AI coding agents in Q4 2026: Codex, Claude Code, Google Antigravity and Grok Build compared — what shipped Aug–Oct, how each one bills, and our trend calls

This is a research piece first: what OpenAI, Anthropic, Google and SpaceXSI actually shipped for coding agents from August to October 2026, read from their changelogs, docs and announcements on 2026-10-28, and what we think it means for the next 6–12 months. Cursor and GitHub Copilot appear as context only. Buying advice comes at the end. Vendor benchmark numbers are the vendors’ own.

The timeline: August to October 2026

The pace is the first finding. Claude Code published 79 dated releases in those twelve weeks (v2.1.221 → v2.1.312), the Codex CLI 35 (0.146.1 → 0.165.0) on top of weekly desktop-app builds, Google’s Antigravity app went from v2.6.0 to v2.28.0, and Grok Build reached v1.0.70 on October 23. Anything you read about features is a snapshot.

OpenAI Codex: one meter, five models, three speeds

Codex is now one half of a shared agent allowance: ChatGPT Work and Codex use the same pricing, credits and limits. Inside it, OpenAI built a model ladder in five weeks and priced speed separately.

• GPT-6.1 Terra (Oct 26) is the efficiency play. Its API price ($1 / $5 per million tokens) is half of GPT-6.1 Sol’s, the Plus estimate is 35–285 local messages per five hours against 15–160 for Sol, and OpenAI says Terra solved 50.3% of Terminal-Bench 4.0 at max effort, while recommending Sol or Astra for the hardest coding.
• Speed is a multiplier, not a feature. Fast uses included usage at 2.5x and purchased credits at 2x; Ultrafast (GPT-6 Astra and, since Oct 8, GPT-6.1 Sol) uses 8x and 6x, and only Pro $500 among self-serve plans has it.
• Pro has no five-hour limit. Plus works in five-hour windows; all three Pro tiers ($100 / $200 / $500) have none, but still have an included allowance and possibly weekly limits. When you run out mid-turn, the turn finishes; then Plus and Pro can buy credits instead of upgrading.
• Cloud and events. Codex Cloud (Sep 29) runs tasks in reusable environments that keep going while your computer sleeps (2 vCPU / 8 GiB on Plus, 4 vCPU / 16 GiB on Pro); GitLab joined GitHub as a source (beta, Aug 19); scheduled tasks can fire on Gmail, Slack and GitHub events (Aug 25). Using Codex with an API key gives you the CLI and SDK but no cloud features.
• Openness in both directions. Codex imports setups and recent work from Claude Code, Claude Cowork and Cursor (Aug 11), and the Oct 27 CLI release added a built-in Google Cloud Vertex AI provider.

More on the plan side in ChatGPT Pro 200 for coding and how far Codex credits go.

Claude Code: cloud sessions, self-hosted runners, mods

Claude Code is included in every paid Claude plan and shares its five-hour and weekly limits with Claude chat; it runs in the terminal, the desktop app, VS Code, Cursor and other forks, and JetBrains IDEs. The default model on Pro, Max, Team and Enterprise is Opus 5.5 (Sep 22, $4 / $20 per million tokens, 40% cheaper to run than Opus 5 per Anthropic).

• Cloud sessions run on Anthropic’s infrastructure on Pro, Max and Team (and Enterprise premium seats), keep going after you close the laptop, and count toward plan limits with no separate charge for the VM. Projects, in public beta on Pro and Max, coordinate a stream of tasks as parallel cloud sessions that share repositories, instructions and memory.
• Self-hosted runners (Aug 7, Team and Enterprise) let web, mobile and desktop sessions run on your own machines or containers; Remote Control (Pro, Max, Team, Enterprise; not with API keys) drives a session that stays on your computer.
• Extensibility. Mods (Oct 1) let plugins change Claude Code’s behaviour and interface; through October the hooks kept growing (prompt, agent-spawn, autocomplete). Auto mode, which decides permission prompts, got a guided setup on Oct 19. See our mods guide.
• Team-first review. Code Review went GA for Team and Enterprise on Oct 15. It has no flat price: reviews are billed on tokens through usage credits, separately from plan usage, to the organisation.
• Limits moved twice. On Sep 14 the weekly Claude Code limit changed in a way Anthropic called “a 17% reduction” versus the day before; on Sep 22 five-hour limits were raised (size not published). Enterprise now buys seats at $20 a month plus usage at API rates.

The wider Anthropic picture is in our August–October Anthropic review.

Google: Gemini CLI became Antigravity CLI

Google’s biggest coding change happened before this quarter and shapes it: on May 19 it announced that Gemini CLI would be folded into Antigravity, and from June 18 Gemini CLI and the Code Assist IDE extensions stopped serving Google AI Pro, Ultra and free individual accounts. Gemini CLI lives on for Code Assist Standard and Enterprise licences and paid API keys (v0.66.0 with nightlies through October 28), and its README now says it is not optimized for the Gemini 4 series. For everyone else the tool is Antigravity: a desktop app, a CLI written in Go, IDE extensions and an SDK sharing one agent harness.

• Gemini 4 Argon (Oct 13) is rolling out in the Antigravity app and CLI. Antigravity’s post and model list put it on Google AI Pro and Ultra (not on Free or AI Plus); Google’s launch-day post mentioned Ultra. API price: $2 / $10 per million tokens through November 30, $4 / $20 from December 1. Google says Argon in Antigravity raised its own engineers’ agentic code submissions by more than 35% and cut rollbacks by nearly a third. Antigravity’s coding interface also sits inside the Gemini macOS app.
• Quota model: every plan gets the CLI and scheduled tasks; Google AI Pro gets a quota refreshed every five hours until a weekly limit, Ultra the highest quota plus third-party models; beyond that, AI credits at Gemini Enterprise consumption pricing, with a Never / Always overage switch.
• The app moved fast: Remote Control (Aug 20), /plan (Sep 22), a plugin marketplace (Sep 28), Automations (Oct 7), a built-in browser panel (Oct 27). Business use goes through Google Cloud from $30 per seat a month or pay-as-you-go.
• Jules, Google’s asynchronous coding agent, last updated its public changelog on March 9, 2026; the visible momentum is in Antigravity.

We do not sell Google plans; the rest of Google’s quarter is in Google Gemini, October 2026.

SpaceXSI: Grok Build, and Cursor inside The X Subscription

Grok Build is SpaceXSI’s coding agent: a CLI powered by Grok 4.7 (Sep 21, $2 / $6 per million tokens) that reads AGENTS.md and supports plugins, hooks and MCP servers, plus a web and mobile app builder available on every plan since Aug 19. The bigger move is commercial. Since Oct 12, The X Subscription bundles Grok, Grok Bot, Cursor and X Premium: Plus $8, Premium $30 (4x Plus usage), Super $100 (15x) and Ultra $200 (50x), with one pool for Grok models across Grok and Cursor. Premium adds code hosting, AI code reviews and cloud agents; third-party models in Cursor are billed separately at API list prices. We do not sell these plans.

GitHub Copilot, for context, keeps playing aggregator: Claude Haiku 5.5 (Oct 7), Gemini 4 Argon (Oct 13) and GPT-6.1 Terra (Oct 26) all landed in Copilot within days of launch.

How each one bills

None of the four publishes how many tasks a plan buys. OpenAI gives ranges per model, Anthropic gives per-session multiples (Max 5x / 20x, Team Premium 6.25x Pro), Google and SpaceXSI give relative tiers. Real capacity depends on the model, context size, caching and how long the agent runs, so measure with your own work for a week before upgrading.

Our read: five trends for the next 6–12 months

These are our analysis, not vendor statements. Each lists the signals, a confidence word and what would prove it wrong.

1. One billing shape: pool, windows, overflow (likely)

Signals: Codex (five-hour windows on Plus, weekly limits, credits), Claude Code (five-hour and weekly limits, usage credits) and Antigravity (five-hour quota until a weekly limit, AI credits) now share a structure; SpaceXSI pools usage across two products. Our read: plans will differ by pool size and overflow price, not by feature lists. Wrong if: a major vendor goes back to fixed message counts or a flat “unlimited” coding plan.

2. Speed becomes a paid tier (likely)

Signals: Fast at 2.5x and Ultrafast at 8x of included usage, Ultrafast only on Pro $500 and extended to GPT-6.1 Sol on Oct 8; Anthropic sells a faster Opus 5.5 mode at a higher API price. Our read: others will price speed explicitly, and the top consumer tier will be defined by speed as much as by volume. Wrong if: Ultrafast reaches Pro $200 at no extra multiplier.

3. Cheaper workhorses carry most agent work (likely)

Signals: GPT-6.1 Terra at half of Sol’s API price with roughly twice the Plus estimate; Opus 5.5 40% cheaper to run than Opus 5 as Claude Code’s default; Argon at $2 / $10 until Nov 30 (then $4 / $20); Grok 4.7 at $2 / $6. Our read: the same plan will do more routine work, while the flagship models (Astra, Fable) stay metered heavily. Wrong if: vendors cut included allowances in step with each cheaper model.

4. Agents run without you, on the same pool (likely)

Signals: Codex Cloud and event-triggered tasks, Claude Code cloud sessions, Projects and self-hosted runners, Antigravity Remote Control and Automations, cloud agents in X Premium; Claude’s cloud sessions explicitly count toward plan limits. Our read: background work will spend a growing share of each plan, and review (already billed separately on Claude’s Code Review) is the first piece to get its own price. Wrong if: vendors start metering background agents outside plan limits as a standard feature.

5. Switching gets cheaper (possible)

Signals: Codex imports Claude Code and Cursor setups; AGENTS.md, skills, hooks, plugins and MCP work in Codex, Claude Code, Antigravity CLI and Grok Build; Anthropic gave Agent Skills to the Agentic AI Foundation on Oct 22; Codex CLI now talks to Vertex AI; Copilot adds every new model within days. Our read: lock-in moves from the tool to the account and its allowance, so monthly plans and quick switches make sense. Wrong if: a vendor ties a must-have model to its own tool only, or stops importing others’ configurations.

What developers should expect

• Read the usage page, not the plan card. Codex shows limits at chatgpt.com/codex/settings/usage and /status in the CLI; Claude shows them under Settings → Usage.
• Pick the model per task. Terra or Luna for routine edits, Sol or Opus 5.5 for real features, Astra or Fable for the hardest jobs; the price gap is several-fold.
• Turn on background agents deliberately. They save time and spend the same pool.
• Keep scripts model-agnostic. GPT-5.5 left Codex on Oct 14 and GPT-5.3-Codex-Spark on Sep 14; hard-coded model names break.

What to buy now

Comparisons: ChatGPT Pro 200 vs Claude Max 20x, Claude Max 6.25x vs ChatGPT Pro 200, and credits or upgrade when Codex runs out.

How ordering works: a top-up goes onto your own ChatGPT or claude.ai account; we check its sign-in data before you pay, and you pay nothing if the check fails. A ready-made account is hand-delivered in about 30 minutes to 4 hours after you verify a delivery email; one account per person, change the password on first sign-in. Codex credits are added to your own Plus or Pro. Payment is USDT on TRC20 or BEP20; to renew, message Telegram @ccnb1937 with your order ID. Ready-made accounts are not refundable once delivered; if anything goes wrong, contact us promptly.

Risk, briefly: OpenAI’s and Anthropic’s terms apply per user, and access from regions they do not support may be restricted. ClaudeMax is independent, not OpenAI, Anthropic, Google or SpaceXSI and not an authorized reseller; vendors set plans and limits and may change them again.

Also in this batch: Google Gemini, October 2026 and Meta Muse, October 2026; the 10-26 research trio: OpenAI, Anthropic and the Q4 industry overview.