After six months of running all three premium AI subscriptions in parallel for actual paid work, the verdict is lopsided enough that I am writing it down before Google has a chance to change it.
Tomorrow’s Google I/O 2026 keynote could move the board, but going in, one of these three is doing most of the heavy lifting and the other two are not as close as the marketing copy suggests.
What I Actually Run These Things For
This is not a benchmark post. The three tools sit on my desk because I write for a living and ship code for a living, and the question I care about is which one earns its monthly bill on a Tuesday at 3pm when I am tired and the deadline is real.
The setup is boring. ChatGPT Plus, $20 a month, on GPT-5.5 since the April 23 launch. Claude Pro, $20 a month, running Opus 4.7 since its release on April 16. Gemini, $20 a month under the Google AI Pro tier, on whatever Google labels its top consumer model that week. Sixty dollars a month, roughly the cost of a streaming service stack, in exchange for what is supposed to be three competing intelligences.
Six months of receipts later, two of those bills feel like rent and one feels like a tax.
ChatGPT: the Consumer Surface, Stretched Thin
ChatGPT is the one I open when I want to point a camera at a thing and ask what it is, when I want to talk to my phone in the car, or when I need an image generated on the spot. OpenAI has been building a consumer dashboard rather than a thinking tool, and on that axis it is winning. Voice mode is genuinely good. The built-in image generator is the only one of the three that produces a usable image inside the same conversation. The autonomous web agent will sometimes go book a thing, and the success rate is real if you keep the task narrow.
What it has stopped doing well, at least for me, is the actual cognitive work. The release of GPT-5.5 was supposed to settle the coding question. It did not. On long technical edits across multiple files, the model still hallucinates function signatures, still loses track of which file it is editing by the third turn, and still treats my instructions as suggestions. None of that is fatal in a one-shot chat. It is fatal in a four-hour debugging session.
The price gap is the other problem. ChatGPT Pro is $200 a month, ten times the Plus tier, for what amounts to higher message limits and a Pro reasoning mode. There is a real product there for a specific buyer. For most paying customers, it is a wall the consumer tier hits and then keeps hitting.
Gemini: Cheap, Integrated, and Still Looking for the Win Column
Gemini does the thing nobody else does. It lives inside Google’s stack. With permission, it reads my Gmail, summarizes my Docs, watches my Calendar, and the Workspace integrations are, after eighteen months of false starts, finally what they should have been at launch. The pricing on the API is also the most aggressive in the market, with Gemini 3 Pro at $2 per million input tokens against $5 at OpenAI and Anthropic.
I use Gemini’s image model under the hood for almost every published hero image on this site, which is its own form of vote. The image quality is excellent and the per-image cost is a fraction of the competing options.
Everything else is harder to defend. On extended reasoning the model is inconsistent. Hand it a thirty-page brief and the summary is fine. Ask it to push back on a flawed argument and it tends to soften, hedge, and look for the middle. That is a posture, not a capability gap, but the posture costs me work because I want the AI to argue with me, not flatter me.
The other thing I notice: Gemini is gated by Google more than the others. Refusals are more frequent, the safety language is more corporate, and the model will sometimes simply not engage with topics that Claude or ChatGPT will discuss freely. Some of that is reasonable product caution. Most of it feels like product-counsel anxiety leaking into the assistant.
Claude: the One Doing the Work
The honest answer, and the one I expected to feel less comfortable writing, is that Claude is the one I open when something has to actually ship. Opus 4.7 tops SWE-bench Verified at 87.6 percent in the Anthropic launch announcement, the highest mark from any frontier model released to the public, and that ranking matches what I see in practice. It edits across files. It holds context across long sessions. It writes long-form copy that reads like it was written by a person who reads.
Anthropic’s release of Opus 4.7 came with an unusual second admission. In the same announcement window, the company conceded to Axios that an unreleased internal model, codenamed Mythos, is materially more capable and is being held back under a restricted program called Project Glasswing. CNBC reported the same day that Mythos showed roughly a 90x improvement over Opus 4.6 on exploit development, including the discovery of a 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg bug during testing.
I do not love every implication of that disclosure. A company telling the public that its real frontier is hidden because the public version of the frontier is too dangerous to ship is a story about corporate self-restraint, and corporate self-restraint is exactly the kind of governance question that should not be decided behind a closed door at a single private lab. The disclosure is also, on its face, more honest than what OpenAI or Google have offered about their internal capabilities. Both things can be true.
For now, the public Claude is the public-frontier model in actual practice, and the gap is widest on the work that produces real economic output: code, analysis, structured writing, multi-step reasoning. The 70 percent developer preference figure cited in the Opus 4.7 launch reads less like marketing copy and more like a stylized fact.
What Tomorrow Could Change
Google I/O 2026 opens tomorrow, and the rumored Gemini Omni model, which is expected to ship a single pipeline for text, image, and video generation, is the closest thing Google has had to a real frontier swing in two years. If the demo is as broad as the leaks suggest, the consumer tier of Gemini could collapse the gap with ChatGPT on multimedia, and the price advantage on the API gets sharper. None of that fixes the reasoning posture, but it is a real product move.
What it does not change, at least not on Day One, is the work-product question. Google has not, in this cycle, shipped a model that holds up against Opus 4.7 on the tasks that pay my rent. Until that ships, the $20 I send Anthropic each month is the cleanest line on the credit card statement, the $20 to Google is the lowest-risk hedge on a deeply integrated future, and the $20 to OpenAI is the bill I think hardest about renewing.
Tomorrow could change one of those three. It could also just confirm them.