GPT-5.5 vs Claude Opus 4.8 vs Gemini 3.1 Pro: which is actually cheapest?

Two of these three models cost the same per token. One of them costs nearly twice as much to actually use. That is not a contradiction — it is the whole point, and it is why comparing pricing pages is the wrong way to pick a model.

Everything below is measured, sourced and dated. Where a number is vendor-reported rather than independent, we say so.

The list prices

Start with the sticker, because that is where everyone starts. All figures are per 1M tokens from the vendors' own pricing pages, retrieved July 2026.

On paper the story is simple. Gemini is cheap, Claude is slightly cheaper than GPT on output, and everything else is a rounding error. Every one of those conclusions is wrong once you measure what these models actually do.

What they actually cost to run

Artificial Analysis runs the same evaluation suite against every model and publishes the dollar cost of doing so. This is the closest thing the industry has to a like-for-like cost benchmark, because the workload is identical.

Opus 4.8 costs 1.9× more to run than GPT-5.5 while charging 17% less per output token. Gemini 3.1 Pro costs a fifth of Opus. If you had picked by price per token you would have picked wrong.

Why sticker price lies

Cost per task is price per token multiplied by tokens per task. Vendors publish the first number and hide the second. Four things move it.

1. The tokenizer decides how many tokens your text becomes

The same paragraph is a different number of tokens on each vendor's tokenizer. Anthropic's newer models are notably token-hungry.

2. Reasoning tokens are billed but often invisible

All three vendors charge reasoning at the output rate. You cannot see most of it.

3. Models differ enormously in how much they say

Some models take 45M tokens to finish a job. Others take 120M. That is a 2.7× difference in output volume for the same work.

4. Long-context and caching rules diverge

One vendor doubles your input price past a threshold. Another charges rent on your cache. The third does neither.

Tokenizers: the same text is not the same tokens

Anthropic states on its own pricing page that Opus 4.7 and later, Fable 5, Mythos 5 and Sonnet 5 use a newer tokenizer producing "approximately 30% more tokens for the same text". Claude Opus 4.8 is on that tokenizer. Same sticker price, more tokens per unit of text.

An independent measurement using Anthropic's own count_tokens endpoint found the inflation is not uniform. English prose came in at 1.20×. Technical documentation hit 1.47×. A real instruction file measured 1.45×. Weighted across seven realistic coding samples the average was 1.325×.

That matters because technical English is exactly what sits in a system prompt. If your prefix is documentation and instructions, budget for 45% more tokens, not 30%.

What we could not verify

There is no published side-by-side count of identical text run through the GPT-5.5, Opus 4.8 and Gemini 3.1 Pro tokenizers. Anyone showing you a neat three-way table has estimated it. The one hard multilingual datapoint we found: a localisation study measured Claude encoding Tamil at 1.19 characters per token against Gemini's 4.24 — roughly a 3.5× cost difference for the same output. If you serve non-English users, tokenizer choice is a first-order cost decision.

Hidden thinking tokens

All three vendors bill reasoning tokens at the output rate. None of them show you most of what you paid for.

OpenAI adds a warning worth reading twice: "you could incur costs for input and reasoning tokens without receiving a visible response". A reasoning model that fails to answer still bills you.

Every vendor now exposes the billed count. Read it. On OpenAI it is usage.output_tokens_details, on Anthropic usage.output_tokens_details.thinking_tokens, on Google usage.total_thought_tokens. Most teams have never looked.

Verbosity: how many tokens does the model need to finish?

This is the variable that decides the bill, and only one organisation publishes it consistently. Artificial Analysis reports the output tokens each model emits to complete the identical suite.

Opus emits 2.7× the output of GPT-5.5 for the same work. At $25 versus $30 per million that difference is not close to being recovered.

The generational proof

The cleanest evidence that token efficiency beats token price comes from OpenAI's own product line. Artificial Analysis on the GPT-5.5 launch: per-token pricing doubled from GPT-5.4 to $5 / $30, but a ~40% reduction in token use largely absorbed the hike — the net cost to run the Index rose only about 20%.

A model can double its price and barely move your bill. Or hold its price and quietly double it.

Long context inverts the ranking

Above a certain prompt size, the cheap model stops being cheap and the expensive one stops being expensive.

At a 500k-token prompt, Opus 4.8 becomes cheaper than GPT-5.5 on both input and output — $5 against $10, $25 against $45. Anthropic's caching and batch discounts still apply across the full 1M window too.

So the honest answer to "which is cheapest" is: it depends on your prompt size, and the crossover is real. If you are processing large documents, whole codebases or long conversation histories, run the numbers again above the threshold. They will not look like the pricing page.

Caching rules differ, and it matters more than people think

All three cut cached input by 90%. What differs is what they charge you to get there.

Google's storage fee is the line item nobody models

At $4.50 per 1M tokens per hour, a 200,000-token cached context held across an eight-hour working session costs $7.20 in storage alone, before a single read. Gemini's headline price is genuinely the lowest on this page. Its cache is the only one that charges rent.

Anthropic charges to write, not to store

The 1.25× write fee is real but bounded: Anthropic's own break-even is a single read at the five-minute TTL. For an agent hammering the same prefix, that is one request. For a low-traffic app whose cache expires between users, you may be paying 1.25× input forever and reading it back never.

Speed is a cost, not a feature

Combine the two columns and the picture sharpens. Opus emits 120M tokens at 61.9 tokens per second. Gemini emits 56M at 148.1. Opus needs roughly five times more wall-clock generation time for the same suite. In an agent loop, wall-clock time is engineer time, queue depth and timeout risk — all of which cost money that never appears on the API invoice.

What do you get for the money?

Cost only matters relative to capability, so here is the honest quality picture — with a large caveat attached.

Read that table with suspicion. Every SWE-bench Verified score above is the vendor's own number on the vendor's own harness, and none are independently verified. When the same models are run on Scale's standardised harness the scores drop by 17–21 points. The harness is the variable, not just the model.

The one comparison we would stake something on is Artificial Analysis's, because the workload is identical and the cost is measured rather than claimed. Their summary of GPT-5.5's launch, verbatim: "GPT-5.5 (medium) scores the same as Claude Opus 4.7 (max) on our Intelligence Index at one quarter of the cost (~$1,200 vs $4,800) — although Gemini 3.1 Pro Preview scores the same at a cost of ~$900."

Why we do not rank these models on quality

Because nobody can, honestly. Every headline coding score in circulation was produced by the vendor that sells the model, on a harness the vendor built. The only fair comparison is one where the workload, the harness and the scorer are the same for everyone — and on cost, that is what Artificial Analysis provides.

So which is cheapest for your workload?

Do not forget the small models

The biggest cost lever on this page is not choosing between three flagships. It is not using a flagship at all.

OpenAI's gpt-5.4-nano is 24× cheaper on output than GPT-5.5. Claude Haiku 4.5 is 5× cheaper than Opus 4.8. Gemini Flash-Lite is 8× cheaper than Gemini 3.1 Pro. No amount of vendor-shopping between flagships recovers a 24× ratio.

An AI gateway makes this practical, because you can route per request without three SDKs and three invoices. How gateways charge for that convenience is covered in our gateway guide, and the current landed prices are on our model price comparison.

Routing is a config change, not a rewrite

The objection to routing is usually operational: three vendors means three SDKs, three keys and three invoices. That objection is what an OpenAI-compatible gateway removes — one base URL, one balance, and the model chosen per request by changing a string.

How to measure this yourself in an afternoon

1. Take 50 real requests from your production logs. Not synthetic prompts — real ones.
2. Run them through each candidate model with identical settings.
3. Record four numbers per request: input tokens, cached input tokens, output tokens, and reasoning tokens. Not three.
4. Compute the dollar cost from each vendor's current rates, including the cache write fee or storage fee where it applies.
5. Score the outputs for quality on your own rubric, because a cheap wrong answer costs more than an expensive right one.
6. Divide cost by successful task, not by token. That is your real number.

Nobody else's benchmark is about your workload. Ours included.

Common mistakes

• Comparing price per token. Opus 4.8 is cheaper per output token than GPT-5.5 and costs 1.9× as much to run.
• Ignoring reasoning tokens. They are billed at the output rate and mostly invisible. They are often the largest line.
• Modelling long context at the headline price. Past 272k, GPT-5.5 costs $10 / $45, not $5 / $30.
• Forgetting Google's cache storage fee. $4.50 per 1M tokens per hour turns a long session expensive.
• Trusting vendor-run benchmarks. The same models lose 17–21 points on a standardised harness.
• Choosing a flagship at all. A 24× price ratio to a small model sits unused in most stacks.

FAQ

Which is cheapest: GPT-5.5, Claude Opus 4.8 or Gemini 3.1 Pro?

On measured cost to run the same evaluation suite, Gemini 3.1 Pro at $815, GPT-5.5 at $2,159 and Claude Opus 4.8 at $4,011. On list price the order looks different, which is exactly the point.

Why does Claude cost more if its output price is lower?

Because it emits more tokens. Opus 4.8 produced 120M output tokens to complete the suite; GPT-5.5 needed 45M. A 17% lower price per token does not survive a 2.7× higher token count.

Is Gemini really the cheapest?

On short-to-medium prompts, yes, by a wide margin. Above a 200k prompt its price rises to $4 / $18, and it charges $4.50 per 1M tokens per hour to store a cache — a fee neither competitor has. On long-lived contexts that gap narrows.

Do all three vendors charge for reasoning tokens?

Yes, all three bill them at the output rate. OpenAI does not return them at all. Anthropic omits them by default on Opus 4.8. Google returns a summary. Every vendor exposes the billed count — read it.

What is the Claude tokenizer change and does it affect Opus 4.8?

Yes. Anthropic states its newer models, including Opus 4.8, produce about 30% more tokens for the same text. Independent measurement found 1.20× on prose and up to 1.47× on technical documentation.

Which model is cheapest for long documents?

Claude. Anthropic applies no long-context surcharge — a 900k-token request bills at the same rate as a 9k one. OpenAI doubles input above 272k; Google raises Gemini above a 200k prompt.

Does prompt caching change the ranking?

It can. All three give 90% off cached reads, but Anthropic charges 1.25× to write the cache and Google charges hourly storage. On a hot prefix hit many times per minute, Anthropic's write fee amortises to nothing. On a cold one, it does not.

Are the SWE-bench scores trustworthy?

Treat vendor-reported SWE-bench Verified numbers with caution — none are independently verified, and the same models score 17–21 points lower on a standardised harness. Compare models only within the same harness.

Is GPT-5.5 better value than Claude Opus 4.8?

On measured cost per unit of benchmark performance, yes — roughly half the cost for a comparable score. On very long prompts the ranking flips, because GPT-5.5 doubles its input price above 272k tokens and Claude does not.

How much cheaper is a small model?

Dramatically. On output tokens, gpt-5.4-nano is 24× cheaper than GPT-5.5, Claude Haiku 4.5 is 5× cheaper than Opus 4.8, and Gemini Flash-Lite is 8× cheaper than Gemini 3.1 Pro. Routing beats vendor-shopping.

Does the tokenizer difference apply to non-English text?

It varies enormously. Anthropic's new tokenizer added only about 1% for Japanese and Chinese prose but up to 47% for technical English. Separately, one study measured Claude encoding Tamil at 1.19 characters per token against Gemini's 4.24 — roughly 3.5× the cost for the same output.

Should I use a gateway to compare these models?

It makes A/B testing far easier — one key, one balance, one line changed to swap models. Just check the gateway's landed price first: in our own comparison, prices for the identical model range from 40% below vendor list to 10% above it.

Which model has the best free tier?

Google. Gemini offers a genuine ongoing free API tier on its Flash-class models, though not on Gemini 3.1 Pro. Anthropic gives only a small unspecified starter credit; OpenAI's free tier is heavily rate-limited.

Why is Opus so much slower?

Artificial Analysis measures Opus 4.8 at 61.9 output tokens per second against Gemini 3.1 Pro's 148.1. Combined with its higher token volume, Opus needs roughly five times more wall-clock generation time to complete the same suite.

The bottom line

Price per token tells you almost nothing. Two of these three models charge the same for input, and one of them costs nearly twice as much to actually run. The variable that decides your bill is how many tokens the model needs to finish the job — and that number appears on no pricing page.

If you want a default: Gemini 3.1 Pro is the cheapest flagship for most workloads by a wide margin. GPT-5.5 is the value pick for agentic work, because it finishes in fewer tokens. Claude Opus 4.8 wins on long context, where it is the only flagship that does not raise its price, and on the hardest reasoning tasks, where you are buying capability rather than economy.

And before you optimise any of that: check whether the task needs a flagship at all. There is a 24× price ratio sitting unused inside OpenAI's own catalogue.

Read next

• Claude API pricing 2026 — every multiplier and the 30% tokenizer tax.
• Prompt caching explained — the biggest lever on any model’s bill.
• What AI agents really cost — why input, not output, is the bill.

Sources

• OpenAI — API pricing documentation. Retrieved 14 July 2026. List prices, cached input, batch, long-context tier, priority pricing.
• OpenAI — Reasoning models guide. Retrieved 14 July 2026. Reasoning-token billing.
• Anthropic — Claude Platform pricing documentation. Retrieved 14 July 2026. List prices, cache multipliers, tokenizer statement, long-context rule.
• Anthropic — Extended thinking documentation. Retrieved 14 July 2026. Thinking-token billing and omitted display.
• Google — Gemini API pricing. Retrieved 14 July 2026. List prices, cache storage fee, batch discount, free tier.
• Google — Gemini thinking documentation. Last updated 6 July 2026.
• Artificial Analysis — model pages for Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro Preview: cost to run the Intelligence Index, output tokens, verbosity, output speed. Retrieved 14 July 2026.
• Artificial Analysis — GPT-5.5 launch analysis, April 2026. Token-efficiency and cost commentary.
• Independent tokenizer measurement via Anthropic's count_tokens endpoint, April 2026.
• Scale SEAL — SWE-bench Pro standardised harness results, 2026.
• Multilingual tokenizer efficiency study (Tamil vs English chars-per-token), April 2026.