Gemini 4 Argon vs Claude Opus 5.5: Argon Is Cheaper Until November 30, and Slower on Every Call

Gemini 4 Argon is on promotional pricing, and the promotion has an end date:

From December, the two list at exactly the same rate. We ran both through the same three script-graded tasks, five times each.

Both passed all 15 calls. Neither got an answer wrong, so the comparison comes down to cost and speed. Three findings:

1. Today, Argon is cheaper. Opus 5.5 cost 2.84× Argon as Promptster displayed it at the time, and 1.36× on our estimate of what Google actually bills.
2. From December 1, Argon's estimated bill is 1.47× Opus 5.5's, because Argon bills hidden thinking tokens that Opus does not need on this work.
3. Every Argon call was slower than every Opus 5.5 call on the same task. Argon's median was 2.6× to 5.2× slower, and two of its calls took more than 37 seconds.

Quality: no separation

Both returned bare JSON with the right order id, the new email, 89.5 as a number and "high" priority. Both reached 580 and ended with the required Answer: line. Both wrote exactly three short bullets and nothing else.

Cost

Until 2026-10-22, Promptster displayed Gemini costs from candidatesTokenCount, the visible answer. Gemini reports reasoning separately as thoughtsTokenCount, and Promptster's displayed cost left it out. It now counts it, so the display matches our billed estimate. Our billed estimate re-prices output as total tokens minus input, which counts the thoughts. Anthropic includes Opus 5.5's thinking in output_tokens, so for Opus the two figures are the same.

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

The billed estimate is our reconstruction from Google's usage object, not an invoice. It assumes thought tokens bill at the output rate.

Where Argon's hidden tokens came from

Argon did not think on every task. On JSON and arithmetic its usage object reported no thought tokens at all, and its costs there were well below Opus 5.5:

The bullets task is where it changed. On four of five runs Argon reported 401–799 thought tokens before writing 42–46 tokens of bullets. Its billed estimate on that task is 9.2× its displayed cost, and higher than Opus 5.5's even at promo prices.

Opus 5.5 thinks too, but less: 84–96 thinking tokens on every arithmetic run, 103–114 on every bullets run, none on JSON. It is billed for all of them, and they are already in the Opus column.

The practical point: on prompts like the JSON and arithmetic tasks, Argon is cheaper per token and stays cheaper after the promo (Opus 5.5 would cost 1.1× Argon on arithmetic at December rates). On prompts that trigger its thinking, it can cost more than Opus 5.5 today. Before 2026-10-22 you could not see which kind of prompt you had from Promptster's displayed cost alone. The display now counts the thoughts, and thoughtsTokenCount in the raw response shows how many there were.

Latency

Median wall-clock time per call, with the full range across five runs:

The ranges do not overlap on any task. Argon's fastest call on each task was slower than Opus 5.5's slowest.

Opus 5.5 was also consistent: its arithmetic calls all landed within 0.14 s of each other. Argon's spread from 4.5 s to 21.8 s on the same prompt. On the bullets task two of Argon's five calls took 37.3 and 37.4 seconds. Those runs reported 447 and 767 thought tokens, while the run with the most thinking (799) took 13.4 s, so thinking alone does not explain the slow calls.

Our 10-word token probe ("Reply with only the word OK…") showed the same thing without any reasoning involved. Argon returned "OK" in 2.3, 13.8 and 15.7 seconds; Opus 5.5 in 0.80–0.88 s. Argon's usage object reported no thought tokens on any probe run, so the delay there is not thinking. The probe also showed Argon counting the 10-word prompt as 43 input tokens, against Opus 5.5's 25.

Latency is from one machine in one session on 2026-10-19 and is not a Google SLA; host-side latency can change week to week. But the gap here held on every call we made.

Which to use

Batch work, today. Argon at promo prices is the cheaper option on every task except the one that made it think, and both models passed everything. If nobody waits on the answer and you can check thoughtsTokenCount, Argon is the cheaper choice until November 30.

Anything interactive. Opus 5.5. A median of 2.5 s against 13.3 s on a short formatting task is not a close call.

Anything you are budgeting past November. Plan at $4 / $20 for Argon, and budget on its billed estimate rather than the displayed cost. On this battery that puts it above Opus 5.5.

Re-test in December. We will. The promo ends, the latency may settle, and the question becomes a straight comparison at identical list prices.

We made a similar point about Claude Sonnet 5's introductory pricing in August: a comparison that quotes a promo rate without its end date is out of date the day it expires.

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, temperature 0.7 and max_tokens 1,500, run 2026-10-19 through Promptster's own provider handlers. Claude 5-family models reject a custom temperature, so none was sent to Opus 5.5; Argon received 0.7. Costs are from shared/pricing.ts at list rate with no cache or batch discounts; December figures double Argon's measured cost to the $4 / $20 rate. The token probe ran three times per model.

For how both compare with open-weight models, see Benchmarking open-source vs. closed-source models.

Try it yourself

• Run the three-bullet instruction task — the task that made Argon think, no account needed
• Run the 10-word token probe — watch the latency

To run Argon and Opus 5.5 side by side on your own prompts and see each provider's token counts, that's Promptster.

All figures from content/blog/data/q4-2026-battery/ and content/blog/data/q4-2026-token-overhead/ (summaries plus one raw file per call, including each provider's usage object verbatim), run 2026-10-19. Displayed cost uses Promptster's normalised token counts; billed estimate re-prices output as total minus input where the provider reports a larger total. Per-pass costs are the 15-call total divided by five. Argon is priced at its promotional $2 / $10, which ends 2026-11-30.