AI Field Guide
Chapter Four

The Tools,
Not the Models

Claude Code, Google Antigravity, Codex, Cursor and the rest — where the real competition is happening, and whether you should be looking over the fence.
Compiled 14 July 2026 · this is the chapter that matters most to how you actually work

Chapter FourFirst, the correction


Antigravity is not a model

You listed it alongside Claude, Kimi and DeepSeek. It doesn't belong there. Google Antigravity is an agent-first development platform — an IDE, a CLI, and an SDK for orchestrating coding agents. It is powered by Gemini models (and, notably, by Claude models too).

So it doesn't compete with Claude-the-model. It competes with Claude Code. And that turns out to be a far more interesting comparison for you, because it's a comparison between the tool you live in all day and its most credible challenger.

The mistake is worth making, because correcting it reframes the question. You don't really want to know "which model is best" — you have that from Chapters Two and Three. You want to know whether the thing you spend eight hours a day inside is still the right thing. That's this chapter.

Chapter FourWhat changed in Claude Code while you weren't looking


Before comparing outward, it's worth noting what shipped in the tool you already pay for — because based on your usage history, you are not using most of it.

Dynamic Workflows — the big one, shipped May 2026

Claude now writes a JavaScript orchestration script that fans out subagents: up to 16 concurrent, 1,000 per run, executing in a background runtime while your session stays responsive, with a grader loop that sends subagents back to revise until the output meets a rubric.

The proof point is genuinely startling. Jarred Sumner used it to port Bun from Zig to Rust — roughly 750,000 lines, 99.8% of the test suite passing, in eleven days.

Intermediate results live in script variables, not your context window — which is why it can go so wide without drowning. You trigger it with the keyword ultracode: in a prompt, or /effort ultracode for a whole session.

Also new since you'd have last looked: Routines (scheduled agents on Anthropic's infra that survive your laptop being closed — this is what your AI Brief cron actually wants to be), Agent Teams (3–5 peer agents that can argue with each other to disprove one another's theories), Channels (push events from Telegram/Discord/webhooks into a session), and /goal (a separate evaluator model re-checks a completion condition after every turn and restarts Claude until it holds).

Chapter FourThe master comparison


ToolSurfacesModelsOrchestrationPriceStandoutWeakness
Claude Code
Anthropic
CLI, VS Code, JetBrains, Desktop, Web, iOS, Slack, CI Sonnet 5, Opus 4.8, Fable 5 (Anthropic only) Best in class. 1,000 subagents/run, grader loops, 3-level hierarchy $20 / $100 / $200
no free tier
Orchestration depth + richest extensibility (Skills, Hooks, MCP, SDK) Most expensive floor; opaque limits; single-vendor models
Antigravity 2.0
Google
Desktop app, Go CLI, SDK, Managed Agents API Gemini 3.5 Flash (default), Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.6, GPT-OSS Manager Surface, ~5 parallel agents, per-subagent model choice, scheduled tasks Free tier
$20 / $100
Artifacts + browser surface — the agent runs your app, clicks it, hands back a recording Two months old; free tier dies fast; Google killed Gemini CLI with 30 days' notice
Codex
OpenAI
Web, CLI, IDE, iOS, Slack, GitHub GPT-5.6 Sol / Terra / Luna Local + parallel cloud tasks Free · Go $8 · $20 · $100 / $200 Cheapest serious on-ramp; ~4× fewer tokens than Claude Code Thinnest extensibility of the big three
Cursor
Anysphere
AI-first editor (VS Code fork) + CLI Composer 2.5 + 40+ third-party Agent mode; shallower $20 / $60 / $200 Best editor ergonomics; Composer 2.5 is 3rd on the index at 10–60× lower cost Pricing restructured repeatedly; you must live in their editor
Copilot
GitHub
VS Code, JetBrains, github.com, CLI Multi-vendor via Agent HQ Control plane over other people's agents Free · $10 · $39 · $100 Inline completions are free and burn zero credits on all paid plans Not the leader on any single axis
Amp
Sourcegraph
Terminal + every major IDE Multi Named subagents (Oracle, Librarian, Painter) Free, ad-supported
$10/day API cap, zero markup
Most transparent pricing in the industry Ad-supported; smaller ecosystem

Chapter FourClaude Code vs Antigravity, head to head


The framing the neutral reviewers keep landing on, and it's a good one:

Claude Code is optimised for controlled collaboration. Antigravity is optimised for autonomous execution. Claude Code feels like an AI developer; Antigravity feels like an AI engineering department.
 Claude CodeAntigravity 2.0
ShapeA CLI that dissolves into your existing setupA desktop app you adopt as your workspace
Orchestration ceiling1,000 subagents, grader loops, 3-deep hierarchy~5 parallel agents, but per-subagent model choice
VerificationYou review diffs. Chrome surface for debuggingArtifacts — the agent runs the app in a browser and returns screenshots and recordings as evidence
ModelsAnthropic onlyMulti-vendor — including Claude Sonnet 4.6 and Opus 4.6
ExtensibilitySkills, Hooks, MCP, plugins, Agent SDKAGENTS.md, Skills, Hooks, MCP, plugins — near-parity on paper, thinner in practice
Cost to try$20/mo minimum$0
Maturity341 releasesTwo months old as a standalone app
Do not decide on the benchmarks you'll see quoted

The most-cited head-to-head figures (Claude Code at 87.6% SWE-bench Verified vs Gemini 3.5 Flash at 76.2% Terminal-Bench) are not comparable — two different benchmarks, and the Claude number uses an outdated model. Whoever quotes those at you hasn't read them.

My honest recommendation for you

Don't migrate. Your entire operating system is built out of Claude Code primitives — the erp-builder and core-2.0-dashboard skills, your MCP servers, your launch.json setups, your memory files, your deploy-to-Pages habits. Antigravity has analogues for most of it, but you'd be re-authoring all of it, and its orchestration ceiling (~5 parallel agents) is well below what Dynamic Workflows already gives you today for free.

But do spend one evening on it — and here's the specific reason

Antigravity's genuinely differentiated capability is Artifacts and the browser surface: the agent runs your app, clicks through the UI, and hands back a screenshot or a screen recording as evidence it worked.

Look at your own tool usage. You called preview_screenshot 567 times and preview_eval 1,434 times in two weeks. Your memory file records that "framer mount-anims get stuck here so keep static." Visual verification of React UIs is, empirically, one of your biggest time sinks. Antigravity is the only mainstream tool that has built a first-class answer to exactly that problem.

It costs nothing and one evening. Run one self-contained job on it — a Sandow or Elite Mobile screen where "does the UI actually work" is the hard part — and see whether the Artifacts loop beats your current screenshot-and-squint cycle.

Two things to watch before you'd ever take it seriously

  1. Whether the free tier survives a real day. Reported consensus is that "a handful of consecutive agent requests is often enough to end the day's allowance." Fine for evaluation, not for work.
  2. Whether Google does it again. They killed Gemini CLI with roughly 30 days' notice in June 2026 and broke people's CI pipelines. You have production deploys on Cloudflare. A vendor that retires a CLI mid-flight is a vendor you keep at arm's length from anything load-bearing.
The realistic end state

Both, not one. Claude Code stays your primary engine — orchestration depth, code quality, and the skill library you've already built. Antigravity, if it earns it, becomes the thing you reach for on UI-heavy prototyping where visual verification is the bottleneck. That's exactly the split the neutral comparisons land on, and it happens to map cleanly onto the split in your own work.

Chapter FourThe rest of the field, briefly


Chapter FourThe uncomfortable observation


The genuinely important finding of this chapter isn't about Antigravity at all.

You are paying for a Ferrari and driving it in first gear

Claude Code shipped Dynamic Workflows (1,000 orchestrated subagents), Routines (scheduled cloud agents), Agent Teams, /goal, hooks, and worktrees. In 1,318 prompts across two weeks, you have used essentially none of them.

The gap between Claude Code and Antigravity is real but modest. The gap between the Claude Code you're using and the Claude Code that exists is enormous. Closing that is worth far more to you than switching tools — and it's what Chapters Five and Seven are about.

Sources. Claude Code docs (code.claude.com) & changelog · Anthropic Dynamic Workflows announcement · Google I/O 2026 developer keynote · Antigravity docs & the Gemini CLI transition notice · OpenAI Codex pricing docs & rate card · Cursor docs · GitHub Copilot billing changes · Artificial Analysis · InfoQ · MarkTechPost · MindStudio. Antigravity's top pricing tier is secondary-source only and unverified against a Google page.