Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
TL;DR Highlight
Gemini 4 Carbon takes #1 on Intelligence Index v5 with 59 and is second in coding agents.
Who Should Read
A backend or agent developer weighing a model switch and comparing LLM API costs
Core Mechanics
- #1 on Intelligence Index v5 with 59
- It has the top score in 4 of 9 evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, AutomationBench-AA.
- It does not lead in coding or knowledge work.
- Second to Opus 5.5 in coding agents
- With Antigravity CLI it scores 59 on the Coding Agent Index, 3 points behind Claude Code with Claude Opus 5.5 (max, 62).
- It ranks second on Terminal-Bench 4.0 (64%) and FrontierCode v1.1 (51%).
- Lower cost per task than Opus 5.5
- At standard pricing it costs $7.76 per Intelligence Index task, about 23% less than Claude Opus 5.5 (max, $10.04).
- It uses 123k output tokens per task, 29% fewer than Opus 5.5 (173k), and costs about 1.2x GPT-6 Astra (max, $6.25).
- 9 points higher than Argon at the same price
- Standard pricing is $4/$20 per 1M input/output tokens, the same as Gemini 4 Argon.
- Cost per task is in line with Argon ($7.68) while the Index score is 9 points higher.
- 54 on AA-Omniscience v1.1, 11% hallucination
- That is 6 points above Claude Opus 5.5 (max, 48) and 8 above Argon (46).
- Accuracy rises 10 points over Argon to 58%, while hallucination rises from 4% to 11%.
- The 11% is second lowest behind Argon, versus 50% for both Claude Opus 5.5 (max) and GPT-6 Astra (max).
- Rolled out only to selected users
- Context is 1M tokens, with text, image, video and speech input and text output.
- The 50% launch discount has no confirmed end date, and cached input gets 95% off ($0.20 per 1M).
Community Reactions
- Doubts about benchmarking an unreleased model
- Commenters said that without public access there is no way to confirm the later release matches the benchmarked model.
- They argued benchmarks should verify the tested model is the released one.
- Doubts about the AA index itself
- Commenters said 60% of the index is private and cannot be replicated, and that benchmark consulting for labs is a conflict of interest.
- One noted the v5 update came two days before a new model took #1.
- Hands-on experience with Argon
- One user said it read 19th-century handwriting well but transcribed only the first column of two-column pages, costing $20 per retry.
- Others reported very low usage limits and insecure generated code.
- Opinion that Gemini's strength is non-coding
- One commenter said it explains scientific concepts best.
- Another said it kept track of agents without losing the thread when supervising them.
How to Apply
- Run it on your own eval set
- When you get access, compare accuracy, tokens and cost against your current model on 20-30 real prompts.
- Use public index scores only as a reference and decide on your own data.
- Shortlist it for hallucination-sensitive tasks
- Compare hallucination rates against your current model in fact-checking Q&A features.
- In the source, Carbon's 11% is lower than Opus 5.5 and GPT-6 Astra (50% each).
- Include cache discounts in cost math
- If you resend long system prompts or documents, recompute monthly cost with the 95% cached-input discount.
- Use standard pricing, since the 50% launch discount has no end date.
Try It
# Paste into Claude or ChatGPT (build a model adoption decision frame from the source numbers)
I run a service that uses LLM APIs. Below are public figures for the new model Gemini 4 Carbon.
- Intelligence Index v5: 59
- Cost per task $7.76 (Claude Opus 5.5 $10.04)
- Output tokens per task 123k (Opus 5.5 173k)
- AA-Omniscience hallucination rate 11% (Opus 5.5 50%)
- Pricing $4/$20 per 1M input/output tokens, cached input $0.20
My service: {{monthly requests}} requests per month, {{input tokens}} input tokens and {{output tokens}} output tokens per request, current model {{current model}}
1) Make a table comparing monthly cost with my current model.
2) List 5 items I should verify myself.
Next: ask for a 20-case test template for each verification item.Terminology
Related Papers
We're releasing the weights of Mistral Large 4
Mistral released Large 4 weights, claiming top open-weight scores on coding benchmarks
Gemini 4 Carbon
Google unveils Gemini 4 Carbon, a cybersecurity-focused model, but access is limited to select cyber defenders.
Dayjob: Tax – evaluating AI agents on 50 professional tax assignments
On 50 professional tax tasks, the best AI agent passed only 10.8%
Anthropic S-1
Anthropic filed an S-1 to list Class A shares on Nasdaq as 'ANT'.
A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
A fine-tuned Inception v3 beat GPT-6 Astra and Claude Fable 5.1 on 200 skin lesions
America has lost 200k of jobs in media, film and the arts. How much is AI?
US creative-industry jobs fell by over 200k in four years, but AI's share of the blame is hard to prove
No signs of AI payoff outside the tech sector
Revenue per worker and margins rise at the Mag 7 but stay flat across the S&P 493
LLMs are making formal software verification feasible at a much larger scale
Claude proved no-crash properties for 123 open-source libraries and found bugs along the way.