01LLMs & Foundation Models

Overview

This is the head-to-head on the large language models that power everything else in this study — the "brains" your agents, skills, and chat windows run on. The question this report answers isn't "which app should I open," it's "which model should I reach for, and when." By mid-2026 the frontier has split into clear lanes: a flagship reasoning tier for hard problems and writing, a fast/cheap tier for routine work, a precision coding tier, a design-and-multimodal tier, and a research-grounding layer. No single model wins every lane — the operators getting the most leverage route deliberately.

To keep this category focused on raw model capability, the application-layer write-ups live elsewhere: Claude as a writing tool is in Category 02 (Content), the coding harnesses (Claude Code, Codex, OpenCode) in Category 14 (Vibecoding & Workflow Automation), and the autonomous agents in Category 15 (AI Agents & Autonomous Tools). Here we compare the underlying models head-to-head.

A note on families: each provider ships a lineup, not a single model. Rather than card every point release, we consolidate one card per provider and cover the tier choice in prose — which member of the family to pick for which job.

Highly Recommended

Claude (Anthropic)

Highly Recommended ⭐ James Pick

Best for: Daily-driver reasoning and writing, deep debugging, cost-tiered routing across one model family

Claude is the runaway default inside this community and a benchmark leader in coding and general reasoning on the wider market as of June 2026 — community consensus on dev forums, HN, and G2 keeps coming back to reliability, instruction-following, and code quality. The family spans four tiers, and the skill is matching the tier to the task. Opus 4.8 is the default flagship for general-purpose reasoning, code generation, and instruction-heavy work. Sonnet 4.6 is the cost/quality daily driver — strong on straightforward tasks at materially lower cost, which is why it powers always-on agent setups. Haiku 4.5 is the cheap routing tier: route simple classification, data analysis, and high-volume batch work here and reserve the frontier models for strategic thinking. Fable 5 (currently unavailable) is the newest deep-reasoning tier — built for multi-step debugging, autonomous long-horizon tasks, and finding issues other models miss.

The cost discipline matters more than the leaderboard. A lot of operators overspend by running flagship models on work a cheap tier would nail; the members getting real leverage tier deliberately — Haiku for the routine 80%, Opus or Fable 5 for the hard 20%.

"A lot of folks spend more money than they need to because they're using the top frontier models for things that Haiku could do well." — James Dickerson

"The benefit of Sonnet 4.6 is a good cost and quality." — James Dickerson

"It's very strong at debugging and finding issues. It's night and day from Opus models." — James Dickerson

Sentiment: The community's runaway default and the assistant members are most attached to. Tier deliberately — Haiku to route, Sonnet for daily work, Opus as flagship, Fable 5 (currently unavailable) for the hard debugging — and the cost-to-quality ratio is hard to beat.

ChatGPT / GPT-5.5 (OpenAI)

Highly Recommended ⭐ James Pick

Best for: Primary coding agent (precision and codebase understanding), fast bug diagnosis, general chat

OpenAI's GPT-5.5 (and the Codex 5.5 coding variant built on it) drew extremely positive May–June 2026 reception across X, Reddit, HN, and dev forums for coding precision, codebase comprehension, and security review — with reports of edge cases where it outperforms the Claude flagship. For James this flipped his coding workflow: roughly 95% of his coding now runs through Codex/GPT-5.5, which he finds "more precise" and "smarter" at understanding what's happening inside a codebase. Earlier GPT-5/GPT-5.1 releases earned the same praise for fast bug diagnosis — the long-running "AI ping-pong" pattern of building in one model and having GPT zero in on the specific bug. The consumer ChatGPT app remains the most familiar general assistant on the market.

One critique worth flagging from James: ChatGPT's chat interface has drifted toward sycophancy — telling users what they want to hear, which can quietly reinforce a flawed assumption. The fix has been to work against the raw model (GPT-5.5) or in a coding harness rather than the chat app for high-stakes thinking.

"95% of my coding right now is in Codex versus Claude Code. It's more precise, it's smarter, and it understands what's going on better within the codebase." — James Dickerson

"GPT-5 is super smart and fast at diagnosing issues with code." — James Dickerson

Sentiment: The precision coding tier of choice in mid-2026, with a familiar consumer app on top. Strong, broad market enthusiasm — watch the chat app's tendency to agree with you, and verify against the raw model when it matters.

Recommended

Perplexity

Recommended ⭐ James Pick

Best for: Research grounding, fact-checking, real-time web access, deep context-gathering inside agentic workflows

Perplexity is the research layer of the modern stack, and community sentiment (2025–June 2026) across Reddit, HN, and X is overwhelmingly positive — praised for search relevance and fact-grounding, with strong adoption among developers wiring it in via MCP. For James it's the foundational research MCP: when an agent is told to "research this deeply," routing through Perplexity returns far better insights and deeper analysis than a bare web search. It's less a chat model you talk to and more a grounding engine you point your other models at.

"Perplexity MCP is my go-to. I power all my research and context-gathering with it." — James Dickerson

Sentiment: The default research-grounding layer for agentic workflows. Wire it in via MCP and it measurably upgrades the quality of everything downstream.

Gemini 3.1 Pro (Google)

Recommended

Best for: Visual design and mockups, multimodal/visual understanding, cost-competitive reasoning, Gmail-integrated workflows

Google's Gemini 3.1 Pro (May–June 2026) earns its strongest community marks for visual design — improved mockup generation and layout understanding — and is the model James reaches for via Google Stitch when he wants design output that Claude Code struggles to match. Its predecessor, Gemini 3 Pro, remains a useful cross-model debugging fallback: when a coding agent gets stuck, Gemini (or GPT-5.1) often produces a solution worth pasting back in. Beyond design, the family is genuinely multimodal, deeply integrated into Gmail and Google Workspace, and consistently cheaper than the Anthropic tier — a recurring point in its favor for cost-conscious teams.

"Try Stitch 3.1 Pro, that's what I've used." — James Dickerson

Sentiment: The design-and-multimodal pick, cost-competitive and Workspace-native. Strong for visual work and as a debugging second opinion; a solid alternative rather than a primary coding agent.

Situational

Grok (xAI)

Situational

Best for: X-integrated research, real-time social signal, a secondary in-browser assistant

Grok's differentiator is its native tie to X — useful for pulling real-time social signal and trends, and several members keep it open as a secondary browser-based prompt tool alongside their primary stack. Community reception is mixed-to-solid: capable and fast, but not the model most operators reach for first when reasoning depth or code precision is the job.

Sentiment: A handy X-integrated secondary tool rather than a primary brain. Worth having open for social-signal research; reach elsewhere for heavy reasoning or coding.

Up & Coming

DeepSeek

Up & Coming ⭐ James Pick

Best for: Low-cost reasoning, exploratory analysis where you want frontier-ish thinking on a budget

DeepSeek's latest (DeepSeek 4) lands as a genuinely strong, low-cost reasoning model — the kind of option that delivers a lot of the thinking quality of a frontier model at a fraction of the price. James flags it specifically as a reasoning model worth checking out, though without a single locked-in use case yet — it's on the radar as a cost-efficient alternative as it matures.

"DeepSeek 4 is a really good reasoning model, something to check out." — James Dickerson

Sentiment: A rising low-cost reasoning option. Strong value signal and worth testing as a budget alternative; adoption is still early.

Kimi

Up & Coming

Best for: Cost-saving Sonnet-class alternative for everyday agent tasks

Kimi (2.6) is positioned as a low-cost alternative that delivers a lot of the capability of a Sonnet-class model — useful as the everyday model inside an agent harness when you want to keep spend down without dropping to a weak tier. It's a "try this to save money on routine work" option rather than a frontier contender.

"You can also try lower-cost models like Kimi 2.6, which give you a lot of the capabilities of a Sonnet." — James Dickerson

Sentiment: A credible low-cost Sonnet-class stand-in for routine tasks. Good for keeping agent costs down; not a frontier replacement.

Ollama (local models)

Up & Coming

Best for: Running open models (Llama and others) locally, cost-free and private inference

Ollama is the simplest way to run open-weight models — Llama and friends — locally on your own machine. The appeal is cost-free, private inference with no per-token bill and no data leaving your environment, which makes it attractive for high-volume routine work or privacy-sensitive tasks. Local open models trail the hosted frontier on raw capability, so the trade-off is cost and control versus peak quality.

Sentiment: The go-to for free, private, local inference. A genuine lever for cost and privacy on routine work; expect a capability gap versus the hosted frontier.

James's Model Evolution (Aug 2025 → Jun 2026)

The fastest way to read this category is to watch how one heavy operator's primary coding model shifted over ten months. The throughline: Claude Code anchored building through 2025; Codex/GPT-5.5 took over the coding seat in mid-2026 for precision; and by June 2026 the workflow is multi-model — Fable 5 for deep debugging, Codex 5.5 alongside it, Gemini 3.1 Pro for design — run through model-agnostic harnesses. Design has been the persistent weak spot for Claude Code throughout, which is what pushed the diversification.

"The big unlock for me was Opus 4.6 and Codex 5.3 working side by side. I'm seeing a massive increase in code quality and output quality." — James Dickerson

What Our Community Uses

Alongside the market research above, we surveyed Vibe Marketer members (May 18 – June 7, 2026) and drew on six months of community session transcripts. Here's how this category actually looks inside the community.

Claude is the runaway default

When members listed the AI assistants they use regularly for marketing work, Claude led every other model by a wide margin — and it's also the model members are most attached to. When the study-wide survey asked "What's the single tool you'd be most devastated to lose access to tomorrow?", the overwhelming majority of answers were Claude or Claude Code. ChatGPT, Gemini, and Grok all show up as secondary tools, but none come close to Claude's mindshare here.

Claude(4.18) 22

ChatGPT(3.79) 13

Gemini(3.75) 11

Perplexity(4.19) 8

Grok(3.79) 6

Copilot(3.17) 3

A few patterns stand out beyond the raw counts. Claude doesn't just win on usage — it carries one of the highest average ratings in the set (4.18), and members are blunt about why: "Claude simply outperforms everything on the market." (Scott Cowan) — "Claude is capable enough to work on most of the tasks." Perplexity posts the single highest average rating (4.19) despite a smaller footprint, consistent with its role as a quietly indispensable research layer rather than a chat window. ChatGPT plays a defined second-seat role and was also the most-dropped assistant — several members tried it and switched to Claude ("I use ChatGPT for images only"). Gemini and Grok round out the secondary tier, valued by their users but rarely the primary brain.

The takeaway mirrors James's own evolution: Claude anchors the stack, the other frontier models get pulled in for specific lanes — design, debugging second opinions, social signal, research — and very few members run just one model for everything.

🧩 Skills That Fit Here

Skills aren't separate tools — they're reusable instructions that make your AI do a job your way, every time, on top of whichever model or platform you point them at — they work across Claude, ChatGPT, and other modern models, not just one. On a foundation-models page that matters because the skill is the playbook and the model is the engine: the same skill can route its routine steps to a cheap tier like Haiku and escalate the hard reasoning to a flagship tier.

• Start Here — orients you on which model tier to reach for before you spend a token on the wrong one.
• Direct Response Copy — runs the same persuasion framework whether it's executing on Opus, GPT-5.5, or a cheaper tier.
• SEO Content — turns research (often grounded through Perplexity) into structured drafts a mid-tier model can produce at volume.
• Creative Engine — orchestrates a multi-model flow, handing visual work to Gemini and reasoning to the flagship.

🧩 The community's Vibe Skills pack ($199, one-time) bundles these. James's own advice is to "get deep into Claude Code and figure out how to leverage skills" — the model is the engine, the skill is what makes it produce consistent, on-brand output.

Pricing note: Model pricing and tiers change frequently — subscription prices, per-token API rates, and which models sit in each tier all move month to month. Verify current rates on each provider's pricing page before committing, and remember that the cheapest viable tier for a given task usually beats defaulting to the flagship.

All data sourced from community discussions, published reviews, independent benchmarks, and provider announcements from April–June 2026. James's quotes provide use-case color; verdicts are grounded in community and market sentiment. Tool list limited to models with meaningful recent signal; models not appearing above lacked sufficient sentiment data to assess with confidence.