ChatGPT vs Claude
Both flagships now cost exactly $10/$50 — but Anthropic's tokenizer fits 555k words into a million tokens where the old one fit 750k, so identical rates aren't identical bills. Plus the retirement date on Haiku 4.5 that's weeks away.
Ask ChatGPT or Claude
in 2026 and most answers still describe a world where one was better at code and the other better at writing. That world is gone. Put the two current lineups side by side and the striking thing is how deliberately they now mirror each other — same context class, same output ceiling, and price ladders that land within pennies of one another at every rung.
Verified against Anthropic's and OpenAI's own model documentation on 5 September 2026.
| Claude | ChatGPT / OpenAI | |||
|---|---|---|---|---|
| Flagship | Fable 5.1 | $10 / $50 | GPT-6 Astra | $10 / $50 |
| Workhorse | Opus 5 | $5 / $25 | GPT-5.6 Sol | $4 / $20 |
| Balanced | Sonnet 5 | $2 / $10 | GPT-5.6 Terra | $2 / $12 |
| Cheap tier | Haiku 4.5 | $1 / $5 | GPT-5.6 Luna | $0.20 / $1.20 |
Prices per million input / output tokens.
Two lineups, four rungs each, and at the top the numbers are simply identical. That convergence is the story — and it means the things that actually separate them are no longer on the price sheet.
The identical $10 / $50 flagships are not identically priced
Claude Fable 5.1 and GPT-6 Astra both cost $10 per million input tokens and $50 per million output. They are not, however, the same price for the same work, because a token is not the same amount of text on both sides.
Anthropic's current tokenizer, introduced with Claude Opus 4.7, fits roughly 555,000 words into a million tokens. The previous one fit about 750,000. Anthropic publishes this plainly in its own context-window documentation, and the implication is rarely followed through: the same document, priced at the same headline rate, consumes more tokens on Claude than the sticker comparison implies.
So identical per-token pricing is not identical per-page pricing, and every $10 vs $10
table you will read this year is comparing units that do not match. If cost is genuinely the deciding factor, the only honest measurement is to run your own representative prompts through both and compare the invoice — not the rate card.
This cuts the other way too. It means Claude's cheaper tiers are also buying you less text per dollar than their numbers suggest, which makes GPT-5.6 Luna's position at the bottom even stronger than the table shows.
Luna has no counterpart in the Claude lineup
The one rung where the mirror breaks is the bottom. Haiku 4.5 at $1 / $5 is Anthropic's cheap option; Luna at $0.20 / $1.20 is five times cheaper on input and four times cheaper on output — and unlike Haiku, it carries the full 1.05M context window, the same 128K output ceiling and the same February 2026 cutoff as its flagship siblings.
Haiku 4.5 makes real concessions for its price: a 200K context window against 1M, 64K max output against 128K, and a February 2025 knowledge cutoff that is more than a year behind the rest of the lineup. It is a previous-generation model kept in the price list, not a small version of Claude 5.
For high-volume, well-specified work — classification, extraction, routing, bulk summarisation — that gap is the most consequential difference between the two vendors right now.
Anthropic tells you when a model dies; OpenAI's model list doesn't
Here is a difference that never appears in comparison tables and matters enormously if you are shipping something.
Anthropic publishes a retirement commitment for every model: Fable 5.1 not sooner than 1 September 2027, Opus 5 not sooner than 24 July 2027, Sonnet 5 not sooner than 30 June 2027, and Haiku 4.5 not sooner than 15 October 2026 — which is weeks away, and the sharpest thing on this page. If you are building on Haiku 4.5 today, you are building on the shortest runway in either lineup.
OpenAI's model documentation lists IDs, prices, context and cutoffs, but no equivalent per-model retirement commitment. That is not a criticism of the models; it is a planning difference. One vendor lets you put a date in your migration calendar and the other asks you to watch for announcements.
Knowledge cutoffs: precision and freshness pull in opposite directions
OpenAI publishes cutoffs to the day — GPT-6 Astra 30 April 2026, the whole GPT-5.6 family 16 February 2026. Anthropic publishes to the month: Fable 5.1 June 2026, Opus 5 May 2026, Sonnet 5 January 2026, Haiku 4.5 February 2025.
The freshest model in either lineup is Claude Fable 5.1 at June 2026. The most precisely specified is any OpenAI model. Which you prefer depends on whether you need to reason about recency or merely to state it — and note that Anthropic distinguishes a reliable knowledge cutoff
from the broader training-data cutoff, which is a more careful claim than a single date implies.
One caution worth carrying: a spread of sixteen months separates Fable 5.1 from Haiku 4.5 inside the same price list. Cutoff is not a vendor-level property. Check it per model.
If you mean the apps, not the API
Most people asking this question mean the chat products, not the token prices.
Anthropic publishes its plan pricing openly: Free at $0, Pro at $20 a month or $17 on an annual plan, Max from $100 a month, Team seats at $25 standard and $125 premium monthly ($20 and $100 annually), and Enterprise at a seat price plus usage at API rates. Claude Code is included on every paid tier and shares your plan's usage limits rather than carrying its own.
OpenAI's ChatGPT tiers follow a comparable shape — a free tier, an individual subscription, a higher-usage individual tier, and business and enterprise seats — but we are not printing figures for them here. OpenAI's pricing pages refuse automated retrieval, and this site's rule is that a number we could not verify at the source does not get published. Check openai.com/chatgpt/pricing directly; it renders normally in a browser.
That is a smaller caveat than it sounds. Both vendors move consumer pricing more often than API pricing, so any figure in any comparison article — including the ones above — should be treated as a pointer to the vendor's page rather than a substitute for it.
How to actually choose
- Running high-volume, well-specified tasks? GPT-5.6 Luna, and it isn't close. Nothing in the Claude lineup competes at $0.20 per million input tokens without giving up context.
- Need the freshest knowledge? Claude Fable 5.1, at a June 2026 cutoff.
- Need to plan a migration calendar? Anthropic's published retirement dates are the only ones you can put in a spreadsheet — and check Haiku 4.5's October 2026 date before you build on it.
- Comparing cost seriously? Ignore both rate cards. Run your own prompts and compare the bill, because the tokenizer difference means the rate cards are not measuring the same thing.
- Everything else? These two are close enough on paper that paper cannot separate them. Ten minutes with your real prompts will.
For a wider view, Claude vs Gemini vs ChatGPT adds the third lineup and the pricing philosophies behind all three. And before you commit to any number in any comparison, including this one, check why they go stale so fast.
The same pairings, as structured comparisons
These head-to-heads also exist as structured pages carrying their own verification dates, so you can see what has moved since this was written:
- Claude Opus 5 vs GPT-5.6 Sol — the workhorse rung, and why the tokenizer widens the gap
- Claude Opus 5 vs Claude Sonnet 5 — picking a tier inside the Claude lineup
- Claude Haiku 4.5 vs Claude Sonnet 5 — why Haiku is a generation behind, not a smaller Claude 5
Verified 5 September 2026 against Anthropic's models overview (platform.claude.com) and pricing page (claude.com/pricing), and OpenAI's models documentation (developers.openai.com). ChatGPT consumer subscription prices are deliberately omitted: openai.com returned HTTP 403 to every automated request, and this site does not publish figures it could not read at the source.