October 2026 in AI models: Gemini 4 Carbon takes first, Haiku 5.5 falls to a tenth of the price, and the index changes scale
Gemini 4 Carbon takes first but few can call it, Claude Haiku 5.5 drops to $0.10 / $0.50, GPT-6.1 Terra reaches Codex, and the index moves to v5.0.
GTM Architect & Growth Operator · Now · 1 November 2026 · 7 min read
On this page
October brought releases that matter if you pay for tokens, and one change that alters how every older score on this site reads. The change comes first.
The scale changed, and the scores moved with it
Artificial Analysis moved its Intelligence Index to v5.0. The methodology page says GDPval-AA was removed in v5.0, having been a constituent up to and including v4.3.2, and that the suite is now four categories of 25% each: Knowledge Work, Coding, Science and General. VentureBeat’s report on the Carbon launch adds that 60% of the new suite uses private tests or withheld answers.
Same model, two scales. These are the models I had on v4.3.2 in earlier pieces, and where they sit now.
| Model | v4.3.2 | v5.0 |
|---|---|---|
| Claude Opus 5.5 | 58 at $5.98 a task | 58 at $10.04 a task |
| Claude Sonnet 5.5 | 56 at $7.62 a task | 54 at $10.02 a task |
| GPT-6 Astra | 53 at $3.26 a task | 57 at $6.24 a task |
| GPT-6.1 Sol | 52 at $0.72 a task | 55 at $1.27 a task |
Opus 5.5 kept its score and got more expensive to run. Sonnet 5.5 lost two points. Sol gained three and the gap to Astra widened from one point to two. None of that is the models changing. It is the ruler.
Do not mix the scales
A v4.3.2 number and a v5.0 number do not belong in one sentence. The earlier comparisons here keep their v4.3.2 figures and say so. Four of them called Opus 5.5 first of 211 or 212 models. On v5.0 it is second of 58, and each now carries a note saying so.
The board on v5.0
| Model | Index | Per index task | List price, in / out per 1M |
|---|---|---|---|
| Gemini 4 Carbon | 59 | $7.76 | $4 / $20 ($2 / $10 introductory) |
| Claude Opus 5.5 | 58 | $10.04 | $4 / $20 |
| GPT-6 Astra | 57 | $6.24 | not rechecked here |
| GPT-6.1 Sol | 55 | $1.27 | $2 / $10 |
| Claude Sonnet 5.5 | 54 | $10.02 | $2 / $10 |
| GPT-6.1 Terra | 51 | $0.96 | $1 / $5 |
| Gemini 4 Argon | 50 | $7.68 | $2 / $10 to 30 Nov, then $4 / $20 |
| Claude Haiku 5.5 | 36 | $0.41 | $0.10 / $0.50 to 100k tokens |
| Grok 4.7 | 35 | $6.18 | $2 / $6 |
| Gemini 3.8 Flash | 31 | $7.77 | $0.75 / $3.75 to 31 Dec, then $1.50 / $7.50 |
Read the third column against the fourth. A low sticker price does not make a cheap job: Gemini 3.8 Flash lists at $0.75 input and costs $7.77 to run the index. GPT-6.1 Sol lists at $2 and costs $1.27.
What each release is worth
Gemini 4 Carbon, 28 October. Score 59, first of 58, one point above Opus 5.5. Google gives access to selected cyber defenders through its Fairwind Program and has named no date for more. Artificial Analysis lists it without a public API. First place on a chart you cannot call is a press release. I put it against the model you can call in Gemini 4 Carbon against Claude Opus 5.5.
Claude Haiku 5.5, 7 October. $0.10 input and $0.50 output up to 100,000 prompt tokens, against $1 and $5 on Haiku 4.5. It scores 36 at $0.41 an index task. It is also verbose: 300 million tokens on the run against a median of 72 million, so the per-token price flatters it. The pair I checked it against is Haiku 5.5 against Grok 4.7.
GPT-6.1 Terra, 26 October in Work and Codex. $1 input, $0.05 cached, $5 output, 1,050,000 tokens of context. It scores 51 at $0.96 a task. Sol scores 55 at $1.27, so four points cost 32% more per task. ChatGPT’s changelog lists Terra for Work and Codex and says it is not in Chat.
Gemini 4 Argon, generally available 13 October. Score 50 at $7.68 a task. Its introductory price of $2 / $10 ends on 30 November and becomes $4 / $20 on 1 December. If you budgeted on the cheap number, the bill doubles in a month.
What I would switch this week
- Nothing, because of Carbon. Until you can call it, it is not a candidate.
- Classification, extraction and routing to Haiku 5.5, behind a test on your own data. Anthropic describes it for exactly that work, and at $0.41 an index task it is the cheapest row in the table.
- Opus 5.5 fast mode, if you still use it on Opus 4.8. Anthropic’s release notes say fast mode on Opus 4.8 is removed on 30 November and point to Opus 5.5.
- Re-check any Argon budget before December.
Dates to put in a calendar
| Date | What happens |
|---|---|
| 30 Nov 2026 | Fast mode for Claude Opus 4.8 is removed. Gemini 4 Argon's introductory $2 / $10 ends. |
| 1 Dec 2026 | Gemini 4 Argon moves to $4 / $20 per million tokens. |
| 31 Dec 2026 | Gemini 3.8 Flash introductory pricing ends. It becomes $1.50 / $7.50. |
What this page will not tell you
It is one index, run by one outfit, on a suite that changed this month and is mostly private. It says which model scores higher on that suite and what the run cost. It does not say which model is right for your prompts. Run a hundred of your own through the two you are weighing before you move traffic.
If you came for the agents rather than the models, the companion piece is what stops your agent.
Questions people asked
Which AI model is best in November 2026?
On Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon scores 59, Claude Opus 5.5 58 and GPT-6 Astra 57. Carbon has no public API, so among models you can call, Claude Opus 5.5 is first at 58. On cost per task GPT-6.1 Sol is the standout: 55 at $1.27, against $10.04 for Opus 5.5.
Is Gemini 4 Carbon available?
Not to most people. Google announced it on 28 October with access limited to selected cyber defenders through its Fairwind Program, and has given no date for wider access. Gemini 4 Argon is the model generally available on the Gemini API.
What is Artificial Analysis Intelligence Index v5.0?
A revised version of the index that removes GDPval-AA and is organised into four categories of 25% each: Knowledge Work, Coding, Science and General. Its scores are not comparable with v4.3.2. Scores on this site from before October use the older scale and say so.
How cheap is Claude Haiku 5.5?
$0.10 per million input tokens and $0.50 output for prompts up to 100,000 tokens, and $0.50 / $2.50 above that. Claude Haiku 4.5 lists at $1 / $5. At the higher prompt size Haiku 5.5 is half the price of Haiku 4.5, below it a tenth.