AI Coding Models & Agent Tools: The July 2026 State of Play
A practical July 19, 2026 snapshot of Claude, GPT, Gemini, and Grok coding models—and the Claude Code, Codex, Antigravity, Jules, and Grok Build agent stacks around them.
The AI coding market no longer has one meaningful leaderboard. By July 2026, every major lab is shipping both a model portfolio and an agent harness: the model supplies judgment, while the harness supplies tools, permissions, parallel workers, persistence, and a place to run.
That distinction matters. A frontier model in a thin terminal loop can be the wrong choice for a long, failure-prone migration. A slightly smaller model inside a mature harness with worktrees, checkpoints, and reliable resume can finish more real work. Choosing an AI coding stack now means choosing both layers.
Snapshot date: July 19, 2026. Model aliases, preview labels, quotas, and product availability change quickly. This article marks preview and experimental surfaces explicitly and links to the vendors’ current documentation rather than treating this table as timeless.
The Short Version
| Provider | Current model ladder | Main coding-agent surface | Important status note |
|---|---|---|---|
| Anthropic | Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5 | Claude Code: CLI, IDE, desktop/web, subagents, dynamic workflows, Agent Teams | Fable 5 is GA; Agent Teams remain experimental |
| OpenAI | GPT-5.6 Sol, Terra, Luna | Codex across desktop, CLI, IDE, cloud, exec, and SDKs | Surface maturity varies; codex exec is stable, some cloud/app-server commands are experimental |
| Gemini 3.5 Flash, 3.1 Pro Preview, 3.1 Flash-Lite | Antigravity and Antigravity CLI; Gemini API Managed Agents; Jules | 3.1 Pro is Preview; consumer Gemini CLI users have been moved toward Antigravity CLI | |
| xAI | Grok 4.5 | Grok Build terminal agent and open-source harness | Grok Build launched in early beta; its source was published July 15 |
There is no universal winner. The useful question is: which combination is best for this workload, under this failure model, at this volume?
Anthropic: Four Model Tiers, Several Kinds of Parallelism
Anthropic’s current model overview presents four broadly available tiers:
- Claude Fable 5 (
claude-fable-5) is the most capable widely released model and is positioned for long-running agents and the hardest judgment work. - Claude Opus 4.8 (
claude-opus-4-8) targets complex agentic coding, large refactors, and enterprise work. - Claude Sonnet 5 (
claude-sonnet-5) is the speed/intelligence balance for everyday implementation. - Claude Haiku 4.5 (
claude-haiku-4-5-20251001) remains the fastest economical tier for fan-out, scans, and high-volume supporting work.
The practical strategy is not “Fable everywhere.” It is cheap fan-out, expensive judgment: let Haiku or Sonnet map files and run mechanical checks, then spend Fable or Opus on architecture, adversarial verification, and final synthesis.
Claude Code now offers multiple parallel-work abstractions, and they should not be conflated:
- Standard subagents get their own context and return a result to the caller. Since Claude Code v2.1.172, subagents can spawn nested subagents to a fixed maximum depth of five. Only the top-level summary needs to reach the main conversation; an agent at depth five no longer receives the Agent tool.
- Agent Teams coordinate independent Claude Code instances with a shared task list and direct teammate messaging. They remain experimental and disabled by default. Teams cost more tokens, do not isolate teammates in worktrees automatically, and do not support nested teams.
- Dynamic workflows move large orchestration into a generated workflow structure, with background execution and
/workflowsobservability. Recent Claude Code releases added workflow sizing and telemetry controls; see the v2.1.202 release. - Background sessions and worktrees are the better fit when separate tasks need full session lifetimes or isolated checkouts rather than a brief delegated answer.
This is the most varied native orchestration menu in the group, but it is not one infinitely nested tree. The five-level ceiling is fixed, and each abstraction has a different lifetime and coordination boundary.
OpenAI: GPT-5.6 Plus the Broadest Codex Surface
OpenAI’s model catalog turns GPT-5.6 into a three-tier decision:
- GPT-5.6 Sol (
gpt-5.6-sol, withgpt-5.6as the starting alias) is the flagship for complex professional reasoning and coding. - GPT-5.6 Terra (
gpt-5.6-terra) balances intelligence, latency, and cost. - GPT-5.6 Luna (
gpt-5.6-luna) targets cost-sensitive, high-volume work.
The tiering resembles Anthropic’s role-based strategy: Sol for ambiguous multi-step work and final decisions, Terra for everyday agents and supporting reviewers, Luna for throughput. The right mapping still needs an evaluation on your own repository; vendor benchmarks do not capture your build time, tool reliability, or project conventions.
Codex’s differentiator is how many product surfaces share the same working model. It is available through the ChatGPT desktop experience, a terminal CLI, IDE integration, and cloud execution. For automation, codex exec is a stable non-interactive command that can emit normal output or JSONL and resume a previous exec session. For embedding, the Codex SDK exposes persistent coding threads in TypeScript and Python; the Python SDK is still beta.
Codex also supports real recursive delegation, but the safe default is intentionally shallow. In the configuration reference, agents.max_depth defaults to 1: the root can spawn direct children, but those children cannot spawn descendants unless you raise the setting. Deeper trees are configurable, not free—they amplify token use, latency, local resource pressure, and runaway fan-out risk. agents.max_threads defaults to six and caps concurrency, but it cannot make a poorly bounded recursive plan predictable.
Use Codex when you want one agent environment to span interactive work, scripted exec, programmatic SDK control, and configurable subagent trees. Treat “supports deeper nesting” as a control surface, not as a quality score.
Google: Gemini 3.5 Flash and the Move to Antigravity
Google’s Gemini model page makes the production status unusually clear:
- Gemini 3.5 Flash is Stable and described as Google’s most intelligent model for sustained frontier performance on agentic and coding tasks.
- Gemini 3.1 Flash-Lite is Stable, optimized for speed, scale, and cost efficiency.
- Gemini 3.1 Pro is still Preview, despite its advanced reasoning and coding position. Preview endpoints can have tighter limits and shorter deprecation windows than stable models.
For most new production integrations, 3.5 Flash is the obvious baseline to evaluate first. Use 3.1 Flash-Lite for high-volume classification, scans, and simple tool calls. Use 3.1 Pro Preview only when its behavior wins your evaluation and you can tolerate preview lifecycle risk.
The tooling story changed more sharply than the model story. Google announced a transition from Gemini CLI to Antigravity CLI. From June 18, Gemini CLI and Gemini Code Assist IDE extensions stopped serving consumer/free, Google AI Pro, and Google AI Ultra access through the old path; Standard and Enterprise customers were not part of that consumer cutoff.
Antigravity CLI keeps key Gemini CLI concepts—skills, hooks, subagents, and extensions repackaged as plugins—but Google warns that the first release does not have one-to-one feature parity. Its direction is a unified, server-backed multi-agent platform with asynchronous background workflows shared with Antigravity 2.0. That makes it attractive for Google-centric agent work, but migration testing is part of adoption, not an optional cleanup task.
Jules is a separate choice, not merely the old CLI with a new name. Jules remains a Public Beta, GitHub-integrated cloud coding agent. It spins up a fresh VM, works asynchronously, and is strongest for scoped repository tasks, scheduled maintenance, CI repair, and pull-request delivery. Choose Jules when “submit work and return later” is the desired interface; choose Antigravity CLI when the terminal and multi-agent interaction are central.
xAI: Grok 4.5 and a Newly Open Grok Build Harness
xAI launched Grok 4.5 on July 16 as its strongest model for coding, agentic tasks, and knowledge work. It is available through the xAI API and is now the default model in Grok Build.
Grok Build is a terminal coding agent with plan review, diffs, AGENTS.md support, skills, plugins, hooks, MCP, parallel subagents, and worktree integration. Its -p headless mode makes it usable from scripts and automation, while ACP provides another integration boundary. The product launched in early beta in May, so reliability and interface churn should be evaluated accordingly.
The notable July update is architectural transparency: xAI open-sourced the Grok Build agent loop and TUI on July 15. The published code includes context assembly, tool dispatch, terminal UI, and the extension system. It can also be built local-first and pointed at local inference. Open source makes the harness inspectable and extensible; it does not, by itself, guarantee that the hosted model, quotas, or beta interfaces will remain unchanged.
Grok Build is compelling when you want a fast terminal-first agent, inspectable harness code, and many parallel worktree workers. Because both Grok 4.5 and the open-source release are days old in this snapshot, start with a representative repository trial rather than a wholesale migration.
Choose by Workload, Not by Logo
| Workload | Strong starting point | Why |
|---|---|---|
| Architecture, ambiguous debugging, adversarial verification | Fable 5 or GPT-5.6 Sol | Spend the frontier tier where wrong decisions are expensive |
| Everyday implementation with human review | Sonnet 5, GPT-5.6 Terra, or Gemini 3.5 Flash | Better balance of intelligence, latency, and throughput |
| High-volume scans and mechanical fan-out | Haiku 4.5, GPT-5.6 Luna, Gemini 3.1 Flash-Lite | Supporting work rarely needs the most expensive judgment tier |
| Deterministic or large Claude-native orchestration | Claude Code dynamic workflows | Workflow structure, background execution, and native monitoring |
| One stack across desktop, CLI, IDE, cloud, and embedded automation | Codex | Broad surfaces plus stable exec and SDK control |
| Google-centric asynchronous multi-agent work | Antigravity CLI | Unified Google agent harness and background workflows |
| GitHub tasks that should run remotely and return a PR | Jules | Cloud VM, scheduled work, CI loop, asynchronous delivery |
| Terminal-first parallel work with an inspectable harness | Grok Build | Headless mode, subagents, worktrees, and open-source agent loop |
A Better Evaluation Than “Which Model Is Best?”
Run the same real task through each candidate stack and record five things:
- Patch correctness: did it fix the issue without regressions?
- Verification discipline: did it run the right tests and report failures honestly?
- Recovery: after interruption, can it resume without redoing or forgetting work?
- Isolation: can parallel workers avoid colliding on files and state?
- Total completion cost: include retries, human review time, and failed runs—not only token price.
Model intelligence dominates short, clean tasks. Harness architecture dominates long, messy ones. The best July 2026 setup is therefore often a portfolio: a frontier model for decisions, cheaper workers for fan-out, and an orchestration boundary chosen for the task’s actual lifetime.
Bottom Line
Anthropic currently offers the clearest ladder from cheap worker to high-judgment model and the widest menu of native Claude coordination patterns. OpenAI pairs GPT-5.6 with the broadest set of Codex surfaces and configurable recursive subagents. Google has a strong stable coding baseline in Gemini 3.5 Flash, but consumer tool users must understand the Gemini CLI-to-Antigravity transition. xAI has made Grok Build unusually inspectable just as Grok 4.5 becomes its default coding model, though the CLI is still young.
Do not buy a benchmark winner in isolation. Pick a model role, a process boundary, and a recovery story. That is what determines whether an agent demo becomes a dependable engineering system.
Official references: Anthropic models, Claude Code subagents, Claude Code Agent Teams, OpenAI models, Codex non-interactive mode, Gemini models, Antigravity transition, Jules FAQ, Grok 4.5, and Grok Build open source.