On October 28, Google DeepMind released Gemini 4 Carbon. Artificial Analysis scored it 59 on its Intelligence Index v5.0 at the high reasoning setting. That is one point above Anthropic's Claude Opus 5.5 at 58, two above OpenAI's GPT-6 Astra at 57, and nine above Google's own previous flagship, Gemini 4 Argon, at 50.
Alphabet closed up 0.15 percent that day. Coverage of the same release cycle also carried a figure of $500 billion in market value erased in a single day. I can't tell you how that loss maps onto a 0.15 percent move in Alphabet, and I suspect nobody writing the headline could either. But the shape of the story is worth sitting with: the highest score on the board, and almost nobody can call the model.
The leaderboard, with the caveat attached
| Model | Intelligence Index v5.0 |
|---|---|
| Gemini 4 Carbon | 59 |
| Claude Opus 5.5 | 58 |
| GPT-6 Astra | 57 |
| Gemini 4 Argon | 50 |
Those are composite scores across nine independent evaluations, taken at high reasoning effort in late October 2026. A nine-point gap between Carbon and Argon is big for a composite of that size. It is also the least useful nine points in the table, because Carbon has no publicly available API.
Google is granting access to selected cybersecurity defenders through its Fairwind Program. That's it. If you have an API key and a terminal, Carbon is a press release to you.
I think that's a defensible choice for a model with that capability profile. Frontier capability shipped to defenders first is a reasonable safety posture, and eval-gated access programs are how labs test real-world abuse resistance before opening the firehose. But it means the ranking does not describe anything you can buy, and every "Google is back on top" take written this week is describing a model behind a door.
OpenAI went the other direction
GPT-6.1 Terra showed up in release notes on October 26 and became available on October 27 to Plus, Pro, Enterprise and Edu users through ChatGPT Work and Codex. Developers get it as gpt-6.1-terra. Published API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, which OpenAI says is 50 percent below both GPT-5.6 Terra and GPT-6.1 Sol.
The per-task numbers are the ones I'd actually plan around. On Agents' Last Exam at maximum reasoning effort, as reported by Computerworld from OpenAI's measurements:
- GPT-6.1 Terra: 60.0 percent at about $1.01 per task
- GPT-6.1 Sol: 58.6 percent at $2.00 per task
- GPT-6 Astra: 59.3 percent at $7.24 per task
- GPT-5.6 Terra: 50.7 percent at $3.58 per task
Terra trails Astra by 0.7 points on that benchmark and costs about a seventh as much per task. OpenAI's own framing is that Terra matches Astra on a professional-work benchmark at roughly one tenth the price per token. Either way, the score is not the interesting part of that table. The dollars are.
Other results from the release: 74.2 percent on OSWorld 2.1 at maximum effort with an estimated $1.19 per task, close to Astra and Sol on MMMU-Pro, and 50.3 percent on Terminal-Bench 4.0 against 21.5 percent for GPT-5.6 Terra. On an internal evaluation built to expose weak responses, Terra's error rate at low reasoning effort was 16.7 percent versus 53.6 percent for its predecessor. At maximum effort it recorded 4.3 percent. Those are OpenAI's internal numbers, not a general error rate, and I'd treat them as directional.
The speed claim is the one I want independently measured: more than three times faster than GPT-6 Astra and GPT-6.1 Sol on computer-use tasks. For agent loops with timeouts and retry budgets, a 3x speedup changes the architecture more than a benchmark point does.
Nobody has published the benchmark you need
Here's what bothers me about this release cycle. The comparisons that exist are vendor-run or incomparable.
Anthropic's launch page puts Claude Opus 5.5 at 54.6 percent on FrontierCode at default medium effort, above GPT-6 Astra's top score of 53.3 percent, alongside a claim of about one-fifth the cost per task. That is Anthropic grading Anthropic. Zapier's AutomationBench, cited by Anthropic, compares Opus 5.5 against Opus 5 and GPT-5.6 Sol, which is the previous OpenAI generation, not the current one. Perplexity's WANDR test has no OpenAI model in it at all.
Entagl's October 28 review of the model wave found no reliable comparable GPT-6 versus Claude 5.5 benchmark. That's the honest state of the art, and it should be uncomfortable for anyone signing a procurement contract this quarter.
The ChatGPT plan tier is worse. Nobody has published a clean study of the exact models inside each plan, Luna against Sol or otherwise. If you are choosing between the tiers, you are choosing on marketing copy.
What I'd actually do with this
- Do not architect around Carbon. No API, no timeline, defender access only. If Google opens it, you re-evaluate then.
- Rerun your own evals at the reasoning effort you'd deploy. Terra's headline costs are maximum-effort figures. Medium effort scored 15 points above GPT-5.6 Terra on Agents' Last Exam at roughly a third of the cost, which is a different product with a different price.
- Price the cache explicitly. $0.05 per million cached input tokens is where agent savings land if your prompt resends the same system context every turn. If it doesn't, the discount doesn't apply to you.
- Measure latency on your own workload before committing. The 3x computer-use speedup, if it holds, matters more than the benchmark spread.
- Keep GPT-6.1 Sol and GPT-6 Astra in rotation for the hard coding work. OpenAI says so itself in the release notes, and that admission is more useful than any of the composite scores above.
Terra has been live since October 27 at gpt-6.1-terra. Carbon is a number in a table.