Claude vs Gemini vs ChatGPT

Three-way comparisons usually collapse into a scoreboard: this one wins at code, that one wins at writing, the third one wins at long documents. Those verdicts have a shelf life measured in weeks, and by the time you read them the ranking has moved.

Something more durable shows up if you compare the three vendors' pricing structures rather than their outputs. Anthropic, OpenAI and Google have arrived at genuinely different philosophies about how to charge you and how much to tell you — and unlike benchmark rankings, those choices are stable enough to plan around.

Verified against all three vendors' own documentation on 5 September 2026.

	Cheapest	Mid	Flagship	Publishes context?	Publishes cutoff?	Publishes retirement?
Claude	Haiku 4.5 $1 / $5	Sonnet 5 $2 / $10	Fable 5.1 $10 / $50	Yes	Yes (month)	Yes
ChatGPT / OpenAI	Luna $0.20 / $1.20	Terra $2 / $12	GPT-6 Astra $10 / $50	Yes	Yes (exact date)	No
Gemini	3.1 Flash-Lite $0.25 / $1.50	3.8 Flash $0.75* / $3.75*	3.1 Pro (preview) $2–4 / $12–18	No	No	No

Prices per million input / output tokens. * Promotional — see below.

Google is the only one whose price has an expiry date printed on it

Gemini 3.8, 3.7 and 3.6 Flash are $0.75 input and $3.75 output through 31 December 2026, then $1.50 and $7.50 from 1 January 2027. That is not a forecast or a rumour; it is stated on Google's own pricing page as a dated change.

If you are modelling unit economics on Gemini Flash today, you are modelling on a rate with a known expiry roughly four months out. A margin that works at $0.75 has to work at $1.50, and anything with a twelve-month payback assumption crosses the boundary before it pays back.

There is a second structural feature nobody plans for: Gemini's Pro tier prices by prompt length. Gemini 3.1 Pro is $2 input for prompts of 200,000 tokens or fewer and $4 above that, with output going $12 to $18. It is a cliff, not a gradient — the same request either side of the boundary costs very differently. That is a trap for precisely the workload a large context window makes attractive.

And a quirk worth knowing: Gemini 3.5 Flash costs $1.50 / $9.00, which is more than the newer 3.8, 3.7 and 3.6 Flash. Version numbers do not predict price in this lineup. Read the table.

OpenAI has by far the widest price ladder

From GPT-5.6 Luna at $0.20 / $1.20 to GPT-6 Astra at $10 / $50 is a fifty-fold spread on input, inside one vendor, with every rung sharing the same 1.05M context window, the same 128K output ceiling, and the same February 2026 cutoff on the 5.6 family.

That uniformity is the useful part. Moving down OpenAI's ladder costs you model quality but nothing structural — you do not lose context, output headroom or recency. Moving down Anthropic's ladder does: Haiku 4.5 drops to a 200K window, 64K output and a February 2025 cutoff. Moving down Google's is hard to reason about at all, because Google publishes neither context windows nor cutoffs per model on its pricing page.

For high-volume production work, Luna has no real competitor in any of the three lineups.

Anthropic is the only one that tells you when a model retires

Every current Claude model carries a published retirement commitment: 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, Haiku 4.5 not sooner than 15 October 2026.

Neither OpenAI's nor Google's model documentation publishes an equivalent per-model commitment. If you need to put a migration date in a planning document rather than watch a changelog, that asymmetry is worth more than a benchmark point or two.

Note the Haiku 4.5 date, though. It is the nearest deadline in any of the three lineups, and it belongs to the model most likely to be chosen for cost reasons in exactly the high-volume systems that are hardest to migrate.

The Anthropic tokenizer means its prices aren't directly comparable

One correction that applies to every table on this page, including mine. Anthropic's current tokenizer — introduced with Claude Opus 4.7 — fits roughly 555,000 words into a million tokens, where the previous one fit about 750,000.

So Claude's per-token rates buy less text than a straight comparison implies, and the gap against OpenAI and Google on identical sticker prices is wider in practice than it looks. Fable 5.1 and GPT-6 Astra are both $10 / $50; they are not the same price for the same document.

There is no way to fix this in a table. The only honest measurement is to run representative prompts through all three and compare the invoices.

Google's flagship is still preview

Something structural shows up in Google's model list that shapes the whole comparison: the stable Gemini models are almost all Flash tiers, while Gemini 3.1 Pro remains preview.

Preview models carry no stability promise. If you need a Pro-class Gemini under a support commitment, check its status before designing around it — the equivalent tiers at Anthropic and OpenAI are generally available with published lifecycles.

Google's lineup is genuinely Flash-first, and the newest stable Flash models are described for exactly the complex and agentic work that would once have implied a Pro tier. That is a real position, not a gap. It just is not the same shape of offer as the other two.

If you mean the apps, not the API

Google publishes the clearest consumer ladder of the three: Free, Google AI Plus at $4.99, Google AI Pro at $19.99, and Google AI Ultra from $99.99 (5× limits) or $199.99 (20× limits), with storage and adjacent Google products bundled in at each step.

Anthropic publishes Free at $0, Pro at $20 a month or $17 annually, and Max from $100 a month, with Team seats at $25 and $125 monthly. Claude Code is included on every paid tier.

OpenAI's ChatGPT tiers follow a similar shape, but we are not printing figures for them. OpenAI's pricing pages refuse automated retrieval, and this site does not publish numbers it could not verify at the source. Check openai.com/chatgpt/pricing in a browser.

How to actually choose

• Running at volume with well-specified tasks? GPT-5.6 Luna. The 50× internal spread means you can drop to it without losing context or output headroom.
• Budgeting beyond this year on Gemini? Model Flash at $1.50 / $7.50, not $0.75 / $3.75, and check whether your prompts cross Pro's 200k cliff.
• Need a migration date you can plan around? Anthropic is the only vendor publishing one — and check Haiku 4.5's October 2026 commitment before building on it.
• Need a supported flagship from Google? Confirm Gemini 3.1 Pro's preview status first.
• Comparing cost seriously? Run your own prompts through all three. The tokenizer difference alone makes the rate cards non-comparable.

For the pairwise versions of this question: ChatGPT vs Claude goes deeper on the two mirrored price ladders, Gemini vs ChatGPT on Google's price structure against OpenAI's, and Claude Opus vs Sonnet vs Haiku on picking a tier once you have picked a vendor.

The same pairings, as structured comparisons

• Gemini 3.7 Flash vs GPT-5.6 Sol — and why a Flash tier against a flagship is not like-for-like
• Claude Opus 5 vs GPT-5.6 Sol — the closest thing to a fair flagship pairing
• Claude Opus 5 vs Claude Sonnet 5 — same specs, 2.5× the price

---

Verified 5 September 2026 against Anthropic's models overview (platform.claude.com), OpenAI's models documentation (developers.openai.com), Google's Gemini API pricing (ai.google.dev) and Gemini subscription plans (gemini.google). Google publishes no per-model context window or knowledge cutoff on its pricing page; those cells are left blank rather than filled from secondary sources. ChatGPT consumer subscription prices are omitted because openai.com returned HTTP 403 to every automated request.