal person trying to pick one, the launch events have not helped. Here is the version without the stagecraft.

What makes this cluster matter is not that the models are good — they are all good — but that they have stopped being interchangeable. A year ago, the honest answer to “which AI should I use?” was “whichever your subscription includes.” That no longer holds. The four releases have sorted themselves into genuinely different tools, and the sharpest dividing line runs straight through software development: one camp is faster and cheaper, the other more careful and reliable. For everyone else — students, researchers, people who write and analyze for a living — the decision now turns on a few specific things each model does better than the rest. What follows is a guide to those things, with the numbers attached.

GPT-5.6: the generalist, now faster and cheaper

GPT-5.6 is really three products, not one: Sol, the flagship; Terra, the mid-tier; and Luna, the cheapest. Those are OpenAI's actual model names, not internal codenames — the company previewed all three on June 26, and CNBC confirmed the public rollout. OpenAI launched the family on July 9, one day after Reuters and Axios reported it had cleared U.S. government approval for a broad rollout, a reminder that the top tiers of these systems now clear export-control-style review before they reach you.

The headline technical claim is speed. Per OpenAI's own preview, GPT-5.6 Sol can run on Cerebras hardware at up to 750 tokens per second — fast enough that waiting on the model stops being part of the experience. On raw capability it is competitive but not dominant: a July 2026 SWE-Bench Pro compilation placed the GPT-5.6 Sol entry around 64.6%, clustered with Grok 4.5 and behind Anthropic's top Claude tiers. Where GPT-5.6 actually wins is breadth and convenience. It is the default inside ChatGPT, it handles the widest range of casual tasks competently, and the three-tier structure means you can drop down to Luna when you are burning volume on something that does not need the flagship.

Try it first if you already live in ChatGPT and want one model that is decent at everything without managing a decision.

Claude Sonnet 5: the careful one for writing, review and agents

Sonnet 5 landed June 30, priced to be roughly cost-neutral with the model it replaced: $2 per million input tokens and $10 per million output tokens as an introductory rate through August 31, then $3 and $15 after that, per Anthropic's platform documentation. It is available across the Claude API, Amazon Bedrock, and Google Cloud's Vertex AI. The pitch is value — a model that approaches the capability of Anthropic's top-tier Opus 4.8 at a fraction of the price. Independent analyst Simon Willison, whose assessments of coding models are among the most-cited in the field, put it cleanly in his June 30 write-up: Sonnet 5's performance is “close to that of Opus 4.8, but at” a lower cost.

Where the Claude line keeps winning is care. Long-form writing, document analysis, code review, and the multi-step agent work where you genuinely need the model to not improvise. Willison has been consistent on this point — last September he called the Claude line “probably the best coding model in the world,” and he has broadly stayed there. Sonnet 5 continues that tradition at a price that makes it reasonable to use as a daily driver rather than a luxury.

Try it first if your work is analytical writing, research, or careful code review — or if you run agents that take real actions on your behalf and you want fewer surprises.

Grok 4.5: the coding challenger trained on Cursor data

Grok 4.5 is the most pointed release of the four. It is xAI's first model built specifically for coding and agentic work, and — notably — the first jointly trained on real-world data from the Cursor code editor, the company xAI is now entangled with through the SpaceXAI consolidation. It is available inside Cursor on all plans. Under the hood it is a roughly 1.5-trillion-parameter mixture-of-experts model on the V9 foundation, per OpenRouter's listing, with a 500,000-token context window. Separate professional-grade testing by Snorkel ran Grok 4.5 against GPT 5.5 and Claude Opus 4.8 across 2,000 expert-graded tasks.

The pricing is the aggressive part. Grok 4.5 costs $2 per million input tokens and $6 per million output, with cached input at $0.50, per xAI's developer documentation. That output price undercuts Sonnet 5 by 40% and Anthropic's top Opus tier by roughly 80% — which is why outlets including DevOps.com framed the release as the opening shot in a coding-agent price war. (A note on the name: xAI rebranded to SpaceXAI on July 6, per Yahoo Finance and Business Insider; the same company makes Grok, and you will see both names in its documentation.)

Try it first if you write code for a living, especially if you are already in Cursor, and you measure your monthly model bill.

Kimi K3: the open-source upset

Kimi K3 is the wild card, and the release that matters most to anyone who cares about open source. Moonshot AI's model is a 2.8-trillion-parameter sparse mixture-of-experts system with openly released weights, a one-million-token context window, and native multimodal vision, per Tom's Hardware and Hugging Face — making it the largest open-weight model ever released, roughly 75% larger than DeepSeek's V4 Pro, per VentureBeat.

The number that got developers' attention came from Arena's Frontend Code evaluation, a leaderboard scored by blind developer testing. Kimi K3 took first place at 1,679 points, a seventeen-spot jump from its predecessor, finishing ahead of Anthropic's Claude, OpenAI's GPT-5.6, and GLM-5.2 — the first time an open-weight model has held that position. It also beat Claude on a GPU-kernel optimization benchmark. The API runs roughly $0.30 per million input tokens, per Tom's Hardware, though it is more expensive to run than its predecessor if you self-host.

Try it first if you want to run a frontier-grade model on your own hardware, you do research that needs a massive context window, or you simply want capability without a recurring cloud bill or a vendor relationship.

The coding split, with names and numbers attached

This is where the four models stop being interchangeable and the choice gets real. Independent testing by merge.dev ran Grok 4.5 and Sonnet 5 head-to-head on the same coding task: Grok finished in roughly 60 seconds against Sonnet's 116, and used fewer output tokens to do it. On raw speed and price, Grok 4.5 wins outright. On the SWE-Bench Pro leaderboard compiled in July 2026, Anthropic's higher-tier Claude entries led the field — Fable 5 at 80.3% and Opus 4.8 at 69.2% — with Grok 4.5 (64.7%) and GPT-5.6 Sol (64.6%) clustered just behind. Kimi K3's edge shows up on Arena's frontend benchmark rather than SWE-Bench.

Translated into a decision: Grok 4.5 is the model to reach for when you want a lot of code generated quickly and cheaply. The Claude line is the one to reach for when the code has to be right — in review, in refactoring, in the kind of work where a confident mistake costs more than a slower correct answer. That is my editorial read, but it lines up with what the named benchmarks show and with where the developers I follow have consistently landed, Willison among them. Faster-and-cheaper versus more-careful is the real fork in this market, and for the first time it is wide enough to be worth choosing deliberately.

For non-developers, the fork is simpler and less dramatic. GPT-5.6 remains the safest all-rounder for the widest range of everyday tasks. Sonnet 5 is the better writing and analysis tool. Grok's edge in coding does not translate into a clear advantage for, say, drafting an email, and Kimi K3's appeal for casual users is mostly curiosity unless you have a concrete reason to self-host.

So which one should you actually use?

If you want one subscription and no decisions, stay on ChatGPT and let GPT-5.6 carry the load. If your work is writing, analysis, or careful code, Claude Sonnet 5 is the safest default and a strong value at its launch price. If you write code for a living and you watch your bills, put Grok 4.5 in your editor this week. And if you want to own the model rather than rent it — or you just want to see how good open source has become — download Kimi K3. The useful news from this unprecedented month is not that there is a new single best model. It is that there is no longer a single best model, and the choice has finally become worth ten minutes of your time.