Claude 5.5 vs GPT-6, Tier by Tier: Claude Was Faster in Every Pair, GPT-6 Was Cheaper in Every Pair

Anthropic's Claude 5.5 family and OpenAI's GPT-6 family line up tier for tier, and in three of the four tiers the list prices are identical. We ran all eight models through the same three script-graded tasks, five times each.

All eight passed all 15 calls. Our grader cannot separate them on accuracy. What it can measure is this: in every pair, the Claude model answered faster, and in every pair, the GPT-6 model cost less. At identical list prices, the cost gap comes from how many tokens each model spends on the same answer.

The pairs

List prices per million tokens (input / output), from shared/pricing.ts:

The upper tier is the one that does not match. OpenAI has no GPT-6 model at $4 / $20, so Opus 5.5 sits at twice the price of the nearest Sol. We paired it with GPT-6.1 Sol, the newer of the two Sol models.

Quality: no separation

Every model returned bare JSON (no code fences) with the right order id, the new email, 89.5 as a number and "high" priority. Every model got 580 and ended with the required Answer: line, and every model produced exactly three short bullets with nothing around them.

That is 120 graded calls with no failures, so the rest of this post is about cost and latency, not accuracy.

Cost: GPT-6 cheaper in every pair

Cost per battery pass (one call of each task), averaged over five passes, at list price:

The upper-tier gap is mostly the price list: Opus 5.5 charges twice as much per token. The other three tiers charge the same per token, so the 20–45% gaps there come from token counts.

Where the extra tokens come from

Two things push the Claude bills up at equal rates.

The same prompt counts as more input tokens. The JSON task's prompt was reported as 209 input tokens by Anthropic and 141 by OpenAI. Our 10-word token probe ("Reply with only the word OK…") showed the same pattern: 25 tokens on every Claude model, 18 on every GPT-6 model. The providers count the same text differently, and you pay for their count.

Claude models wrote more on two of the three tasks. Median output tokens:

Some of that is thinking. Both providers bill reasoning as output and include it in the output count, so the figures above already contain it and Promptster's displayed cost matches what you are billed. Opus 5.5 used 84–96 thinking tokens on every arithmetic run and 103–114 on every bullets run, which is why its bullets answers ran to 174–195 output tokens for roughly 200–230 characters of text. Haiku 5.5 thought for 126–145 tokens on the arithmetic task and not at all on the other two. Fable 5.1 and Sonnet 5.5 reported zero thinking tokens on every call.

On the OpenAI side, Luna does the most reasoning (75–125 reasoning tokens on every bullets run), which is why it has the highest output counts of the GPT-6 models. At $0.50 per million output tokens that barely shows in the bill.

GPT-6 used more output tokens than its Claude pair in three places: Astra on bullets, and Luna on JSON and bullets. Luna still came out cheaper overall because Haiku's arithmetic answers were long.

Latency: Claude faster in every pair

Median wall-clock time per call, five runs each:

The Claude model had the lower median on all twelve task-and-tier combinations. The widest gap is at the top: Fable 5.1 answered the bullets task in a third of Astra's time. The narrowest is the mid tier, where Sonnet 5.5 and GPT-6 Sol are within 0.3 s of each other on every task.

GPT-6 Sol beat GPT-6.1 Sol on latency on every task while costing slightly less. If you want an OpenAI model to put against Opus 5.5 purely on speed, 6 Sol is closer, at 4.37× cheaper than Opus 5.5 per pass.

Picking a tier

Top tier. Fable 5.1 cost 20% more than Astra and was roughly twice as fast. For an interactive product that trade favours Fable; for batch work, Astra is the cheaper of two models our battery cannot separate.

Upper tier. Opus 5.5 cost 3.95× as much as GPT-6.1 Sol for the same graded results, and was faster. This battery gives no reason to pay that premium; the reason would be work it cannot measure, such as long agentic tasks. Test that yourself first.

Mid tier. The closest pair: same list price, latency within 0.3 s, Sonnet 5.5 about 45% more per pass. Pick on what this battery does not measure, such as tool use or your own prompts.

Small tier. Both cost less than a hundredth of a cent per call. Haiku 5.5 was the fastest model in the whole comparison on two of the three tasks. Luna was the cheapest. At a million battery passes that is $298 against $219.

Method

Three script-graded tasks: strict JSON extraction from a messy support ticket (four keys, exact match, code fences fail), a four-step arithmetic word problem with answer 580, and an instruction-adherence task asking for exactly three bullets of 12 words or fewer with no preamble. Five runs per task per model, requested at temperature 0.7 and max_tokens 1,500, run 2026-10-19 through Promptster's own provider handlers. GPT-5-and-later and Claude 5-family models reject a custom temperature, so the handlers send none to any model in this post, and OpenAI calls go out with reasoning_effort: medium. Costs are from shared/pricing.ts at list rate with no cache or batch discounts. Latency is from one machine in one session, not a host SLA.

For the last head-to-head between these two families, see Claude Opus 5 vs GPT-5.6 Sol; for the wider picture, Benchmarking open-source vs. closed-source models.

Try it yourself

• Run the three-bullet instruction task — the task with the widest latency gap between the pairs, no account needed

To put any Claude tier against any GPT-6 tier on your own prompts and see the token counts that drive the bill, that's Promptster.

All figures from content/blog/data/q4-2026-battery/ (summary plus one raw file per call, including each provider's usage object) and the 10-word token probe in content/blog/data/q4-2026-token-overhead/, run 2026-10-19. Costs at list price from shared/pricing.ts; for every model in this post the output count already includes reasoning, so displayed and billed cost are the same. Medians over five runs; ranges are min–max across the five.