ChatGPT vs Claude: Which AI Is Better in 2026? Models, Pricing, Strengths & Weaknesses Compared

Last updated: 23 September 2026. Both platforms shipped a new flagship this month, so any comparison written before September is describing models that are no longer current.

This is a comparison of what you can actually buy today. OpenAI released GPT-6 Astra on 3 September 2026. Anthropic released Claude Opus 5.5 on 22 September 2026, one day before this article was published. Those two models, plus the cheaper tiers underneath them, are what a real decision comes down to.

There is no single winner here, and any article that gives you one is selling something. The two platforms have diverged: they publish different benchmarks, price different parts of the stack aggressively, and have made opposite bets about what an AI product is for. What follows is where each one has documented advantages, where each one has documented problems, and how to pick per task rather than per brand.

A note on sourcing before we start. Every price and score below is attributed, and we flag where the two vendors' numbers are not directly comparable. We also flag the figures we could not verify. That matters more than usual in this comparison, because one of the headline benchmark results of the month turns out to measure infrastructure rather than a model.

ChatGPT vs Claude: Quick Comparison

The most useful thing in that table is the pricing row. Anthropic's current flagship costs 40% of OpenAI's current flagship per token. That single fact reframes most of the rest of the comparison, and we return to it under cost-effectiveness.

What Is ChatGPT?

ChatGPT is OpenAI's consumer and business product; the models underneath it are sold separately through the OpenAI API. Confusing the two is the most common mistake in comparisons like this one, because several things people call "ChatGPT features" are product features that have nothing to do with the model.

The current model lineup has two generations in active service:

• GPT-6 Astra (3 September 2026) is the flagship, built for long-horizon agentic work: computer use, coding, research and multi-step tasks. It has a 1,050,000 token context window and 128,000 token maximum output. It became the Codex CLI default on 4 September.
• The GPT-5.6 series (9 July 2026) covers the rest of the range in three tiers: Sol (the most capable of the three), Terra (the mid tier) and Luna (the cheapest). OpenAI positioned Terra as competitive with the previous generation at roughly half the price.

In the ChatGPT product, you do not pick these by name. Since March 2026 the model picker shows three labels, "Instant", "Thinking" and "Pro", and selects the underlying model automatically. This is a genuine usability improvement and a genuine transparency loss, and we treat it as both below.

Product capabilities that belong to ChatGPT rather than to any model: native image generation, voice mode, memory across conversations, Projects, file and document analysis, web browsing, agent mode, and the connector ecosystem. Codex, OpenAI's coding agent, has been available on every ChatGPT plan including Free and Go since June 2026, which is a meaningful distribution advantage.

The ecosystem has also been volatile. The Atlas browser, launched in October 2025, was discontinued on 9 August 2026 and its agentic browsing folded back into ChatGPT and Codex. The standalone Sora web and app experiences were discontinued on 26 April 2026. OpenAI's stated direction is consolidation into a desktop application plus a Chrome extension. If you are building a business process on a specific OpenAI surface, that history is worth weighing.

What Is Claude?

Claude is Anthropic's assistant, and unlike ChatGPT the consumer product and the API expose much the same model names. The family currently runs:

• Claude Opus 5.5 (22 September 2026), the newest model, at $4 / $20 per million tokens with a 1M token context window. Anthropic says it performs at a level similar to Claude Fable 5.1 on most work, costs about 40% less to run than Opus 5, and produces output more than 30% faster.
• Claude Fable 5.1 (1 September 2026), the most capable widely released model, at $10 / $50, for the most demanding reasoning and long-horizon agentic work.
• Claude Sonnet 5 at $2 / $10, the balanced workhorse, and Claude Haiku 4.5 at $1 / $5 for high-volume work. Haiku is the one model in the family with a 200K rather than 1M context window. Anthropic has signalled that Sonnet 5.5 and Haiku 5.5 will follow within weeks.

Claude's product surface is narrower than ChatGPT's and more work-shaped. Every paid tier includes Claude Code (the terminal and IDE coding agent), Claude Cowork (an agent that reads your files, runs scheduled tasks and returns finished work), Claude Design (visual and prototype creation, released April 2026) and Claude Science (research workflows, announced June 2026), all drawing on one shared usage pool. In September 2026 Anthropic merged the Claude chat and Cowork front ends into a single interface.

The important structural point for developers: MCP, the Model Context Protocol, is Anthropic's. Anthropic published it in November 2024, OpenAI adopted it in March 2025, and Anthropic donated it to the Agentic AI Foundation under the Linux Foundation in December 2025. It is now the de facto way to connect models to tools and data across vendors. This is a case where Anthropic's advantage has become everyone's advantage, so it is no longer a reason to pick Claude, but it does mean tool integrations you build are portable.

Model-by-Model Comparison

Below are the current models on both sides with published figures. Prices are standard first-party list rates per million tokens, checked 23 September 2026. Cached input bills at roughly 10% of standard rates on both platforms, and both offer a batch tier at about half price.

Two pricing caveats that matter and that most comparisons get wrong:

GPT-5.6 prices moved. The series launched on 9 July 2026 at higher rates and was cut on 30 July. Terra launched around $2.50 / $15 and is now $2 / $12; Luna launched around $1 / $6 and is now $0.20 / $1.20. If you see the launch numbers quoted as current, the source is stale.

GPT-5.6 Sol has two prices. Sources disagree, and both are right. The standard list rate is $5 / $30. OpenAI also lists a promotional $4 / $20 rate available at least through 21 November 2026. Any total cost of ownership built on the promotional rate has an expiry date attached.

We chart output price rather than input because output is what dominates a real bill on reasoning models: thinking tokens are billed as output, and both flagships think by default.

Where ChatGPT Excels

ChatGPT's strongest claim is breadth. One $20 subscription covers text, native image generation, voice, document analysis, browsing, agent mode and a coding agent. Nothing in Anthropic's lineup matches that span, and for a solo operator or a small marketing team the value of not assembling a toolchain is real.

Published reasoning and maths scores are the highest on offer. OpenAI reports GPQA Diamond at 96.0% and FrontierMath Tier 4 at 97.6% for Astra. Anthropic does not publish a comparable GPQA figure for Opus 5.5, so this is an area where OpenAI is ahead on the record even if the underlying gap is unknown.

Image generation is a genuine capability gap, not a preference. Claude cannot produce raster images at all. If your work involves generating visual assets, this is not a close call, and it is the single clearest functional difference between the two platforms.

Computer use and speed on agentic tasks. OpenAI reports Astra at 72.6% on OSWorld 2.0 against 65.7% for Sol, and completing those tasks in roughly 40 minutes on average versus about 75 minutes for the predecessor. Speed on long agentic runs compounds: a task that takes half as long costs less and fails less often from timeouts.

Distribution. Codex on the Free and Go tiers means a developer can evaluate OpenAI's coding agent at no cost. Claude Code requires a paid plan.

Where ChatGPT Has Limitations

You often cannot tell which model answered you. The simplified "Instant / Thinking / Pro" picker routes automatically, and confirming which model handled a specific query requires a Configure setting most users never open. For casual use this is fine. For regulated work, for benchmarking your own prompts, or for debugging a quality regression, not knowing what ran is a real operational problem.

Rate limit documentation is uneven. OpenAI documents per-model allowances on the Pro page, while the Plus page describes only that subscriptions "may include usage limits such as message caps". On Plus it is genuinely difficult to know which limit you have hit.

Product churn. Atlas shipped in October 2025 and was gone by August 2026. Standalone Sora was discontinued in April 2026. The capabilities mostly survived inside other surfaces, but teams that built around the standalone products had to migrate.

Citation reliability when not grounded. A Columbia Journalism Review citation audit found ChatGPT hallucinating citations at 67% with web search disabled, against 37% for Perplexity. The practical lesson is narrower than "ChatGPT hallucinates": it is that you should not ask any model for sources without search enabled.

Price at the top of the range. Astra at $10 / $50 is 2.5 times Opus 5.5 per token. That is defensible if Astra is the only model that can finish your task, and expensive if it is not.

Where Claude Excels

Claude leads the coding benchmarks both vendors actually published for their current flagships. On Terminal-Bench 4.0, Anthropic reports Opus 5.5 at 66.4% against OpenAI's reported 57.9% for Astra. On FrontierCode v1.1 it is 54.4% against 53.3%. Both sets are vendor-reported on each vendor's own harness, so read the FrontierCode gap of one point as a tie and the Terminal-Bench gap as meaningful but not decisive.

Price-to-capability at the frontier. Opus 5.5 arrived at $4 / $20 with Anthropic claiming near-parity with Fable 5.1 on most work, and cache reads dropped to $0.20 per million tokens, a 60% cut. For a workload that re-reads a large stable context on every call, which describes most coding agents and most document workflows, cheap cache reads matter more than the headline token price.

Long-context and document work. A 1M token window across the whole family except Haiku, up to 600 images or PDF pages per request, and citation-linked document grounding. On the independent vals.ai leaderboard, Claude Opus 5 leads SWE-bench Verified at 97.0%.

Instruction-following over long outputs. This is the most consistent theme in user discussion, and it is qualitative rather than benchmarked: Claude is described as holding a constraint set through a long response, where a constraint given at the start is less likely to drift by the end.

One coherent work surface. Claude Code, Cowork, Design and Science on a shared usage pool, with a single merged interface as of September 2026. Fewer products, less churn.

Where Claude Has Limitations

Usage limits are Claude's most-criticised feature, and the criticism is documented rather than anecdotal. Claude Code enforces a 5-hour rolling window plus a weekly cap on active compute hours, and hitting either throttles access. The specific complaint is that the Max plan's advertised "5x" and "20x" multipliers apply to the 5-hour sessions rather than to the weekly cap, so effective throughput is lower than the naming implies. The Register covered developer complaints in January 2026; a class action, Karl Kahn v. Anthropic, was filed in California federal court on 15 June 2026 arguing the advertised Max value was never real. Anthropic doubled the 5-hour limits across all paid plans on 6 May 2026. Treat the litigation as an allegation, not a finding, but treat the underlying pattern as well evidenced.

No image generation. Repeating it here because it is a hard boundary. Claude produces SVG, diagrams, charts and interactive artifacts, which covers more ground than people expect, but it will not give you a photograph or an illustration.

A narrower consumer product. No voice mode comparable to ChatGPT's, a smaller connector set, and less third-party tooling built specifically for it.

Fewer published academic benchmarks. Anthropic's launch materials favour task-completion benchmarks over GPQA-style academic tests. That is arguably the more honest choice, but it makes Claude harder to compare and means "Astra scores higher on GPQA" is unanswerable rather than refuted.

Haiku's 200K context. The cheapest tier does not share the 1M window, which is easy to miss when designing a cost cascade.

What Developers and Users Say About ChatGPT vs Claude

Across recent developer discussion, the most consistent theme is that experienced users stopped choosing and now run both. The common pattern described is Claude for code and long-form writing, ChatGPT for images, voice and general breadth, with a paid subscription to each.

Recurring positive themes for Claude centre on code quality, codebase comprehension and instruction adherence, particularly the observation that Claude holds a stated constraint through a long response. Recurring positive themes for ChatGPT centre on versatility, image generation and the size of the integration ecosystem.

The recurring criticism of Claude is usage limits, by a wide margin. The recurring criticisms of ChatGPT are inconsistency between sessions and uncertainty about which model handled a request.

On the coding-agent question specifically, a shorthand has settled in developer discussion: Claude Code for architecture and review, Codex for volume. That the two companies have made this interoperable is telling. OpenAI shipped a Codex plugin that runs inside Claude Code and added direct import of Claude Code configurations in August 2026.

One caution about the numbers you will see quoted. A figure of "70% of developers prefer Claude for coding" circulates widely, attributed to an unnamed survey. We could not trace it to a named publisher, sample size or methodology, so we are not repeating it as a finding. Several comparison articles also cite search-volume growth for "claude code" as evidence of quality; search volume measures interest, not satisfaction. Community sentiment is genuinely useful for finding out what to test, and genuinely unreliable as evidence of which is better.

ChatGPT vs Claude: Benchmark Comparison

The honest headline is that a clean head-to-head does not exist, because the two vendors no longer publish the same benchmarks. OpenAI has stopped reporting SWE-bench Verified, the most widely recognised coding benchmark. Anthropic does not report GPQA Diamond for Opus 5.5. Across the current flagships we found exactly three benchmarks where both vendors published a number.

Opus 5.5 leads two of the three; Astra leads Terminal-Bench-Science 0.1 at 64.6% against 58.7%. Each vendor ran its own harness and chose its own effort setting, so these are not controlled comparisons.

Where only one vendor reports, the numbers are still informative but not comparative:

The benchmark result you should be most sceptical of

GPT-6 Astra's ARC-AGI-3 score is the clearest example this year of a benchmark number that measures infrastructure rather than a model. The headline figure is about 99.9%. That run used a "provider adapter" harness that preserves opaque reasoning state between calls, and is reported at roughly $19,000 in compute. Run the same model through a standard stateless API loop and scores fall to a range of roughly 17% to 63%, depending on reasoning tier.

One analysis attributed 37 points of the score to the harness rather than the weights. Reporting also disagrees on the headline itself: some sources cite 98.6% rather than 99.9%, which appears to reflect different evaluation sets. We cannot resolve that from the published material.

The reason this matters to a buyer is simple. The score you can reproduce is the one from the harness you actually have. If you call the API in a normal loop, the stateless number is your number. This is not a criticism unique to OpenAI, and it generalises: a benchmark result is a claim about a model plus a harness plus an effort setting plus a cost, and vendors report the combination that flatters them.

Benchmark performance and real-world experience come apart routinely. By 2026 the frontier models sit within a few points of each other on most coding tests, which means workflow fit, context handling, latency and rate limits decide outcomes far more often than a benchmark delta does.

ChatGPT vs Claude Pricing

All figures below checked 23 September 2026 against published pricing.

Consumer and business plans

The consumer tiers have converged almost exactly, which tells you both companies have found the same price points. Two real differences: ChatGPT has a cheaper $8 entry tier, and Claude's team plans require a five-seat minimum against ChatGPT's two. Every paid Claude tier bundles Claude Code, Cowork, Design and Science on a shared usage pool, so the $20 comparison is not like for like.

API pricing

On both platforms, cached input bills at roughly 10% of standard rates and the batch tier at about half. OpenAI applies a 10% uplift for regional data-residency processing on models released on or after 5 March 2026.

Which AI Should You Use for Different Tasks?

One category deserves separating from the rest. "Which AI should I use" and "which AI cites my business" are different questions. Your customers are asking ChatGPT, Claude, Gemini and Perplexity about your category right now, and which assistant you subscribe to has no bearing on whether you appear in those answers. That depends on how your entity, schema and citations are structured, which is what AI visibility tracking and the difference between SEO, AEO and GEO are about. If you want to know where you currently stand, our free AI visibility scan queries the live answer engines and reports what they actually say.

Which Is More Cost-Effective?

The cheapest model per token is frequently not the cheapest way to finish the job. Cost per completed task is the number that matters, and it includes retries, review time, and the tooling you build to compensate for weaker output.

Casual users. Both free tiers are usable. If you want one paid product, ChatGPT's $8 Go tier is the cheapest real entry point and the breadth suits general use. Claude Pro at $17 to $20 makes more sense if your work is writing and code.

Heavy individual users. Near-identical pricing at $100 and $200. The deciding factor is not price but throughput, and this is where Claude's usage-limit criticism is most relevant: verify the weekly cap against your actual pattern before committing, because the Max multipliers apply to session windows rather than the weekly ceiling.

Developers. Claude is the cheaper build at the frontier today: $4 / $20 against $10 / $50, plus $0.20 cache reads. For an agent that re-reads a large stable context on every turn, the cache-read rate can dominate the bill, and that is a 60% cut. Against that, Codex is free to evaluate on any ChatGPT plan.

Startups. Test both, but model the cost cascade rather than the flagship. GPT-5.6 Luna at $0.20 / $1.20 is the cheapest capable tier either vendor offers and has no equivalent in Anthropic's lineup. Claude Haiku 4.5 at $1 / $5 is five times the price, with a 200K context. If most of your volume is classification or extraction, OpenAI's low end is materially cheaper.

Enterprise teams. Seat pricing has converged, so decide on data handling, admin controls and procurement rather than sticker price. Note Claude's five-seat minimum.

High-volume API applications. Batch at roughly half price and cached input at roughly 10% exist on both platforms and are the largest levers available. Measure before choosing a model: one well-cached call to a capable model routinely beats three uncached calls to a cheap one, and lower effort settings on a newer model often match higher effort on an older one at a fraction of the spend.

A pattern worth internalising: both vendors cut prices on their own frontier within weeks this year. Terra and Luna were cut on 30 July. Opus 5.5 arrived 20% below Opus 5 on 22 September. Any architecture that would be expensive to move between vendors is a bet that today's price ranking holds, and this year it has not held for more than a quarter at a time. Keep the model name in configuration, not in code.

Frequently Asked Questions

Is ChatGPT better than Claude? Neither is better overall. On current published evidence, Claude Opus 5.5 leads the coding benchmarks both vendors reported and costs 40% as much per token as GPT-6 Astra; ChatGPT leads on published academic reasoning scores, image generation, voice and ecosystem breadth. Most experienced users run both and pick per task.

Is Claude better than ChatGPT for coding? On the benchmarks both vendors published for their current flagships, yes, narrowly. Anthropic reports Claude Opus 5.5 at 66.4% on Terminal-Bench 4.0 against OpenAI's 57.9% for GPT-6 Astra, and 54.4% against 53.3% on FrontierCode v1.1. Both figures are vendor-reported on each vendor's own harness, and developer discussion generally favours Claude Code for architecture and review with Codex for high-volume work.

Which is better for writing, ChatGPT or Claude? There is no benchmark that settles this. The recurring theme in user discussion is a preference for Claude's prose and its tendency to hold a stated constraint through a long response. ChatGPT offers a wider stylistic range and generates images alongside the copy. Run the same brief through both.

Which is cheaper, ChatGPT or Claude? At the flagship level Claude is cheaper: Opus 5.5 at $4 / $20 per million tokens against GPT-6 Astra at $10 / $50. At the cheapest tier OpenAI is cheaper: GPT-5.6 Luna at $0.20 / $1.20 against Claude Haiku 4.5 at $1 / $5. Consumer subscriptions are effectively identical apart from ChatGPT's $8 Go tier.

Which has the larger context window? They are effectively tied. GPT-6 Astra offers 1,050,000 tokens; Claude Opus 5.5, Fable 5.1 and Sonnet 5 offer 1,000,000. Both cap output at 128,000 tokens. The exception is Claude Haiku 4.5 at 200,000.

Which AI is better for business? It depends on whether visual output is in scope. If you need images, voice or the widest connector ecosystem, ChatGPT. If your work is documents, code, analysis and policy-constrained responses, Claude. Seat pricing is nearly identical, so evaluate data handling and admin controls rather than cost.

Which AI is better for developers? Claude is the cheaper frontier build today and leads the published coding comparisons, and MCP originated with Anthropic. OpenAI's advantages are a free path to evaluate Codex on any plan, a cheaper low-end model, and a larger third-party ecosystem. Because both support MCP, tool integrations you build are largely portable.

Can ChatGPT and Claude use tools? Yes, both. ChatGPT has agent mode, web browsing, code execution and a large connector set; Claude has Claude Code, Cowork, web search and web fetch. Both support the Model Context Protocol, which Anthropic created in November 2024, OpenAI adopted in March 2025, and which is now governed by the Agentic AI Foundation under the Linux Foundation.

What is the difference between ChatGPT and Claude models? ChatGPT is OpenAI's product; its models are GPT-6 Astra and the GPT-5.6 series (Sol, Terra, Luna). Claude is both Anthropic's product and its model family: Fable 5.1, Opus 5.5, Sonnet 5 and Haiku 4.5. One practical difference is that in ChatGPT you usually cannot choose the model by name, because the picker shows "Instant", "Thinking" and "Pro" and routes automatically, whereas Claude exposes model names directly.

Should a business use ChatGPT or Claude? Start by listing your tasks, then check the one hard boundary: Claude cannot generate images. If image generation is required anywhere in your workflow, you need ChatGPT or a dedicated image tool regardless of everything else. If not, the decision comes down to cost per completed task on your own work, which means running a real week of it through both.

Is Claude more accurate than ChatGPT? There is no current, controlled, like-for-like accuracy comparison of GPT-6 Astra and Claude Opus 5.5. Older-generation comparisons showed large hallucination differences, and a Columbia Journalism Review audit found ChatGPT hallucinating citations at 67% with web search disabled. The reliable general finding is that any model is much more accurate with search grounding enabled than without, on either platform.

What are the newest ChatGPT and Claude models? As of 23 September 2026: OpenAI's newest is GPT-6 Astra, released 3 September 2026. Anthropic's newest is Claude Opus 5.5, released 22 September 2026, with Claude Fable 5.1 from 1 September 2026 positioned above it for the hardest work. Anthropic has signalled Sonnet 5.5 and Haiku 5.5 within weeks.

How to actually decide

Pick the hard constraints first, because they eliminate most of the debate. Image generation rules Claude out of a workflow entirely. A five-seat minimum rules out Claude Team for a pair of founders. A weekly compute ceiling rules out Claude Max for some heavy users. Regulated work that requires knowing which model produced an output argues against ChatGPT's automatic routing.

What is left after the constraints is a cost-per-completed-task question, and that one you have to measure on your own work. Run a real week through both. Benchmarks will not tell you, and neither will this article.

Both vendors shipped a flagship in the last three weeks and both cut prices on their own frontier this quarter. Whatever you choose, keep the model name in configuration rather than in code, and expect to revisit the decision within a quarter.

Further reading from us on the models in this comparison: GPT-6 Astra and what agent models change for business, Claude Fable 5 and workflow replacement, and the Claude Opus prompting playbook. For the wider picture on how discovery is shifting, see The State of AI Search 2026.

Sources and method

Prices, model names, release dates and context windows were checked on 23 September 2026. Benchmark figures are vendor-reported unless attributed otherwise, and we have said so at each chart. Where the two vendors' figures are not directly comparable, or where sources disagree, we have described the disagreement rather than picking a number.

One limitation worth stating plainly: the research for this article was conducted through web search and secondary reporting, because OpenAI's and Anthropic's own documentation domains were unreachable from our research environment. Every figure here is corroborated across multiple independent sources, but we have not read the vendors' pricing pages directly. Check the live pricing pages before making a commercial commitment, and tell us if you find a discrepancy.