Three flagship AI models in eight weeks. GPT-5.4 on March 5, Gemini 3.1 Pro on April 8, Claude Opus 4.7 on April 16. The frontier moved twice in less than two months and most people are reasonably wondering which one to actually pay for.
This isn't a benchmark-only post. Benchmarks are useful but they don't tell you what to pick when you're sitting in front of an empty chat box. So below: what's genuinely new in each release, what each one is actually best at right now, and the free open-weights models that are surprisingly close to the paid frontier if you don't want to subscribe to anything.
GPT-5.4 — the computer-use specialist
GPT-5.4 is OpenAI's first release where browser-and-OS agent work is the headline feature, not a side project. The model set new state-of-the-art numbers on OSWorld (computer-use benchmark) and WebArena (web-task benchmark), with reported success rates roughly 12-15 percentage points above the previous best.
What that means in practice: if you're building agents that click around websites, fill in forms, or operate desktop apps, GPT-5.4 with the new Computer Use API is currently the model to use. The general-chat experience is a smaller incremental upgrade over GPT-5.
Pricing held steady — $5/$15 per million tokens (input/output) for the base model, with the Computer Use API at a premium. Free tier on ChatGPT now includes limited GPT-5.4 access with Plus tier ($20/month) getting unrestricted use.
Gemini 3.1 Pro — the science model
Google's headline number for Gemini 3.1 Pro was 94.3% on GPQA Diamond, which is a graduate-level science benchmark and a meaningful jump over the previous frontier. Translation: if you're working in chemistry, physics, biology, or anything quantitative-research-adjacent, Gemini 3.1 is currently the strongest pick.
The 2-million-token context window remains. Long-document work — feeding it an entire textbook or a sprawling legal contract — is still where Gemini's UX is the most polished, partly because Google's Workspace integration makes the "drop a folder of PDFs" workflow effortless.
Pricing: $2.50/$10 per million tokens, the cheapest of the three flagships by a meaningful margin. Free tier on AI Studio remains generous if you're a developer.
Claude Opus 4.7 — the long-running agent
Anthropic positioned Opus 4.7 around extended autonomous work — agents that run for hours rather than minutes. The technical change underneath is a substantial improvement to the model's ability to retain context and stay coherent across very long task chains. Opus 4.7 has been demonstrated keeping a coherent thread across 8+ hours of agent work, which is roughly double what 4.5 could reliably do.
The conversational experience also improved — the model is noticeably better at following multi-step instructions and at saying "I don't know" instead of confidently making things up. It's currently my personal pick for code-heavy work, though that's preference rather than benchmark fact.
Pricing: $15/$75 per million tokens, by far the most expensive. The 1-million-context tier costs more again. Claude Pro at $20/month is the sensible entry point unless you're building APIs.
Quick rule of thumb
- Coding, writing, long-running agents: Claude Opus 4.7
- Computer-use / browser agents: GPT-5.4
- Science, long-document analysis, anything cost-sensitive: Gemini 3.1 Pro
- Image generation in the same chat: GPT-5.4 (still has the most polished image stack)
- Casual chat, open-ended creative work: Personal preference; all three are excellent
The open-weights story (which is surprising)
The story underneath the flagship news is that the open-weights models had an unusually strong April too. Two releases stood out:
Qwen3.6-Max-Preview (Alibaba, April 20). The flagship Qwen model, open-weights, performance close to GPT-5 on most reasoning benchmarks. Runs on your own hardware if you have the VRAM.
Kimi K2.6 (Moonshot AI, April 21). Strong all-rounder, particularly good at long-context reasoning. Also open-weights.
The gap between the open-weights frontier and the closed-weights frontier is now small enough that for many tasks — code completion, document summarisation, structured data extraction — running an open model locally is genuinely competitive with paying for a flagship.
How to run these locally on Windows
You don't need to set up a Python environment or write any code. Two options that just work:
LM Studio — friendly Windows app, browse Hugging Face from inside it, click "Download" on a model, and you're chatting locally a few minutes later. Realistic minimum: 16 GB RAM and an 8 GB VRAM card. With a 12 GB card you can run the medium-sized open-weights models comfortably.
Ollama — terminal-first, lighter weight. ollama run qwen3 downloads and runs the model. Pair with Open WebUI for a browser chat interface if you want.
Neither tool sends anything to the cloud. Once the model is downloaded, you can pull the network cable and everything still works.
A note on cost over a year
Quick mental math, because this stuff adds up. ChatGPT Plus + Claude Pro + Gemini Advanced is $60/month if you keep all three, which is $720/year. A serviceable used RTX 3090 with 24 GB VRAM is around $700-800 right now and runs the open-weights frontier comfortably. The break-even is shorter than it used to be.
That doesn't mean you should always run locally. Frontier models will keep moving and the closed labs will keep being slightly ahead on the absolute peak. But "always pay for the frontier" stopped being the obvious answer in 2026.
What about voice and image?
Speech and image generation are sister markets that have moved in parallel to the text-LLM frontier. For voice, open-source XTTS v2 / F5-TTS / IndexTTS-2 are now genuinely competitive with ElevenLabs for most everyday use. For images, Stable Diffusion XL and Flux Schnell run locally and produce results that are hard to distinguish from Midjourney for most prompts.
If you're interested in the voice side, I write more about that on this blog — covered in the open-source voice cloning roundup and the ElevenLabs pricing comparison.