Inter-office memo
- To
- Everyone choosing an AI model
- From
- Mo, Chief Executive
- Date
- October 28, 2026, 4:13 AM CDT
- Re
- Intelligence Index v5
Gemini 4 Carbon now leads the Intelligence Index v5. Opus 5.5 leads what you can use
On October 27, 2026, Artificial Analysis released version 5 of its Intelligence Index, now built 60% on private tests. Google’s Gemini 4 Carbon, announced October 28, leads at 59, though only cyber defenders can use it yet. Of the models you can use, Claude Opus 5.5 at max effort leads at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55, which costs $1.27 a task at API prices against Opus’s $10.04.
What changed in version 5
Artificial Analysis published version 5 of its Intelligence Index on October 27, 2026, “our biggest upgrade since v4 launched in January.” The index is its single score for how capable a model is, and it now averages nine tests in four equal parts:
| Part | Tests, with their weight in the index |
|---|---|
| Knowledge work, 25% | AA-Briefcase 15%, AutomationBench-AA 5%, GDP.pdf 5% |
| Coding, 25% | FrontierCode 2.0 15%, Terminal-Bench 4.0 10% |
| Science, 25% | Terminal-Bench-Science 15%, CritPt v2 hard subset 10% |
| General, 25% | AA-Omniscience v1.1 15%, HLE-Diamond 10% |
Two tests are new. FrontierCode 2.0 is Cognition’s private set of 132 coding tasks in real open-source repositories, “graded on whether a repository maintainer would merge the agent’s change.” Terminal-Bench-Science is 70 scientific research tasks worked in a terminal. HLE-Diamond, a 1,000-question version of Humanity’s Last Exam, replaces the original, and AA-Omniscience and CritPt were revised. GDPval-AA, SciCode and AA-LCR are gone.
Most of it is now private. In Artificial Analysis’s words, “60% of the Index is now based on evaluations with private test sets or held-out answers,” and FrontierCode’s tasks “stay private to avoid contamination,” so models can’t have trained on them. The tests also take more work: the median model wrote 59,000 output tokens per task, up from 35,000 in version 4.3.
Scores from version 4 don’t carry over. Kimi K3.1 Pro scored 55 on version 4.3 and scores 46 now, while Claude Opus 5.5 is at 58 on both.
Who leads
The ten highest-scoring models on October 28, at the effort Artificial Analysis charts for each, with the cost of one index task at API prices:
| Model | Score | Cost per task | Where you can use it |
|---|---|---|---|
| Gemini 4 Carbon (high) | 59 | $7.76 | Not yet: cyber defenders in Google’s Fairwind Program |
| Claude Opus 5.5 (max) | 58 | $10.04 | Claude Pro ($20 a month) and Max |
| GPT-6 Astra (max) | 57 | $6.24 | ChatGPT Plus ($20 a month) and Pro |
| GPT-6.1 Sol (max) | 55 | $1.27 | ChatGPT Plus and Pro, in Work and Codex |
| Claude Sonnet 5.5 (max) | 54 | $10.02 | Every Claude plan, Free included |
| GPT-6.1 Terra (max) | 51 | $0.96 | ChatGPT Plus and Pro, in Work and Codex |
| Gemini 4 Argon (high) | 50 | $7.68 | Google AI Pro ($19.99 a month) and Ultra |
| Muse Wonder Full (max) | 50 | Not given | Muse Code Power ($50 a month) and Maximum |
| Claude Fable 5.1 (max) | 49 | $13.04 | Claude Max, or Pro with usage credits |
| Kimi K3.1 Pro | 46 | $3.68 | Kimi’s apps, Kimi Code and the Kimi API |
The scores and costs are from Artificial Analysis’s leaderboard. The plans are from Claude’s pricing, ChatGPT’s pricing, Codex pricing, Google’s AI plans, Meta’s Muse Code page, Moonshot’s announcement and Google’s announcement of Carbon. Artificial Analysis marks Muse Wonder Full as having no public API, and gives no cost for it.
Artificial Analysis now prices both Gemini 4 models at $4 per million input tokens and $20 per million output, which is what Argon costs on the API from December 1, 2026. Until November 30 it’s $2 and $10, by the Gemini API’s pricing page, so a task at today’s price costs less than the table shows. When this memo was first published, Artificial Analysis charted Argon at its launch price, $3.84 a task.
Gemini 4 Carbon took first place on October 28, a day after the index launched. Google says it will reach developers and its plans “as soon as possible,” but for now only a subset of cyber defenders can use it (our memo on Carbon). Anthropic and OpenAI hold the next five places, and Gemini 4 Argon follows. Further down, Muse Wonder scores 43 at max effort, Claude Haiku 5.5 36, Grok 4.7 35 at high effort, and Gemini 3.8 Flash, the one model Antigravity’s free plan keeps after November 2, 31.
No model leads every test. Gemini 4 Carbon leads HLE-Diamond (69%), Terminal-Bench-Science (65%), AutomationBench-AA, which automates business apps through their APIs (78%, a fraction ahead of Argon), and AA-Omniscience. Opus 5.5 leads FrontierCode 2.0 (46%), and GPT-6 Astra the CritPt physics problems (72%). Claude Sonnet 5.5 leads AA-Briefcase, a test of office work with files; Kimi K3.1 Pro leads Terminal-Bench 4.0 (67%); and Muse Wonder Full leads GDP.pdf, questions answered from long PDFs.
What a task costs, and why effort matters
Cost per task comes from each model’s “input, cache hit, cache write, reasoning, and answer token prices,” so it’s what you’d pay per token on the API. Among the leaders, Artificial Analysis notes that GPT-6.1 Sol “scores within 4 points of Claude Opus 5.5 (max) at about an eighth of the cost per task.” OpenAI charges $2 per million input tokens and $10 per million output for Sol; Anthropic charges $4 and $20 for Opus 5.5. Of the top ten, GPT-6.1 Terra is the cheapest per task and Claude Fable 5.1 the most expensive, more than 13 times as much.
The effort setting, how long a model may think, moves the scores more than it did in version 4.3. Here are the three leaders at each level:
| Effort | Claude Opus 5.5 | GPT-6 Astra | GPT-6.1 Sol |
|---|---|---|---|
| Low | 35, $0.98 | 43, $1.42 | 35, $0.16 |
| Medium | 48, $2.45 | 51, $2.73 | 46, $0.36 |
| High | 52, $3.46 | 55, $3.24 | 50, $0.59 |
| Extra high | 57, $6.05 | 55, $3.66 | 52, $0.76 |
| Max | 58, $10.04 | 57, $6.24 | 55, $1.27 |
Two things stand out. Opus 5.5 at extra high scores 57 for 60% of max’s cost. And defaults matter: Claude Code starts Opus 5.5 at medium, where it scores 48, three points below GPT-6 Astra at medium.
On a plan there’s no bill per task, but there are limits. Anthropic says how much you can do depends on “the model you choose,” and OpenAI’s Codex pricing that “model choice, context, reasoning, tool use, retrieval, and caching all affect usage.” On ChatGPT Plus, Codex allows 15 to 160 local messages every five hours with GPT-6.1 Sol, against 5 to 45 with GPT-6 Astra.
Open models fall further behind
The best open-weights model, Kimi K3.1 Pro, is 13 points behind the leader, Gemini 4 Carbon, and 12 behind Opus 5.5; in version 4.3 the gap was 3. And its weights aren’t out: Moonshot says they “will be released by Nov 4, 2026,” and they weren’t on its Hugging Face page on October 28.
Of models with weights you can download today, DeepSeek V4.1 Flash 1017 and GLM-5.3 score highest, at 32. Artificial Analysis put them “26 points behind the leader,” Opus 5.5 at the time, “and below Claude Haiku 5.5 (max, 36).” Science is where they fall furthest: “most open weights models we benchmarked score under 10%” on Terminal-Bench-Science. Of the smaller models people run on their own computers, Qwen3.8 27B scores 20 at extra high effort. Recursive’s Flint isn’t on the leaderboard.
What the scores don’t tell you
- Fallback. The Claude models were scored “with fallback.” When Anthropic’s safety classifiers decline a request, its API can send it on to “the fallback model Anthropic recommends,” so some answers may come from another Claude model. Artificial Analysis’s announcement doesn’t say how often that happened.
- Your agent. The index tests models in setups of Artificial Analysis’s choosing, such as its own harness, Stirrup, not in the agents people use. In Claude Code, Codex or Antigravity, the same model can score differently; the Coding Agent Index measures those pairings.
- Images, speech and other languages. The methodology calls the index “a primarily text-based, English-language evaluation suite.”
- How far a plan goes. The index prices a task; it doesn’t say how many tasks a week a plan allows.
How to choose
- For the hardest work, Claude Opus 5.5. It leads the models you can use, and FrontierCode, and comes with Claude Pro and Max. Try extra high before max.
- For the most per dollar, GPT-6.1 Sol. It’s within four points of Opus 5.5 at an eighth of the API cost, and on ChatGPT Plus it allows about three times as many Codex messages as Astra.
- For science and research, GPT-6 Astra. It leads both science tests.
- On a Google plan, Gemini 4 Argon, for now. It scores 50, a fraction behind Carbon on AutomationBench, and comes with Google AI Pro and Ultra. Google hasn’t said which plans Carbon will join.
- For open weights, Kimi K3.1 Pro once its weights arrive. Until then, it’s a hosted model like the rest.
Using more than one
The leaders split the nine tests between them, and their costs differ by more than ten times, which argues for keeping more than one at hand. Enjoy runs Claude Code, Codex, Antigravity, Muse Code, Grok Build and OpenCode side by side on your computer, each signed in with the account you already have. You can switch agents in the middle of a conversation, and the new agent gets the conversation so far, and you can set the effort in your message, such as “quick one, low effort.”
Enjoy is a monthly subscription, with the first seven days free (pricing). It doesn’t add Opus to a ChatGPT plan or GPT to a Claude one: each agent works on its own lab’s plan.
We’ll update this memo when Kimi K3.1 Pro’s weights arrive, Gemini 4 Carbon reaches Google’s plans, or Artificial Analysis changes the index again.
8.Questions.
What is the smartest AI model right now?
- By Artificial Analysis’s Intelligence Index v5, it’s Google’s Gemini 4 Carbon, which scored 59 on October 28, 2026, though only a subset of cyber defenders can use it so far. Of the models you can use, Claude Opus 5.5 at max effort leads at 58, then GPT-6 Astra at 57 and GPT-6.1 Sol at 55. No model leads all nine of the index’s tests.
Is Claude Opus 5.5 better than GPT-6 Astra?
- By one point overall, 58 to 57, and it depends on the work. Opus 5.5 is ahead on the coding test FrontierCode 2.0 and the exam HLE-Diamond; Astra is ahead on both science tests. Astra also costs less, $6.24 a task at API prices against $10.04.
Which AI model is the best value?
- GPT-6.1 Sol, on this index. At max effort it scores 55, within four points of Opus 5.5, for $1.27 a task, about an eighth of the cost. GPT-6.1 Terra scores 51 for $0.96. Both come with ChatGPT Plus, in Work and Codex.
What is the best open-weights AI model?
- Kimi K3.1 Pro, at 46, but Moonshot hasn’t released its weights yet; it promises them by November 4, 2026. Of the models you can download today, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, below Claude Haiku 5.5 at 36.
Why did some models’ scores drop?
- Because version 5 changed the tests, so its scores don’t compare with version 4’s. Kimi K3.1 Pro went from 55 on version 4.3 to 46 while Opus 5.5 stayed at 58, which is why the gap between the best open and closed models grew from 3 points to 12, and to 13 once Gemini 4 Carbon took the lead.
Does the cost per task apply to my ChatGPT or Claude plan?
- No. Artificial Analysis prices each task from the model’s API token prices. On a plan you pay a flat price, and a stronger model or a higher effort uses up more of your limits instead.