Gemini 4 Carbon Costs $7.45 a Coding Test. Google’s 15-Day-Old Argon Costs $17.64 at the Same Price.

What Google shipped, and when

Google made Gemini 4 Argon generally available in the Gemini API on October 13, 2026. On October 28, it announced Gemini 4 Carbon. Google’s own post says Argon launched two weeks earlier.

Carbon’s pitch is coding, agents and security. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench for patching. It also says Carbon gains 89 Elo on WebDev Arena to reach #4 at 1768.

The price is where it gets strange. Google lists Carbon at exactly Argon’s price. Same input rate, same output rate, same jump after the introduction. If two models cost the same per token, the newer one being better is normal. Being better and burning far less money per task is the part worth a closer look.

Terminal-Bench 4.0: same price, 58% less per test

Vals.ai runs Terminal-Bench 4.0 with the same harness, Mini-SWE-agent, for every model, and averages 3 runs. Both Gemini models are priced at the $4 / $20 standard rate on Vals’ model pages, so the costs below are apples to apples.

*Our math: cost per test divided by accuracy. Source for every other number: Vals.ai Terminal-Bench 4.0 leaderboard, updated October 27, 2026.

Argon sits 8th on the 57-model board. Carbon sits 4th, 2.02 points behind Opus 5.5. Per solved task, Argon is the most expensive of the six models in the table, at $30.64.

The token counts only explain part of the gap. Argon uses 34% more tokens per test than Carbon but costs 137% more. Divide cost by tokens and Argon pays about $1.06 per million tokens to Carbon’s $0.60, at identical list prices. Vals does not split those tokens into input, output and cached, so we can’t say which one drives it. A larger share of output tokens, which cost 5x input, or fewer cache hits would both do it.

The other leaderboards are less dramatic

One benchmark is one benchmark. On the two broad indexes, Carbon still wins, but the cost gap shrinks or disappears.

On Artificial Analysis, Carbon spends nearly twice the output tokens of Argon and costs about the same per task, 8 cents more. On the Vals Index it costs 21% less per test. Only on Terminal-Bench does the gap hit 58%.

And Argon still wins somewhere that matters to vibe coders: Vibe Code Bench v1.1, where Vals has it at 91.91% to Carbon’s 89.06%. If you build whole apps from a prompt, the older model is not obsolete.

One caution on Argon’s Vals page: the header shows $13.37 per Vals Index test, while a September 30 update on the same page lists $15.68 at a 68.90% score. Vals does not explain the difference. We use the header figures.

The December 1 catch

Today, Argon bills at $2 / $10 through the Gemini API. Every Vals and Artificial Analysis cost above uses $4 / $20, so the Argon runs on those boards would cost half that until November 30. Starting December 1, Google’s pricing page doubles every Argon rate: standard, Batch, Flex, Priority and cached input.

Carbon’s introductory price has no announced end date. Google says it will reach developers, businesses and consumers “as soon as possible,” after cyber defenders and Fairwind Program partners. So for now, the cheaper model to run is the one you can’t call.

Fun facts

• Google says Carbon has a 0.0% attack success rate on Gray Swan’s prompt-injection benchmark, down from Argon’s 0.7%.
• In Google’s internal testing, Carbon cuts unsupported claims by about 2x over Argon and shows more than a 2x uplift in secure code generation success.
• Argon’s Vals page lists 100.00% on IOI, the olympiad programming benchmark, rank #1 of 46.
• Google’s models page describes Argon with a unified 1M context window and long decodes up to 1M tokens.

What this means if you code with AI

Price per token is not price per task. Carbon and Argon share a price card and differ by 2.4x per Terminal-Bench test. Before you pick a model for a coding agent, look at cost per test or cost per solved task, not the per-million line on the pricing page.

Budget Argon at double after November 30. If you run Argon in an agent today, your December bill doubles at the same usage. Use November to measure what it costs you per merged task, and compare it with Kimi K3.1 Pro, which beats Carbon on Terminal-Bench at $4.57 a test.

Don’t switch your stack for a model you can’t call. Carbon is not in the public Gemini API yet. Plan for it, but keep shipping on what works.

Match the model to the job. For terminal-heavy agent work, Carbon’s numbers are strong. For prompt-to-app vibe coding, Argon’s Vibe Code Bench score is still higher. One leaderboard rarely settles it for your repo.

FAQ

Where MegaBrain fits

Full disclosure: we sell an AI coding agent, and we did not run these benchmarks. MegaBrain’s terminal coding agent puts many models behind one plan: free models on Free, included models such as GLM, Kimi and MiniMax on the paid plans, Claude, GPT and Gemini from Pro, plus your own API keys and local models, and Auto picks one for each task. See the plans. MegaBrain is not affiliated with Google, Vals.ai or Artificial Analysis.

Sources

All pages opened October 29, 2026.

• Google — “Introducing Gemini 4 Carbon” (October 28, 2026): benchmarks, pricing, rollout
• Google AI for Developers — Gemini API pricing: Gemini 4 Argon rates through November 30 and from December 1, 2026
• Google AI for Developers — Gemini API models list
• Vals.ai — Terminal-Bench 4.0 leaderboard (updated October 27, 2026)
• Vals.ai — Gemini 4 Carbon model page
• Vals.ai — Gemini 4 Argon model page
• Artificial Analysis — Gemini 4 Carbon (High)
• Artificial Analysis — Gemini 4 Argon (High)
• 9to5Google — “Google announces Gemini 4 Carbon” (October 28, 2026)