Featured AI Coding Models Claude Code Codex Antigravity Grok Build Comparison

AI Coding Models & Agent Tools: The October 2026 State of Play

An October 21, 2026 snapshot of Claude, GPT, Gemini, and Grok coding models—and the Claude Code, Codex, Antigravity, and Grok Build agent stacks around them.

October 21, 2026 • 14 min read • By Claude World

The AI coding market still has no single meaningful leaderboard—and since our July 19 snapshot, nearly every model on that page has been superseded. Anthropic shipped Fable 5.1 and a 5.5 generation, OpenAI moved Codex to GPT-6, Google made Gemini 4 Argon generally available, and xAI released Grok 4.6 and 4.7. Every lab still ships both a model portfolio and an agent harness: the model supplies judgment, the harness supplies tools, permissions, parallel workers, persistence, and a place to run.

The two layers are also more loosely coupled—Antigravity now offers Claude Opus 5.5 and Sonnet 5.5 to eligible Google Cloud users—so you can mix them more freely, with more defaults to check when you do.

Snapshot date: October 21, 2026. Model aliases, preview labels, quotas, and product availability change quickly. This article marks preview, beta, and experimental surfaces explicitly and links to the vendors’ current documentation rather than treating this table as timeless. CLI versions checked: Claude Code 2.1.305, Codex CLI 0.163.0, and Grok Build 1.0.62.

The Short Version

ProviderCurrent model ladderMain coding-agent surfaceImportant status note
AnthropicFable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 (Mythos 5.1 via trusted access)Claude Code: CLI, IDE, desktop/web, cloud and self-hosted sessions, subagents, dynamic workflowsOpus 5.5 is the default on Pro, Max, Team, Enterprise, and the Anthropic API
OpenAIGPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna (GPT-5.6 is older)Codex in the ChatGPT desktop app, CLI, IDE, and cloud; codex exec; SDKsGPT-6.1 Sol is the Codex default; GPT-5.5 left ChatGPT-sign-in Codex on October 14
GoogleGemini 4 Argon, 3.8 Flash, 3.5 Flash-Lite (3.1 Pro is still Preview)Antigravity app and CLI; Gemini API managed agent (Preview); Jules (Public Beta)Argon GA since October 13 on the paid tier; 3.5 Flash is deprecated and auto-routed
xAIGrok 4.7, Grok 4.7 Fast, Grok 4.6, Grok 4.5Grok Build terminal agent and open-source harnessGrok 4.7 is the default; 4.7 Fast runs only in Grok Build and Cursor; no official CLI changelog

There is still no universal winner. The useful question is unchanged: which combination is best for this workload, under this failure model, at this volume?

Anthropic: A New Ladder and a More Autonomous Claude Code

Anthropic replaced its whole ladder between September 1 and October 7:

  • Claude Fable 5.1 (claude-fable-5-1, September 1) is the top tier, which Anthropic’s model guide recommends “for demanding reasoning and long-horizon agentic work.” It is never the default; select it with /model fable.
  • Claude Opus 5.5 (claude-opus-5-5, September 22) is the new center: “If you’re unsure which model to use, start with Claude Opus 5.5 for most workloads.”
  • Claude Sonnet 5.5 (claude-sonnet-5-5, September 28) is “a faster, lower-cost complement to Claude Opus 5.5.”
  • Claude Haiku 5.5 (claude-haiku-5-5, October 7) is the first Haiku-class model with adjustable effort, and it “pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work.”
  • Claude Mythos 5.1 is not a higher tier: “Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards.” Mythos 5.1 is available only through Anthropic’s trusted access programs.

Cheap fan-out, expensive judgment still holds, but the line moved: Opus 5.5 “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5,” at $4/$20 per million tokens against Fable 5.1’s $10/$50. Sonnet 4.5 retires from the Claude API on November 30.

Claude Code 2.1.305 is more opinionated too. On the Anthropic API, opus, sonnet, haiku, and fable resolve to the new models, and Opus 5.5 is the default on Pro, Max, Team, and Enterprise; Bedrock, Google Cloud’s Agent Platform, Claude Platform on AWS, and Microsoft Foundry still map some aliases to older models (haiku is Haiku 4.5 on all four). All four current models accept low through max, the 5.5 models defaulting to medium. Auto mode now starts interactive terminal and VS Code sessions that configure no mode, while claude -p and Agent SDK runs start in Manual. And ultracode is a setting rather than an effort level: /effort ultracode or "ultracode": true has Claude plan dynamic workflows—“Dozens to hundreds of agents per run”—at the session’s current effort.

The other boundaries stay distinct: subagents return one result to the caller, with forks on by default and nesting three layers deep, while Agent Teams remain “experimental and disabled by default.” claude self-hosted-runner runs cloud sessions on a Team or Enterprise organization’s own infrastructure, mods (October 1) are “small TypeScript functions that change how Claude Code works”—unsandboxed, so “only install mods from sources you trust”—and Code Review reached general availability on October 15.

OpenAI: GPT-6 Takes Over Codex

OpenAI replaced the GPT-5.6 trio with a GPT-6 family in under four weeks:

  • GPT-6 Astra (gpt-6-astra, September 3) is the frontier tier, “the world’s most intelligent and aligned model.” Codex’s model guide says: “Choose Astra when a task needs the strongest capability across steps and tools.”
  • GPT-6.1 Sol (gpt-6.1-sol, September 29) is the recommended default and the Codex CLI default: “an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices.”
  • GPT-6 Sol and GPT-6 Luna (September 22): Sol is now the “Previous generation workhorse model,” while Luna handles clear, repeatable tasks and supports effort up to Max, but not Ultra.
  • GPT-5.6 Sol, Terra, and Luna are the older generation.

For Codex users signed in with ChatGPT, GPT-5.4 and GPT-5.4 mini retired on August 31 and GPT-5.5 on October 14; the API and API-key sessions keep them. GPT-5.3-Codex-Spark was deprecated on September 14 and removed from the desktop app, CLI, and IDE extension. Codex itself now lives in the ChatGPT desktop app beside Chat and ChatGPT Work, and its documentation moved to learn.chatgpt.com, where old developers.openai.com/codex links redirect.

Codex CLI 0.163.0 (October 15) adds an agent command center (codex agents or /agents; /subagents switches between a session’s subagents, and /agent is gone), managed Git worktrees on by default, a background server, and mid-response steering. codex exec remains the Stable automation entry point, codex app-server is still Experimental, and the planning tool has been off by default since 0.152.0. Subagent settings now include a default subagent model and reasoning effort, and custom agents inherit the parent’s permissions, provider, and MCP servers without being able to override them. The effort ladder runs from Light (low in the CLI) to Max and Ultra, which “uses subagents to handle separate parts of a complex task in parallel” and has no API equivalent.

For security work, Daybreak gates cyber-permissive access through Trusted Access for Cyber: Daybreak Blue for most authorized defensive work, Daybreak Red with separate approval. Codex CLI 0.163.0 turns Daybreak on by default for eligible accounts, and OpenAI is “strongly encouraging Daybreak customers using Codex to switch from full-access mode to auto-review mode.”

Google: Gemini 4 Argon Is GA, and July’s Baseline Is Deprecated

Google’s model page has a new top and a new middle:

  • Gemini 4 Argon (gemini-4-argon) became “globally available” on October 13, after a September 30 announcement gave trusted cyber defenders first access. It is Stable in the API but “not available on the Free tier,” shares one 1,048,576-token window across input, thinking, and output, and offers low, medium (default), and high thinking levels. Introductory pricing of $2/$10 per million tokens ends November 30, rising to $4/$20.
  • Gemini 3.8 Flash (GA September 2) is the workhorse for software engineering and agentic tasks, and Gemini 3.5 Flash-Lite (GA July 21) the low-cost tier; for new projects, Google points to “3.5 Flash-Lite or 3.8 Flash.”
  • Gemini 3.1 Pro is still Preview, and Gemini 3.5 Pro has not shipped.

Gemini 3.5 Flash—July’s recommended baseline—and 3.7 Flash were deprecated on October 8. Their endpoints now route automatically to gemini-3.6-flash and gemini-3.8-flash. If you pinned gemini-3.5-flash, you are already running a different model: rerun your evaluations.

The tooling consolidated around Antigravity. The Antigravity app reached v2.24.0 (October 15) with per-model reasoning effort, Remote Control, /plan review, a plugin marketplace, subagent messaging, and Automations. The Antigravity CLI (agy) reached v1.4.0 (October 16) with --effort, no default timeout for headless -p runs, and an Auto mode that “approves file edits and safe commands without asking and still prompts before irreversible actions.” Both offer Claude Opus 5.5 and Sonnet 5.5 to eligible Google Cloud (ADC) users. Argon is rolling out in Antigravity to paid subscribers, though Google’s announcements differ on whether that includes Pro or only Ultra.

Older paths are narrowing. Gemini CLI is still maintained, but its README says it “has not been optimized for Gemini 4 Argon,” and new Gemini Code Assist Standard and Enterprise subscriptions ended on October 9. For API-driven agents, the managed Antigravity agent (Preview) “uses the same harness as the Antigravity IDE,” and Jules is still in Public Beta, with no official changelog entries since March 9.

xAI: Grok 4.7 Leads, and Grok Build Fills Out Its Lifecycle

  • Grok 4.6 (grok-4.6, August 12) “builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work.”
  • Grok 4.7 (grok-4.7, September 21) is xAI’s “most capable model for coding and knowledge work” and the Grok Build default. It launched in Grok Build, Cursor, and the xAI API at Grok 4.6’s price: $2/$0.50/$6 per million input/cached/output tokens below 200K prompt tokens, and $4/$1/$12 for the whole request at or above it. Both have a 500K-token context window.
  • Grok 4.7 Fast (grok-4.7-build-fast in the CLI): “Grok 4.7 Fast is the same model and costs 2x the standard token rates, or 1.5x for long-context requests. It’s available only in Cursor and Grok Build, not on the public xAI API, and is billed through your plan there.”

Reasoning effort adds xhigh to low, medium, and high (the default) on Grok 4.6 and later; on Grok 4.5, “requests with xhigh are treated as high.”

Grok Build 1.0.67’s help adds grok clone, grok worker (register a machine as a Cursor private or cloud-agent worker), grok du for state disk usage, grok usage <SESSION_ID> for token and cost reports, and a worktree lifecycle of create, detach, salvage, clean-artifacts, and redirect. That worker command is itself the argument: in 1.0.62 it was grok cursor-worker, and by 1.0.67 it had been renamed to grok worker and mirrored under grok leader worker, with no release note anywhere to say so. But xAI publishes no official per-version CLI changelog—no GitHub Releases, no tags—so the installed binary’s --help is the authoritative record, and the open-source mirror trails the build you run.

Model-Aware Planning as an Orchestration Mode

Model-aware planning is a routing contract, not a provider ranking. It decomposes one outcome into bounded task nodes, then chooses a model, harness, reasoning effort, and authority boundary for each. The provider roles below are starting hypotheses, not permanent winner labels—this quarter superseded nearly every model the July edition recommended.

  1. Write the task contract before choosing a model. Record the objective, acceptance criteria, required artifacts, dependencies, and side effects. Split independent discovery from dependent implementation, and keep irreversible remote actions as explicit approval gates rather than ordinary child tasks.
  2. Route by demonstrated working strength. Claude is a useful starting candidate for hierarchical decomposition, long-context synthesis, and script-held orchestration; Codex for repository-centered implementation, diff/test loops, and headless codex exec runs; Grok Build for terminal-first exploration and parallel worktree runs. These are roles to test, not a global scorecard; any provider can own a node when evaluation evidence supports it.
  3. Choose and pin model and reasoning effort per node. Spend high-capability models and higher effort on architecture, ambiguous debugging, security review, adversarial verification, and final synthesis; use faster models and lower effort for searches, inventory, formatting, and mechanical checks. Per-node controls now exist—Claude Code’s Agent tool takes effort, and Codex has default subagent model and effort settings—but the knobs differ: Haiku 4.5, still the haiku alias outside the Anthropic API, is not effort-capable; Codex’s Ultra has no API equivalent; Grok 4.5 treats xhigh as high. As OpenAI’s docs put it, “Reasoning efforts don’t map exactly between model generations.” Record the actual values, and never translate one provider’s knob into another by assumption.
  4. Freeze authority before delegation. Defaults moved toward autonomy: Claude Code’s interactive sessions start in auto mode, and the Antigravity CLI added an Auto mode. Each node therefore states whether it is read-only or may edit, which paths and tools are allowed, the sandbox and network boundary, and whether credentials are in scope. A handoff cannot widen the caller’s authority—Codex custom agents, for example, cannot override the parent’s permissions—and native permission and sandbox controls enforce that boundary where a prompt saying “be careful” does not.
  5. Make handoffs verifiable. Send the receiving agent a bounded packet: goal, relevant files, constraints, expected output, current commit or diff, tests already run, unresolved risks, and the evidence needed to claim completion. Claude Code now reads AGENTS.md when a project has no CLAUDE.md, so shared instructions travel further, but the receiver still rechecks repository state instead of trusting a transcript or a vendor session ID.
  6. Return evidence to one controller. One orchestrator owns dependency order and the final verdict: it reconciles concurrent changes, runs the affected tests, inspects the final diff, and reports blockers honestly. A confident summary is not completion evidence; the acceptance criteria and resulting artifact are.

Use this mode when task nodes genuinely need different judgment, cost, or execution boundaries. For a small edit with one obvious verification step, a single well-scoped agent remains the clearer plan.

What Changed Since July

AreaJuly 19October 21
Anthropic modelsFable 5, Opus 5, Sonnet 5, Opus 4.8, Haiku 4.5Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5; the July models are legacy
Claude Code2.1.226; Haiku 4.5 without effort; nesting described as fixed at five levels2.1.305; Opus 5.5 default; 5.5 models default to medium; auto mode starts interactive sessions; ultracode is a setting; nesting defaults to three layers
OpenAI modelsGPT-5.6 Sol, Terra, LunaGPT-6 Astra, GPT-6.1 Sol (default), GPT-6 Sol, GPT-6 Luna
CodexDesktop, CLI, IDE, cloud, exec, SDKsCLI 0.163.0 with command center, managed worktrees, and background server; GPT-5.4, 5.3-Codex-Spark, and 5.5 retired from ChatGPT-sign-in Codex
Google modelsGemini 3.5 Flash (Stable), 3.1 Pro Preview, 3.1 Flash-LiteGemini 4 Argon (GA, paid tier), 3.8 Flash, 3.5 Flash-Lite; 3.5 Flash auto-routed
Google toolsAntigravity CLI replacing consumer Gemini CLI; Jules in Public BetaAntigravity app v2.24.0 and CLI v1.4.0, both offering Claude models; new Code Assist subscriptions closed; Jules still in Public Beta
xAIGrok 4.5 default; Grok Build in early beta, source opened July 15Grok 4.7 default, 4.7 Fast, 4.6; xhigh on 4.6+; Grok Build 1.0.62
WorkloadStrong starting pointWhy
Architecture, ambiguous debugging, adversarial verificationFable 5.1, GPT-6 Astra, or Gemini 4 ArgonSpend the frontier tier where wrong decisions are expensive
Everyday implementation with human reviewOpus 5.5 or Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash, or Grok 4.7Each vendor’s current default or workhorse
High-volume scans and mechanical fan-outHaiku 5.5, GPT-6 Luna, Gemini 3.5 Flash-LiteSupporting work rarely needs the most expensive judgment tier
Large Claude-native orchestrationClaude Code dynamic workflows or ultracodeThe plan lives in a resumable script, not one context window
One stack across desktop, CLI, IDE, cloud, and embedded automationCodexChatGPT desktop app, Stable codex exec, command center, managed worktrees
Google-centric asynchronous multi-agent workAntigravity app and CLIAutomations, subagent messaging, plugins, and Argon access
GitHub tasks that should run remotely and return a PRJulesAsynchronous cloud delivery; still Public Beta
Terminal-first parallel work with an inspectable harnessGrok BuildHeadless mode, worktree lifecycle commands, open-source agent loop

Then test the shortlist as the July edition recommended: run the same real task through each stack and record patch correctness, verification discipline, recovery, isolation, and total completion cost—every quarter, because defaults, aliases, and auto-routing now change underneath existing configurations.

Bottom Line

Anthropic has the freshest complete ladder, and Claude Code 2.1.305 ships more autonomy by default, so permission configuration now deserves as much attention as model choice. OpenAI moved Codex to GPT-6 with a near-Astra default in GPT-6.1 Sol. Google has a GA frontier model in Gemini 4 Argon and one tooling center in Antigravity, but auto-routed deprecations may already have changed July integrations. xAI has a stronger default in Grok 4.7 and a much larger Grok Build surface, yet without an official CLI changelog you verify the release yourself.

Do not buy a benchmark winner in isolation. Pick a model role, a process boundary, and a recovery story—and recheck all three every quarter.

Sources