We ran the same 5 prompts across all three on day one. Here are the results, with the winner for Indian-context use cases. (Spoiler: it's not the one with the highest benchmark.)
July 2026 was a wild month. OpenAI shipped GPT-5.6 (July 9), Anthropic shipped Claude Sonnet 5 (June 30, still fresh), Google shipped Gemini 3.1 Pro (early July). All three are flagship-tier. All three are within 5% of each other on most benchmarks. All three are bad in different ways. Here's the actual data from 5 prompts we ran on day one.
We picked prompts that mirror what 2BFT actually does: (1) Translate a Stashed product page from English to Tamil + Hindi, (2) Write a 200-word WhatsApp message to a jewellery shop owner explaining how to use Nano Banana 2 for catalogue shoots, (3) Debug a 50-line Python script that's hitting a weird MCP error, (4) Summarize a 30-page Indian government PDF in plain English, (5) Write a 6-image Instagram carousel for the Maruthi Jewellers Eid collection.
Each prompt was run three times on each model. The best result per prompt is scored below.
GPT-5.6 won prompt #1 (Tamil + Hindi translation) cleanly. Specifically: it kept the formal register when asked, didn't transliterate, and produced a 3-paragraph version that read like a native Tamil speaker wrote it. Claude Sonnet 5 was 80% as good, Gemini 3.1 Pro was 90% as good. Sonnet 5 had a slight "translation-ese" feel; Gemini 3.1 Pro hallucinated one phrase ("cultural motif of the region" was invented).
Claude Sonnet 5 won prompts #2, #3, and #4. The WhatsApp message came out short, warm, and used exactly the right level of Hindi/Tamil code-mixing for a real Indian shop owner. The Python debug caught the actual bug on the first pass and explained the fix in 3 sentences. The 30-page PDF summary was 8 paragraphs, didn't make anything up, and cited specific sections by page number.
GPT-5.6 was the runner-up on those three — strong, but consistently 20-30% longer than necessary. Gemini 3.1 Pro struggled on prompt #3 (the Python debug) and prompt #4 (the PDF summary) — it over-explained and added things that weren't in the source.
Gemini 3.1 Pro won prompt #5 (the 6-image Instagram carousel). Specifically: the structure of the carousel, the hook on the first slide, and the CTA on the last slide all read like a real Instagram strategist wrote it, not a model. Claude Sonnet 5 was a close second. GPT-5.6 produced something that felt more "LinkedIn carousel" than "Instagram carousel".
Claude Sonnet 5. By a clear margin, on 3 of 5 prompts. The voice, the code-switching, the cultural awareness, the brevity — all of it tuned better for what 2BFT does daily.
But here's the catch: GPT-5.6 is the better default if you're shipping an English-only product for a global audience. Gemini 3.1 Pro is the better default if you're shipping anything image-heavy (Nano Banana 2 is the engine, and it shows).
Our default stack in July 2026: Claude Sonnet 5 for copy, WhatsApp messages, code debugging, anything Tamil + Hindi. GPT-5.6 for the long-context things (it's the cheapest big-context model right now). Gemini 3.1 Pro for anything with images. And the agent framework is OpenClaw, with Claude as the default model for the 105 systems on /agentic-systems.
If you're new to this, start with one model. Use it daily for a month. Then add a second. Don't pay for three on day one.
2BFT is a 2-person studio in Vaniyambadi, Tamil Nadu. We test every major model on day one, ship real agentic systems, and teach builders across India. Free Academy, free newsletter, 200+ curated resources, 100+ copy-ready systems.
