GPT-5.5 vs Claude Opus 4.7: API Pricing, Tokenizers, and the Real Cost per Task

Two frontier models arrived a week apart. Anthropic released Claude Opus 4.7 on April 16 at the same price as Opus 4.6: $5 per million input tokens and $25 per million output tokens. OpenAI introduced GPT-5.5 yesterday and released it to the API today at $5 and $30, double the price of GPT-5.4.

On the price list, Opus 4.7 is the cheaper model. On Artificial Analysis's independent measurement, published yesterday, it is the more expensive one by a wide margin: GPT-5.5 at medium effort matched the score of Opus 4.7 at max effort for about a quarter of the cost.

Both statements are true, and the gap between them is why I compare models by cost per completed task, not cost per token. Prices and product details are reported by OpenAI and Anthropic, the cost and hallucination figures come from Artificial Analysis, and anything else is my arithmetic or my opinion, labelled as such. I have not run either model.

GPT-5.5 and Claude Opus 4.7 pricing against cost per task

Model	Input, per million tokens	Output, per million tokens	Cost to run AA's Intelligence Index
GPT-5.4	$2.50	$15	Baseline for GPT-5.5
GPT-5.5 (xhigh)	$5	$30	About 20% more than GPT-5.4
Claude Opus 4.6 (max)	$5	$25	About $4,970
Claude Opus 4.7 (max)	$5	$25	About $4,406, about 11% less than Opus 4.6

Sources: prices from the GPT-5.5 and GPT-5.4 model pages and Anthropic's announcement. Cost to run measured independently by Artificial Analysis, for Opus on April 17 and for GPT-5.5 on April 23.

Read across the rows. GPT-5.5 doubled its price per token, yet at xhigh it cost only about 20% more than GPT-5.4 to run the index, because AA measured about 40% fewer output tokens. Opus 4.7 kept its price and cost about 11% less than Opus 4.6, using 102 million output tokens where Opus 4.6 used 157 million. In both cases, the bill for the same work moved differently from the price per token.

The tokenizer is a price change nobody announces as one

Anthropic says Opus 4.7 uses an updated tokenizer, and the same input can map to roughly 1.0 to 1.35 times as many tokens, depending on the type of content. The price per token did not change. The number of tokens in your prompt may have.

A hypothetical retrieval request shows the size of it. By my arithmetic, 20,000 tokens of context and a 500-token answer cost about 11 cents on Opus 4.6. If the same context becomes 27,000 tokens on Opus 4.7, the top of Anthropic's range, and the answer stays at 500 tokens, the request costs about 15 cents. That is 31% more for identical text at an unchanged price.

Anthropic is open about the trade. It says Opus 4.7 also thinks more at higher effort levels, which means more output tokens, that the net effect on its own internal coding evaluation was favourable, and that customers should measure the difference on real traffic. That is the right advice, and it concedes that the price list cannot answer the question.

On AA's index, fewer output tokens more than paid for the tokenizer, by AA's own account. Whether that holds for you depends on the shape of your traffic. Agent workloads can read far more than they write: when I looked at the agent-team C compiler in February, its sessions consumed about 14 input tokens for every output token. Retrieval over long documents with short answers looks similar. On traffic like that, the tokenizer applies to every request and the output savings have little to work with. That is my inference, and only your own logs can settle it.

Same score, a quarter of the cost: what Artificial Analysis measured

Artificial Analysis tested GPT-5.5 before release at five reasoning settings, with access from OpenAI. Its headline is that GPT-5.5 leads the Intelligence Index by 3 points. The finding that matters to whoever pays the bill is further down.

Configuration	Cost to run the Intelligence Index	Result
GPT-5.5 (medium)	About $1,200	Same score as Opus 4.7 (max)
Claude Opus 4.7 (max)	About $4,800	Same score as GPT-5.5 (medium)
Gemini 3.1 Pro Preview	About $900	Same score as both
GPT-5.5 (low)	About $500	Close to Opus 4.7 (non-reasoning, high)
Claude Opus 4.7 (non-reasoning, high)	About $1,000	Close to GPT-5.5 (low)

Source: Artificial Analysis, April 23, 2026, on Intelligence Index v4.0, the version in use since January. Independent measurement.

Opus 4.7 charges the same per input token as GPT-5.5 and less per output token. By my arithmetic, the price list works in its favour, so the whole fourfold gap comes from how many tokens each configuration spent reaching the score. Even at xhigh, AA puts GPT-5.5 about 30% cheaper to run than Opus 4.7 at max. Medium is also GPT-5.5's default effort in the API, according to OpenAI's changelog.

Four caveats keep this honest:

The settings differ. This is GPT-5.5 at medium against Opus 4.7 at max. That is the right question, the cheapest configuration that reaches a score, but it compares configurations, not models.
The figure moves. AA's April 17 article put Opus 4.7 (max) at about $4,406 to run the index. Six days later it was about $4,800, roughly 9% higher by my arithmetic, and the April 23 article does not say why. Treat all of these as approximate.
The cheapest route was neither model. Gemini 3.1 Pro Preview reached the same score for about $900.
An index is not your workload. It averages many kinds of task. Your traffic is a few kinds of task repeated many times.

A completed task has to be a correct one

Cost per task means little if the answer is wrong. On AA-Omniscience, AA's test of factual knowledge, GPT-5.5 at xhigh had the highest accuracy in the comparison, 57%. It also had a hallucination rate of 86%, against 36% for Opus 4.7 at max and 50% for Gemini 3.1 Pro Preview.

AA defines that rate as the share of wrong answers among all the questions a model did not get right. It measures what a model does when it does not know: answer anyway, or say so. GPT-5.5 usually answers. OpenAI's system card points the same way from another angle. On conversations users had flagged for factual errors, GPT-5.5's individual claims were 23% more likely to be correct than GPT-5.4's, but its responses contained an error only 3% less often, because it makes more claims per response.

For cost per completed task, a confident wrong answer is the most expensive outcome, because someone pays later to find it. An abstention costs a retry or a human. So the metric I would use is cost per accepted result, with each wrong answer charged at what it costs your business to catch. That changes the ranking for some workloads and not others. A coding agent with a good test suite catches many of its own wrong answers. A support agent quoting a refund policy does not.

Refusals count too. Opus 4.7 is the first model to carry Anthropic's new cyber safeguards, which automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses, while Claude Mythos Preview stays limited. Anthropic offers security professionals a Cyber Verification Program. Without it, a blocked legitimate request is a failed task, and it belongs in the cost.

How I would compare GPT-5.5 and Opus 4.7 on your own workload

I have run this kind of comparison before. On Auto-RAG and MerlinX, I found the orchestrator model was 75% of workflow cost. The cheaper orchestrator model was first shown on a frozen evaluation suite, at 95% confidence, to be not worse than the model it replaced; the rollout then cut cost per conversation by 73% across 840 agents with no measurable quality loss. The unit was the conversation, not the token. For these two models I would set it up like this:

Freeze the suite first. Real tasks from your own traffic, with pass criteria written before either model runs.
Compare configurations, not models. Run each model at two or three effort levels. The cheapest configuration that passes is the answer, whoever makes it.
Take token counts from the bill. Use each provider's reported usage, never a character estimate or one vendor's tokenizer applied to both. A tokenizer change is what an estimate misses.
Measure caching, do not assume it. Hit rates depend on how your prompts are assembled, and OpenAI's changelog says GPT-5.5 supports only extended prompt caching, not in-memory caching.
Charge the failures. Count retries, refusals and wrong answers, each at its real cost.
Ask questions your documents cannot answer. Count abstentions and errors separately. That is the behaviour behind AA's 86% and 36%.
Prove it is not worse before you switch. A single run that looks the same is not evidence. State the confidence level with the result.

So the model that doubled its price is the cheaper one per task, the model that kept its price can charge more for the same words, and the cheapest route to that score belonged to a third company. Every price list involved was accurate.

Sources

Written by Eslam Hamed, a Technical Lead and AI Solutions Architect working on enterprise RAG, agentic AI, and model cost and evaluation.