Claude vs ChatGPT in August 2026

An even-handed comparison of Claude and ChatGPT as of August 2026 - current models, plan pricing, verified benchmark results, and the categories where each one clearly loses.

Claude and ChatGPT are close on raw intelligence and far apart on scope. As of August 2026, Anthropic's Opus 5 tops the independent Artificial Analysis Intelligence Index and Fable 5 tops arena.ai's human-preference board, while OpenAI's GPT-5.6 family wins most agentic-breadth benchmarks, generates images and video, and gives away unlimited text chat for free. Which one you should pay for depends almost entirely on whether you need generative media.

This page reports the losses as carefully as the wins. Every benchmark below names the model version and who ran the test, because in August 2026 almost every published table is run by one of the two vendors against the other's previous generation.

#What models does each company sell right now?

Anthropic ships four callable models; OpenAI ships three general-purpose ones plus a wide media and voice range. The tiers do not line up neatly, which is why vendor comparison tables so often talk past each other.

One structural note that matters more than the tier names: all three GPT-5.6 models carry the same 1.05M-token window, so OpenAI has flattened context across its range. Anthropic has done nearly the same - the 1M window is the default on Fable 5, Opus 5 and Sonnet 5 with no beta header and no long-context surcharge - but Haiku 4.5 is still capped at 200k. The Claude model index has the full specification table.

#Which model wins on benchmarks?

Neither, cleanly. Claude wins novel-problem reasoning and human preference; OpenAI wins agentic breadth and computer use. Here is the honest split, with the source for every row.

#Read those numbers with three caveats

Scaffolding is part of the score. Terminal-Bench measures an agent plus a model. "Claude Code + Fable 5" and "Codex + GPT-5.6 Sol" are different harnesses, so a two-point difference tells you as much about the tooling as the weights. A third-party roundup published on 2 August 2026 put Sol at xhigh effort on 89.5% and Opus 5 at max on 89.1% - a gap inside the error bars, but not a Claude lead.

Vendors benchmark against last generation. OpenAI's GPT-5.6 launch table compares Sol against Claude Opus 4.7-era models and Opus 4.8 on most rows, using Fable 5 only selectively. The 54.8% OSWorld figure above is Opus 4.8, not Opus 5, because Anthropic has never published a numeric Opus 5 OSWorld result - only a relative claim that it beats Fable 5 at a third of the cost.

Anthropic's own transparency is the weaker of the two. OpenAI and Google publish full numeric comparison tables as text. Anthropic published the Fable 5 benchmark table only as an image, and the Opus 5 announcement gives relative claims ("3× the next-best model on ARC-AGI-3", "within 0.5% of Fable 5 at half the cost") instead of absolute figures. That makes Anthropic's claims harder to check independently than its competitors', and it is a fair thing to hold against them. The point is expanded in our Claude review.

Benchmarks that used to settle these arguments - GPQA Diamond, AIME, MMMU and SWE-bench Verified - have effectively been retired at the frontier. None of them appear on OpenAI's GPT-5.6 tables or Google's current model cards. If you see a 2026 comparison leaning on them, it is recycling stale numbers.

#Where does ChatGPT clearly beat Claude?

Four places, and three of them are structural rather than a matter of a few benchmark points.

#1. Generative media - Claude has none at all

Claude models are text and image in, text out. There is no image generation, no video generation, no audio or music generation, and no video or audio input either. OpenAI ships GPT-Image-2 for images, Sora 2 for video (priced per second of output), GPT-Realtime-2.1 for speech-to-speech, and a transcription and translation line. If your work touches any of those, Claude is not a candidate and no amount of reasoning quality changes that. Our feature and plan matrix lists what Claude actually does instead.

#2. The free tier, and the gap above it

Since 6 August 2026 ChatGPT Free offers unlimited text chats with GPT-5.6 Luna, funded by advertising for US users, and ChatGPT Go at $8/month extends that with roughly ten times the file, upload and image allowances. Claude Free is a metered weekly pool with a five-hour rolling refresh, and Anthropic sells nothing between $0 and $20. That leaves an $8–$12 band that OpenAI and Google both occupy and Anthropic does not. What $0 actually gets you on Claude is itemised on the free plan page.

#3. Distribution through Microsoft

Microsoft 365 Copilot puts GPT models in front of every Office seat. Business pricing starts at about $18/user/month on annual billing (promotional through 30 September 2026, regular $21) per Microsoft's pricing page, rising to $32 for Business Premium bundles. Anthropic has no equivalent default-placement channel - it ships Microsoft 365 add-ins including Claude for Excel, but as an install, not as the assistant already in the ribbon for a billion seats.

#4. Agentic breadth and computer use

OpenAI wins OSWorld 2.0, BrowseComp, DeepSWE v1.1, Terminal-Bench 3.0 and Agents' Last Exam on its own published tables, and Codex is a mature competitor to Claude Code rather than a follower. Anthropic's computer use tooling only reached general availability on 19 August 2026, alongside a new browser-use tool. That is late relative to the benchmark record.

#Where does Claude clearly beat ChatGPT?

Also four places, and they are narrower but real.

The fourth is enterprise pricing transparency. Anthropic publishes Enterprise at a flat $20/seat/month plus usage at API rates, with SCIM, audit logs, a Compliance API, custom data retention, HIPAA-ready configuration and IP allowlisting listed openly. OpenAI Enterprise is quote-only. That said, OpenAI advertises customer-managed encryption keys and data residency across ten regions, and Anthropic's public pricing page lists neither - so "better" here depends on which control you actually need.

#How do the plan prices compare?

At $20 and at $100/$200 the two are at parity. The difference is at the bottom of the range.

Sources: claude.com/pricing; ChatGPT Plus and Pro tiers per the OpenAI Help Center; ChatGPT Go per OpenAI's launch post; Business and Enterprise per OpenAI Business Pricing. Full Claude plan mechanics are on the pricing page.

#What about API cost for developers?

Claude is the more expensive family, and it is not close at the bottom of the range. GPT-5.6 Luna at $0.20 in / $1.20 out per million tokens is roughly a tenth of Sonnet 5's input rate and an eighth of its output rate, and OpenAI's marketing explicitly claims Terra and Luna beat Fable 5 on some evals "at around one-sixteenth the cost". On Agents' Last Exam, Luna's 50.3% does beat Fable 5's 40.5%.

Two levers close some of that gap on the Claude side: the Batch API's flat 50% discount, and prompt caching, where cache reads cost 0.1× base input. Both are worked through with worked examples on the Claude API page. One caveat cuts the other way - Sonnet 5's tokenizer emits roughly 30% more tokens for the same text than Sonnet 4.6 did, so the sticker-price saving from its permanent $2/$10 rate is smaller in practice than it looks.

#Claude or ChatGPT: which should you actually use?

By job, not by brand. This table is the short version of everything above.

Plenty of people run both, and at $20 each that is a defensible answer. If you want a wider field than these two - including the cheap open-weight models that are now within a few index points of the frontier - see Claude alternatives. The Google side of the argument, where the gaps are sharper still, is on Claude vs Gemini. And if you have already decided on Claude, the model comparison guide works out which of the four to actually call.

#Frequently asked questions