If you pay for one AI assistant mainly to write and code, the choice in October 2026 is usually Claude or ChatGPT. The two now cost about the same at every level: $20 a month to start, and $100 or $200 for heavy use. The real differences are which models each plan gets, how large the context window is, how usage limits are shared and which coding agent you prefer. Everything below comes from the companies' own pricing pages and help centers, plus public leaderboards that we label as such. Disclosure: 9to5AI drafts stories with Claude's help, so we have kept to the published numbers and attributed vendor claims.
The short answer
- Mostly coding, in long sessions? Both work. Claude Pro includes Claude Code and gives a 1M-token context in chat. ChatGPT Plus includes Codex on web, CLI, IDE and iOS. If you hit limits, the $100 tiers (Claude Max 5x, ChatGPT Pro 100) are the real comparison.
- Mostly writing and editing long documents? Claude's larger context window, and its strong showing on crowd-voted writing leaderboards, give it the edge on paper.
- Want the cheapest paid tier? ChatGPT Go ($8) has no Claude equivalent.
- Want the most powerful model regardless of cost? ChatGPT Pro 500 ($500) offers the fastest GPT-6 Astra. Claude Max ($200) lets you spend up to half your weekly limits on Fable 5.1.
Plans and prices
Last verified: October 8, 2026 (US prices; full rows with sources in our data file)
| Price point | ChatGPT | Claude |
|---|
| $0 | Free: GPT-6 Luna in Chat; Work and Codex limited | Free: Sonnet and Haiku; no Claude Code, Research or Cowork |
| $8 | Go: Luna, limited Work and Codex; may show ads | — |
| $20 | Plus | Pro ($17/month billed annually) |
| $100 | Pro 100 | Max 5x |
| $200 | Pro 200 | Max 20x |
| $500 | Pro 500 (adds Astra "Ultrafast") | — |
| Team, standard seat | Business: $25 monthly / $20 annual, 2–200 users | Team Standard: $25 monthly / $20 annual, 2–150 seats |
| Team, premium seat | $125 monthly / $100 annual (5x usage) | $125 monthly / $100 annual (5x usage) |
| Enterprise | Contact sales | $20/seat/month (annual) plus usage at API rates |
ChatGPT's Pro tiers differ only in how much usage they include and whether Ultrafast is available. OpenAI cut the included allowance for new Pro 200 subscribers on September 29, 2026. Existing subscribers keep the old allowance until October 29. Since October 7, Claude's Max and Team plans also come with monthly API credits ($100 or $200 on Max). Those credits can't be used in the apps or Claude Code.
Models on each plan
ChatGPT now uses different models in different modes. Since October 7, 2026, Chat runs GPT-6 Sol on paid plans and GPT-6 Luna on Free and Go. Work and Codex run a wider set: GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna and the most powerful model, GPT-6 Astra. Plus subscribers get limited Astra use in Work and Codex plus optional credits. Pro 100 and 200 can spend their full allowance on Astra. On Pro, the "Pro" reasoning option in Chat is powered by Astra. Our ChatGPT agent mode explainer explains Chat vs Work.
Claude uses one model list across chat, Cowork and Claude Code. Paid plans get Opus 5.5 (released September 22, 2026), Sonnet 5.5 (September 28) and Haiku 5.5 (October 7). The more expensive Fable 5.1 is billed through usage credits on Pro and Team Standard. Max and premium seats can spend up to 50% of their weekly limits on it. Our guide on which Claude model to use explains the trade-offs, and the Claude app explainer covers the rest of the plan.
Context windows
| ChatGPT (published) | Claude (published) |
|---|
| Free | 27K (Instant); reasoning "varies" | Not documented by Anthropic |
| Go ($8) / Plus ($20) | 54K tokens (Instant), 256K (reasoning) | Pro ($20): 1M tokens for the 5 and 5.5 models in chat |
| Pro 100–500 / Max | 128K (Instant), 400K (reasoning) | 1M (same models) |
| Business / Team | 54K (Instant), 256K (reasoning) | 1M (same models) |
| Enterprise | 128K (Instant), 256K (reasoning) | 1M (same models) |
Last verified: October 9, 2026 (ChatGPT figures from the chatgpt.com/pricing comparison table). These are the figures each company publishes for its consumer and business apps, not its APIs. Claude's advantage matters most if you paste in whole codebases or book-length manuscripts. For everyday use, both are large enough. Anthropic says that when a chat nears the limit, Claude summarizes older messages if code execution is on.
Coding: Codex vs Claude Code
| ChatGPT: Codex | Claude: Claude Code |
|---|
| Included on | Every plan (Free and Go "limited") | Pro, Max, Team, Enterprise |
| Where it runs | Web, CLI, IDE extension, iOS, desktop app; cloud tasks on OpenAI computers | Terminal, IDE, desktop; Chrome integration for testing web apps |
| Limits | Same structure as Work. Plus / Business Standard estimates per 5 hours: Astra 5–45 and GPT-6.1 Sol 15–160 local messages. Pro has no five-hour limit for now | Shared with Claude chat and Desktop: five-hour session plus weekly limit |
| When you run out | Buy credits | Usage credits, or continue on API credits at API rates |
OpenAI's Codex figures are its own estimates, "not fixed limits." Cloud tasks can use more. Anthropic doesn't publish message counts for Claude Code, only that usage is shared with the apps. On a Claude plan, a long agentic coding session uses the same allowance as your writing. For a deeper comparison that includes Cursor, see Claude Code vs Codex vs Cursor and our Codex explainer.
What leaderboards say, and what they don't
The most traceable independent signal is Arena (formerly LMArena, now at arena.ai). Arena ranks models from crowd-sourced, anonymous head-to-head votes, reports confidence intervals and offers a "style control" adjustment.
Arena WebDev leaderboard, page dated October 8, 2026 (top five rows, as displayed by default; checked October 9, 2026)
| Rank | Model | Score | CI (±) | Votes |
|---|
| 1 | Claude Opus 5.5 (max effort) | 1813 | ±15 | 2,294 |
| 2 | GPT-6 Astra (max) | 1786 | ±11 | 5,189 |
| 3 | Claude Sonnet 5.5 (xhigh) | 1774 | ±15 | 2,323 |
| 4 | GPT-6.1 Sol (max) | 1755 | ±15 | 2,102 |
| 5 | Claude Fable 5.1 (max) | 1744 | ±9 | 6,663 |
The WebDev page doesn't say whether its style-control adjustment is applied to the displayed scores. The Text and Creative Writing pages show style control set to "None" by default.
Arena doesn't state the confidence level on the page, so treat the intervals as approximate. Read these results with care. WebDev measures which front-end result voters preferred in a short session. It doesn't measure multi-hour refactors, and it uses top effort settings that cost more than defaults. Opus 5.5 and Sonnet 5.5 also have relatively few votes so far. Anthropic's Opus 5.5 launch page puts Opus 5.5 at 66.4% on Terminal-Bench 4.0 (at xhigh effort, with production safeguards on) against 57.9% for GPT-6 Astra (at high effort, a figure Anthropic says is "as reported by OpenAI"). These are vendor-reported numbers, not independent results; Google's own launch table for Gemini 4 Argon lists a different Astra figure (58.2%), which shows how much these comparisons depend on who runs them and how.
Writing
Neither company publishes a writing benchmark for its apps, so the comparison rests on features and crowd preference.
- Arena Creative Writing (page dated October 8, 2026, style control off by default): Google's Gemini 4 Argon is first at 1519 (±20, 1,053 votes, marked preliminary). Claude Opus 5.5 is a statistical tie in second at 1517 (±17, 1,480 votes). The best-ranked GPT-6 model, GPT-6.1 Sol, is 26th. On the overall Text leaderboard (same date), Claude models hold 7 of the top 10 places, behind Gemini 4 Argon in first. Newer models have far fewer votes than older ones, so their positions can still move.
- Anthropic's claim: Opus 5.5 "puts key information first, uses less jargon, and follows user writing rules." That is a vendor claim.
- OpenAI's latest change: on October 7, 2026, Chat gained "Intelligent UI," which can answer with charts, buttons and small tools instead of plain text. That helps explanations more than prose writing.
- Shared basics: both offer projects with files and instructions, memory and file uploads. Claude's artifacts, including Docs and Slides formats, are on every plan. ChatGPT's Work mode produces finished documents, spreadsheets and presentations on paid plans.
We left out ChatGPT's canvas and Claude's styles because we could not confirm their current status from official pages on October 8, 2026.
Strengths and weaknesses
ChatGPT strengths: an $8 entry tier; Codex on every plan, including Free; a $500 tier for the fastest Astra; mature task automation in Work, with scheduled and webhook-triggered tasks.
ChatGPT weaknesses: different models in Chat and Work/Codex can confuse users; smaller published context windows; Pro 200's allowance was cut for new subscribers.
Claude strengths: a 1M-token context in chat on paid plans; one model list across every surface; strong crowd-vote results for coding and writing; Claude Code included from Pro.
Claude weaknesses: no Claude Code on Free; no plan between Free and $20; one shared allowance means coding can starve chat; Fable costs extra on Pro.
Recommendation by use case
| If you… | Choose | Why |
|---|
| Write long reports or edit book-length drafts | Claude Pro | 1M-token context in chat; strong writing-leaderboard results |
| Code a few hours a week | Either $20 plan | Both include their coding agent. Pick the one whose CLI you prefer |
| Code all day with agents | Claude Max 5x or ChatGPT Pro 100 | Both $100. ChatGPT Pro has no five-hour cap at present. Claude has a shared five-hour session limit plus a weekly limit |
| Want the cheapest paid plan | ChatGPT Go | $8, with no Claude equivalent |
| Want each company's top-tier model with the most headroom | ChatGPT Pro 500 or Claude Max 20x | Astra Ultrafast vs Fable 5.1 within limits |
| Buy for a small team | Either | Seat prices are identical. Compare admin features and existing tools |